Software radio based electrostatic discharge transient signal detection method and apparatus

Abstract

The present application relates to the technical field of electromagnetic compatibility monitoring and transient electromagnetic interference detection, in particular to a static discharge transient signal detection method and device based on software radio, comprising: collecting and digitizing and buffering static discharge transient signals, pre-processing the collected continuous signals, waveform window intercepting and standardizing the burst signals meeting the conditions, sending the standardized data into a double-branch recognition model for recognition processing, splicing the feature vectors of the two branches, and completing signal recognition after processing by a fusion discrimination layer, compared with the prior art, the method can effectively improve the distinguishing ability of ESD signals and interference signals, has high recognition accuracy and good engineering application value.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic compatibility monitoring and transient electromagnetic interference detection technology. Specifically, it is a software-defined radio-based method and device for detecting electrostatic discharge transient signals suitable for engineering field environments. This method and device can achieve stable capture and efficient identification of transient signals while ensuring equipment portability and low cost. Background Technology

[0002] Electrostatic discharge (ESD) is one of the most common sources of transient electromagnetic interference in electronic equipment. It is typically generated by processes such as human contact, frictional separation of objects, or material movement. When accumulated charge is released instantaneously under the influence of a potential difference, it forms a transient electromagnetic pulse with a steep rise time, short duration, and high amplitude, which can lead to equipment malfunctions, communication abnormalities, data errors, or system restarts. Therefore, in electromagnetic compatibility testing and on-site troubleshooting of electronic equipment, effective detection and analysis of ESD events are of practical significance for assessing the immunity level of equipment and locating interference sources.

[0003] Current electrostatic discharge (ESD) detection methods primarily rely on testing equipment such as broadband oscilloscopes, RF receivers, or spectrum analyzers. While these devices can capture and analyze transient signals, they generally suffer from high cost, large size, and stringent requirements for power supply and testing environments, making them unsuitable for applications in engineering fields, mobile testing, and space-constrained scenarios. Existing portable testing devices mainly focus on capturing and recording transient events, lacking the ability to identify and differentiate between different types of transient electromagnetic interference signals. In complex electromagnetic environments, transient interference signals generated by switching power supplies, motors, or relay operations share certain similarities with ESD signals, easily leading to misjudgments and affecting the accuracy of test results.

[0004] Furthermore, the relevant detection systems usually need to be deployed on low-power embedded platforms. Due to limited computing resources, how to achieve real-time detection and accurate identification of electrostatic discharge signals under limited computing power conditions remains a technical problem to be solved in the existing technology. In summary, the existing technology still has the following shortcomings: (1) Existing electrostatic discharge detection equipment mostly relies on high-end testing instruments, which are large in size and expensive, making them unsuitable for portable detection and application in engineering sites; (2) Existing portable detection equipment usually only has the function of recording transient signals and lacks the ability to intelligently identify and distinguish different types of transient electromagnetic interference signals. Summary of the Invention

[0005] This invention addresses the shortcomings and deficiencies of existing technologies by proposing a software-defined radio-based method and apparatus for detecting electrostatic discharge transient signals suitable for engineering field environments. This method and apparatus can achieve stable capture and efficient identification of transient signals while ensuring equipment portability and low cost.

[0006] This invention achieves its purpose through the following measures: A method for detecting transient electrostatic discharge signals based on software-defined radio, characterized by comprising the following steps: Step 1: Acquisition, digitization, and buffering of electrostatic discharge transient signals: Electromagnetic radiation signals in the 5MHz-25MHz frequency band are acquired, and through down-conversion and analog-to-digital conversion, a complex baseband signal is generated. Its discrete expression is as follows: (1), where and These represent the in-phase component and the quadrature component, respectively, and the continuously sampled data is stored in real time based on a ring buffer structure; j in the formula is the symbol of the imaginary unit commonly used in the fields of communication signal processing and electrical engineering, which satisfies j²=-1. Its function is to indicate that the quadrature component Q(n) constitutes the imaginary part of the complex signal, and it is not a technical parameter that needs to be determined separately in this application. Step 2: Preprocess the acquired continuous signal, including first performing amplitude square operation on the I / Q signal in the buffer to obtain the energy sequence signal characterizing the change of electromagnetic field intensity. In order to suppress random noise fluctuations, the energy sequence signal is subjected to moving average filtering. A threshold decision mechanism is used to complete the detection trigger. That is, by setting a threshold by preset or adaptive setting, when the threshold is met, it is determined that there is a potential transient event at that position, and the initial trigger position is recorded. Step 3: Perform waveform window truncation and standardization on burst signals that meet the conditions: After a burst signal is detected, extract a data window of fixed length centered on the trigger position, align the labeled data, and uniformly truncate the length of the output data to ensure that the length of the burst signal segment output each time is consistent, thus forming standardized data; Step 4: Feed the standardized data into the dual-branch recognition model for recognition processing: The first branch is the I / Q raw signal feature extraction branch, which adopts a progressive convolutional modeling method to extract local mutations and global structural features from the temporal dimension; the second branch is the statistical feature learning branch, which is used to calculate the rising edge slope, multi-scale energy proportion and spectral centroid, and performs nonlinear mapping through fully connected layers to achieve feature compression and recombination, and obtain a feature representation with strong discriminative ability; Step 5: Concatenate the depth feature vector output from the first branch with the statistical feature vector output from the second branch. After processing by the fusion discriminant layer, output the probability value of the signal belonging to the electrostatic discharge category through the Sigmoid activation function to complete signal recognition.

[0007] In step 2 of this invention, the amplitude squared operation is performed on the I / Q signal to characterize the change in electromagnetic field intensity, and its expression is: (2) To reduce the impact of noise fluctuations on the detection results, a moving average filter is introduced to smooth the energy sequence. The calculation method is as follows: (3), in, Indicates the first n The smoothed energy value at each location after moving average. Indicates the current point n Counting forward k The energy value of each sampling point Indicates the length of the sliding window. .

[0008] In step 3 of this invention, during waveform truncation, the original data window containing the trigger region is first extracted from the annular buffer based on the output of the burst detection module. The initial window length is set to 100 sampling points to cover the complete transient response process. An adaptive trigger point relocation mechanism based on amplitude constraints is introduced within the window. After relocation, 64 points are symmetrically truncated around the relocated trigger point as the model input. The amplitude of the I / Q data within the window is then calculated. (4), And according to the preset threshold Select the set of valid sample points that meet the criteria. : (5), Based on this, the geometric center of the set is used as the new reference trigger position. : (6), in Indicates the size of the set. This indicates index summation, enabling adaptive correction of the initial trigger point and eliminating time offset issues caused by threshold detection errors or noise fluctuations.

[0009] The standardization process described in step 3 of this invention specifically involves: performing amplitude normalization on the truncated waveform data, assuming the original waveform sequence is... , in and Let represent the minimum and maximum values ​​in the sequence, respectively. Then the normalized data... It can be represented as: (7).

[0010] In step 4 of the present invention, the dual-branch recognition model includes the following I / Q original signal branches: Local transient feature extraction unit: A lightweight one-dimensional convolutional structure with a kernel size of 5 and 32 channels is used to model the signal locally. A depthwise separable convolution method is used to capture short-time-scale features such as amplitude abrupt changes and rising edges. Multi-scale temporal feature extraction unit: It adopts a convolutional structure with a kernel size of 5 and a channel number of 64. It extracts signal features in different time ranges through multi-scale convolutional kernels and uses a depth-separable convolutional structure to achieve multi-scale modeling, which improves the ability to characterize complex transient patterns while maintaining temporal resolution. High-dimensional semantic feature extraction unit: It adopts a convolutional structure with a kernel size of 3 and 128 channels to perform high-level abstraction and compression of the previous features. It generates discriminative feature embeddings by combining depthwise separable convolution with global feature aggregation, thereby realizing the expression of high-level semantic information.

[0011] In step 4 of this invention, the dual-branch recognition model introduces a multi-dimensional feature modeling method based on the ESD discharge mechanism in the statistical feature learning branch, which extracts the pulse envelope. To pinpoint the key time points within the 10% to 90% range of its peak value. and And calculate the slope of the rising edge accordingly. : (8), in: The amplitude at which the signal amplitude reaches 10% of its peak value. The amplitude at which the signal amplitude reaches 90% of its peak value. The point in time when the amplitude first reaches a 10% peak value during the signal's ascent. The point in time when the amplitude first reaches 90% of its peak value during the signal's ascent. To characterize the distribution characteristics of signal energy over time, a multi-scale energy ratio method is used to calculate the signal energy distribution in different time windows (2). k Energy percentage within a scale : (9), of which: The complete transient electromagnetic signal throughout the entire time window. Signals within a sub-window at a certain time. Signal energy Sum all sampling points within the time window. The spectral centroid is used to characterize the central tendency of the spectral energy distribution of an electrostatic discharge signal. Obtained from the following formula: (10), where: :frequency, Corresponding frequency f The spectral amplitude or power spectral density at that location. Weighted frequency Total spectral energy.

[0012] In step 5 of this invention, after completing the two-branch feature extraction, the deep representation features extracted by the temporal structure branch and the physical attribute features constructed by the statistical feature branch are first jointly modeled. The discrimination process of the model is expressed as follows: (11), in: This represents the deep feature vector extracted from the I / Q raw signal branch. This represents the feature representation extracted by the statistical feature branch. This indicates a feature concatenation operation. This is the weight matrix of the fusion layer. For bias terms, This represents the Sigmoid activation function. The feature concatenation operation is used to fuse multi-source information, combining the detailed structural features of the signal extracted by the waveform branch with the statistical features of the signal extracted by the statistical feature branch to form a complete feature representation containing both temporal structural information and physical statistical information, thereby achieving high-precision recognition. The two types of features are complementary in information expression. The I / Q branch focuses on characterizing the transient waveform structural features, while the statistical feature branch focuses on describing energy evolution and spectral distribution characteristics. The combination of the two enhances the ability to distinguish signals under complex interference environments.

[0013] Before training the dual-branch recognition model, this invention also constructs a training dataset: electrostatic discharge signal samples are collected using an electrostatic discharge generator conforming to the IEC61000-4-2 standard under different voltage levels, test distances, and environmental conditions; interference signal samples generated by power strip switches, light switches, or lighter switches are collected; data enhancement is performed on the I / Q dual-channel signals, including at least one of amplitude scaling, phase rotation, noise perturbation, and time jitter, to generate enhanced samples.

[0014] This invention also proposes a software-defined radio-based electrostatic discharge transient signal detection device, characterized by comprising a signal acquisition module, a data processing and recognition module, and a result output module. The signal acquisition module includes a broadband antenna and an SDR receiver. The signal acquisition module sends the acquired I / Q data to the data processing and recognition module. The data processing and recognition module includes an embedded processing platform, a burst detection and waveform truncation processing unit, a dual-branch recognition model feature extraction model, and an ESD recognition model.

[0015] The signal acquisition module of this invention is used to receive electromagnetic radiation signals in the 5 MHz to 25 MHz frequency band through a broadband antenna, and down-converts them into complex baseband I / Q signals through a software-defined wireless device. The embedded processing module is connected to the software-defined wireless device through a USB interface and runs the above-mentioned detection method to realize burst signal triggering, waveform standardization, feature extraction and signal recognition. A display module connected to the embedded processing module is also provided to display the detection results and recognition probability in real time. The embedded processing module adopts a Raspberry Pi platform, and the dual-branch recognition model is deployed locally in a lightweight form without cloud computing.

[0016] In this invention, the sampling rate of the software-defined radio device is 20 MHz and the center frequency is 15 MHz; the embedded processing module maintains a ring buffer in memory to cache the I / Q data of the most recent N sampling points, where N≥100, to support backtracking extraction of the complete transient waveform after triggering.

[0017] The device described above in this invention is a handheld integrated device with independent power supply and operation capabilities, and is suitable for electromagnetic compatibility troubleshooting in engineering sites.

[0018] Compared with existing technologies, this invention aligns the key energy regions of different samples on the time axis through temporal relocation and symmetrical truncation of data, thereby enhancing the consistency of feature distribution and improving the model's discriminative ability. Symmetrical truncation based on the relocation results ensures that the main energy regions of different samples remain aligned on the time axis, effectively reducing the model's sensitivity to time shifts and improving recognition accuracy and generalization ability. Furthermore, by constructing a collaborative fusion mechanism of temporal structural features and physical statistical features, effective coupling and discrimination enhancement of multi-source information are achieved in a low-dimensional mapping space. This improves the ability to distinguish between ESD transient signals and other interference signals without significantly increasing model complexity, maintaining high recognition accuracy and real-time performance, especially under the condition of limited computing resources on embedded platforms. Experimental verification shows that this invention can effectively improve the ability to distinguish between ESD signals and interference signals, possessing high recognition accuracy and good engineering application value. Attached Figure Description

[0019] AppendixFigure 1 This is a schematic diagram of the dual-branch recognition model in this invention.

[0020] Appendix Figure 2 This is a feature curve diagram in an embodiment of the present invention.

[0021] Appendix Figure 3 This is a block diagram illustrating the principle of the detection device in this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments: To address the shortcomings of existing electrostatic discharge (ESD) detection devices in terms of portability, cost, and signal recognition capability, this paper proposes a software-defined radio-based method for detecting transient ESD signals. Based on this method, a low-cost handheld ESD detection and identification device is proposed. The software-defined radio-based ESD transient signal detection method includes the following: Part 1: Transient electromagnetic signal acquisition and trigger detection method based on SDR; Part 2: Adaptive truncation and standardization processing method for transient electromagnetic signal feature waveforms based on trigger point localization; Part 3: Sample construction and data augmentation method for electrostatic discharge signal recognition; Part 4: Construction of a dual-branch multi-feature fusion electrostatic discharge signal recognition model based on depthwise separable convolution; Part 5: Lightweight model deployment and real-time electrostatic discharge signal recognition system implementation method based on embedded platform.

[0023] This example, in its first part, provides a method for acquiring and detecting transient electrostatic discharge (ESD) signals based on Software Defined Radio (SDR). This method is used for continuous monitoring, real-time detection, and waveform capture of transient electromagnetic radiation signals generated by ESD in complex electromagnetic environments. The system consists of functional modules for RF signal reception, data buffer management, and burst signal trigger detection. It achieves efficient acquisition and processing of transient signals through a software-defined radio processing chain. During signal acquisition, the system uses a software-defined wireless device as the front end. By setting the sampling rate to 20 MHz and the center frequency to 15 MHz, it receives electromagnetic radiation signals in the 5 MHz to 25 MHz frequency band in real time. After the radio frequency signal is received by the antenna, it undergoes down-conversion and analog-to-digital conversion to generate a complex baseband signal, the discrete expression of which is as follows: (1), where and These represent the in-phase and quadrature components, respectively. To ensure complete capture of transient signals, the system uses a ring buffer structure to store continuously sampled data in real time. When a trigger event occurs, it can backtrack and extract historical data before and after the trigger point, thereby constructing a transient waveform sequence containing a complete time context. In this design, the system extracts 50 sampling points before and 50 sampling points after the trigger point, forming a standardized burst signal segment of 100 points in length. For transient signal detection, the system first performs amplitude square operation on the I / Q signals to characterize the change in electromagnetic field intensity, the expression of which is: (2). To reduce the impact of noise fluctuations on the detection results, a moving average filter is introduced to smooth the energy sequence. The calculation method is as follows: (3); in, Indicates the first n The smoothed energy value at each location after moving average. Indicates the current point n Counting forward k The energy value of each sampling point This represents the length of the sliding window; in this system, it is taken as... This processing effectively suppresses random noise and improves the stability and robustness of burst detection. Based on this, the system employs a threshold decision mechanism to detect the smoothed signal; when the threshold is met... It can determine the presence of potential transient events at a given location and can also make adaptive adjustments based on the background noise level.

[0024] Upon detecting a burst signal that meets the specified conditions, the system marks the raw data stream as an event and constructs a fixed-length data window centered on the trigger position. The tagged data is then aligned, and the output data length is uniformly truncated to ensure that each output burst signal is a standard data sequence of consistent length. Finally, the system encapsulates the processed burst signal into a message data format and sends the data to the identification unit for recognition via a ZMQ-based communication interface, achieving real-time data interaction and analysis.

[0025] In summary, this example demonstrates how a software-defined radio (SDK)-based signal processing workflow organically combines key steps such as continuous acquisition, energy calculation, moving average filtering, threshold trigger detection, and data window reconstruction to achieve high-precision acquisition and low-latency output of electrostatic discharge transient signals. Compared to traditional continuous full-data acquisition methods, this approach significantly reduces data processing volume and transmission burden through a mechanism of "amplitude mutation detection + local window truncation," while ensuring that key transient features are not lost, thereby improving the system's real-time performance, stability, and engineering application value.

[0026] The second part of this invention includes the following: First, waveform truncation and data structuring reconstruction: During waveform truncation, the system first extracts the original data window containing the trigger region from the circular buffer based on the output of the burst detection module. In this embodiment, the initial window length is set to 100 sampling points to cover the complete transient response process. Unlike the traditional fixed-start point truncation method, this example introduces an adaptive trigger point relocation mechanism based on amplitude constraints within this window. After relocation, 64 points are symmetrically truncated around the relocated trigger point as the model input, and amplitude calculation is performed on the I / Q data within the window. (4), And according to the preset threshold Select the set of valid sample points that meet the criteria. : (5), Based on this, the geometric center of the set is used as the new reference trigger position. : (6); in Indicates the size of the set. The formula represents index summation. By averaging the indices of valid sample points exceeding the threshold, it estimates the center position of the transient signal energy concentration region, thereby achieving adaptive correction of the initial trigger point and effectively eliminating the time offset problem caused by threshold detection error or noise fluctuation.

[0027] Unlike traditional methods that rely on single-point trigger locations, this approach does not directly use the initial trigger point as a time reference. Instead, it statistically models the effective sample set comprised of all sampling points exceeding a threshold within the window. By calculating the distribution center of this set on the time axis, the center location of the transient signal energy concentration region is obtained. This method extends single-point determination to regional statistics, effectively suppressing the impact of noise fluctuations and local anomalies on trigger location, thereby achieving a more stable time reference point estimation.

[0028] Because deep learning models are sensitive to the temporal structure of input data, when similar transient signals are offset on the time axis, the features learned by the model will become scattered, thereby reducing the stability of the classification boundary. This invention uses temporal relocation and symmetrical truncation to align the key energy regions of different samples on the time axis, thereby enhancing the consistency of feature distribution and improving the model's discriminative ability. Based on the relocation results, symmetrical truncation is performed to keep the main energy regions of different samples aligned on the time axis, which can effectively reduce the model's sensitivity to time offset and improve recognition accuracy and generalization ability.

[0029] Secondly, data standardization processing: Data standardization is used to normalize the amplitude of the truncated waveform data to eliminate the influence of different sampling conditions or environmental noise on the signal amplitude range.

[0030] The system employs a minimum-maximum normalization method to process the waveform data, mapping the normalized signal amplitude to a uniform range. Let the original waveform sequence be... , in and Let represent the minimum and maximum values ​​in the sequence, respectively. Then the normalized data... It can be represented as: (7), The above normalization process can keep the amplitude range of different sample data consistent, thereby improving the stability of subsequent feature extraction and signal recognition model training and inference.

[0031] The third part of this example includes the following: To effectively simulate the electrostatic discharge process in a real industrial environment, this example uses a professional electrostatic discharge generator conforming to the IEC 61000-4-2 standard as the excitation source. This device can generate transient electromagnetic pulse signals with good repeatability, and its discharge voltage can be continuously adjusted within the range of ±0.2 kV to ±30 kV.

[0032] During data acquisition, electrostatic discharge signals were collected under different discharge types, test distances, environmental conditions, and discharge voltage levels to improve sample diversity and environmental adaptability. Simultaneously, to construct interference signal samples, typical everyday electrical devices such as power strip switches, light switches, and lighter switches were selected as interference sources, and their transient electromagnetic signals were collected to simulate complex electromagnetic interference situations in real-world applications.

[0033] Based on the above acquisition process, the acquired signal data were categorized to construct a multi-category dataset containing electrostatic discharge (ESD) signals and non-ESD interference signals. The original dataset included 1106 ESD signal samples and 508 interference signal samples.

[0034] Furthermore, to increase the dataset size and enhance the model's adaptability to complex environmental changes, this example designs a data augmentation mechanism for transient electromagnetic signals, specifically for I / Q dual-channel signals. While retaining the original sample category labels, two additional augmented samples are generated for each original sample. In one implementation, the data augmentation mechanism includes at least one of the following processing methods: amplitude scaling enhancement, phase rotation enhancement, noise perturbation enhancement, and time jitter enhancement. These enhancement methods are used to simulate the changes in transient electromagnetic signals during electrostatic discharge under different discharge intensities, propagation paths, receiving conditions, background noise, and triggering deviations, thereby expanding the training sample distribution and improving the robustness of the recognition model in complex electromagnetic environments.

[0035] After the above enhancement processing, the dataset size was expanded to 3318 electrostatic discharge signals and 1524 interference signals. This data construction and enhancement method enriched the different types of signals in terms of sample quantity, environmental distribution, and signal morphology, effectively improving the recognition accuracy, generalization ability, and anti-interference ability of subsequent deep learning models in complex electromagnetic environments.

[0036] Part 4 of this example first includes: I / Q raw signal feature extraction branch: The I / Q raw signal branch is used to automatically extract waveform features from complex baseband signals. Its input is a 64-bit I / Q sequence. To reduce the computational complexity of the model, a three-layer one-dimensional depthwise separable convolutional structure is used for feature extraction. This structure decomposes the convolution operation into depthwise convolution and pointwise convolution, significantly reducing the number of model parameters and computational cost while maintaining the same effect. Normalization and max pooling layers are introduced after each convolution layer to improve training stability and enhance the ability to extract transient signal features. This branch achieves step-by-step feature extraction of electrostatic discharge transient signals through a hierarchical structure. It adopts a progressive convolutional modeling approach to collaboratively represent signal features at different time scales, thereby achieving unified modeling of local mutation features and global structural features.

[0037] Specifically, the I / Q raw signal branches include: (1) Local transient feature extraction unit: A lightweight one-dimensional convolutional structure with a kernel size of 5 and a channel number of 32 is used to model the signal locally. It is preferred to use a depth-separable convolutional method to enhance the response capability to signal change regions and to capture short time scale features such as amplitude change and rising edge. (2) Multi-scale temporal feature extraction unit: A convolutional structure with a kernel size of 5 and a channel number of 64 is adopted. The signal features in different time ranges are extracted through multi-scale convolutional kernels. The depth-separable convolutional structure is preferred to realize multi-scale modeling, which improves the ability to characterize complex transient patterns while maintaining temporal resolution. (3) High-dimensional semantic feature extraction unit: a convolutional structure with a kernel size of 3 and a channel number of 128 is adopted to perform high-level abstraction and compression of the previous features. It is preferred to generate feature embedding with discriminative ability by combining depthwise separable convolution with global feature aggregation, so as to realize the expression of high-level semantic information. Feature normalization and downsampling mechanisms are implemented between the aforementioned units to improve model training stability and suppress redundant information, thereby enhancing the ability to extract key features of electrostatic discharge signals. The parameter settings follow the principle of local-to-global and low-dimensional-to-high-dimensional feature extraction. Larger convolutional kernels are used to enhance the response to transient changes, while smaller kernels are used to improve the accuracy of high-level feature representation. The number of channels is increased layer by layer to enhance the richness of feature representation, thus achieving efficient characterization of electrostatic discharge signals while ensuring controllable computational complexity. Furthermore, the convolutional parameters are not limited to the above values ​​and can be adjusted according to different sampling rates or signal lengths.

[0038] It also includes a statistical feature learning branch, which constructs transient feature vectors based on the ESD discharge mechanism to characterize the dynamic changes of the signal during the discharge process. To compensate for the shortcomings of simple deep learning methods in characterizing physical properties, this invention introduces a multi-dimensional feature modeling method based on the ESD discharge mechanism. By analyzing the signal in multiple dimensions such as the time domain, frequency domain, and energy domain, a feature space with high discriminative power is constructed. Specifically, this is achieved by extracting the pulse envelope. To pinpoint the key time points within the 10% to 90% range of its peak value. and And calculate the slope of the rising edge accordingly. : (8), in: The amplitude at which the signal amplitude reaches 10% of its peak value. The amplitude at which the signal amplitude reaches 90% of its peak value. The point in time when the amplitude first reaches a 10% peak value during the signal's ascent. The point in time when the amplitude first reaches 90% of its peak value during the signal's ascent; Furthermore, to characterize the distribution characteristics of signal energy across time scales, a multi-scale energy ratio method is employed to calculate the signal energy distribution in different time windows (2...k Energy percentage within a scale : (9), in: The complete transient electromagnetic signal throughout the entire time window. Signals within a sub-window at a certain time. Signal energy Sum of all sampled points within the time window; In the frequency domain, the signal spectrum is obtained through Fast Fourier Transform. The spectral centroid is used to characterize the concentration trend of the energy distribution of the electrostatic discharge signal spectrum. Further calculation of the spectral centroid is then performed. : (10) in: :frequency, Corresponding frequency f Spectral amplitude or power spectral density at [location] Weighted frequency Total spectral energy is used to depict the broadband continuous spectrum characteristics of the ESD signal and effectively distinguish it from narrowband interference signals. This branch employs a multilayer perceptron structure composed of multiple fully connected layers, and achieves feature compression and recombination through nonlinear mapping, thereby obtaining a feature representation with stronger discriminative capabilities.

[0039] This example also includes feature fusion and classification methods: A lightweight fusion and discrimination method for multi-source features is proposed. The model concatenates the output vectors of the I / Q branch and the statistical feature branch through a feature fusion layer to form a fused feature vector. Subsequently, feature interaction is performed through a fully connected layer, and finally, the probability value P of the signal belonging to the electrostatic discharge category is output through the Sigmoid activation function, achieving signal classification and recognition.

[0040] After completing the two-branch feature extraction, an adaptive lightweight fusion discrimination mechanism for multi-source heterogeneous features is proposed. Unlike traditional methods that directly concatenate or use a single feature input, this method first jointly models the deep representation features extracted from the temporal structure branch and the physical attribute features constructed from the statistical feature branch. The discrimination process of the model can be represented as follows: (11), in: This represents the deep feature vector extracted from the I / Q raw signal branch. This represents the feature representation extracted by the statistical feature branch. This indicates a feature concatenation operation. This is the weight matrix of the fusion layer. For bias terms, This represents the Sigmoid activation function.

[0041] The feature concatenation operation is used to fuse multi-source information, combining the detailed structural features of the signal extracted by the waveform branch with the statistical features extracted by the statistical feature branch to form a complete feature representation containing both temporal structural information and physical statistical information, thereby achieving high-precision recognition. The two types of features are complementary in information expression; the I / Q branch focuses on characterizing transient waveform structural features, while the statistical feature branch focuses on describing energy evolution and spectral distribution characteristics. The combination of the two can effectively improve the ability to distinguish signals under complex interference environments. The neural network structure diagram is shown below. Figure 1 As shown.

[0042] Most existing transient electromagnetic interference (EMI) detection devices primarily rely on amplitude threshold triggering. This invention, based on threshold triggering to acquire transient waveforms, proposes a method for identifying transient signals by synergistically fusing temporal structural features and physical statistical features. It constructs a dual-branch deep fusion network for electrostatic discharge (ESD) signal identification. By extracting temporal structural features and physical statistical features separately and achieving synergistic coupling within a low-dimensional discrimination space, it solves the problem that existing EMI detection devices can only achieve signal triggering and struggle to effectively distinguish different types of transient interference. Compared to simple feature splicing or single-branch classification methods, this invention, through the synergistic fusion mechanism of temporal structural features and physical statistical features, achieves effective coupling and enhanced discrimination of multi-source information within a low-dimensional mapping space. This improves the ability to distinguish ESD transient signals from other interference signals without significantly increasing model complexity, maintaining high recognition accuracy and real-time performance, especially under the limited computing resources of embedded platforms.

[0043] The constructed fusion model was evaluated on the test set. The overall classification accuracy of the model reached 97.32%. Specifically, for the interference class, the precision was 94.74% and the recall was 94.74%; for the ESD class, the precision was 98.20% and the recall was 98.20%. In terms of overall discriminative ability, the area under the receiver operating characteristic (ROC) curve of this model reached 0.9930. Figure 2 As shown, the model exhibits excellent classification ability and stability under different decision thresholds. These results demonstrate that the method proposed in this invention, based on the fusion of I / Q data and multidimensional statistical features, can effectively improve the ability to distinguish between ESD signals and interference signals, possessing high recognition accuracy and significant engineering application value.

[0044] Building upon the methods described above, this example further presents an embedded system implementation scheme for engineering applications. This system, centered on an embedded computing platform and combined with software-defined wireless devices and a display terminal, forms an independently operable electrostatic discharge detection and identification device. The entire system consists of a signal acquisition module, a data processing module, and a result visualization module. These modules are interconnected via a high-speed data interface, thus forming a complete closed-loop detection and identification process, such as... Figure 3 As shown.

[0045] In terms of hardware implementation, the system uses a Raspberry Pi as the embedded processing core, coupled with a software-defined wireless device as the RF acquisition front end, and receives transient electromagnetic signals from the environment via an external broadband antenna. The software-defined wireless device performs down-conversion and analog-to-digital conversion of the RF signals, outputting complex baseband I / Q data, which is transmitted to the Raspberry Pi platform via a USB interface. An embedded display screen is connected to the Raspberry Pi to realize real-time visualization of the detection process and recognition results, thus forming an integrated hardware system structure of "acquisition-processing-display".

[0046] In terms of software implementation, the system builds a real-time signal processing program based on an embedded Linux operating system, designing a lightweight processing flow around continuous data streams. Addressing the characteristics of electrostatic discharge signals—highly sudden and short-duration—the system implements a burst detection mechanism based on circular buffering and threshold decision-making on the embedded side. Combined with an adaptive waveform processing method, it dynamically adjusts the waveform capture window and processing parameters according to changes in signal amplitude, achieving stable capture and structured representation of transient signals. Based on this, waveform standardization and key feature extraction are completed, and the processing results are input into a locally deployed lightweight recognition model for real-time inference.

[0047] In terms of model deployment, this invention directly deploys a lightweight dual-branch feature fusion network on the Raspberry Pi platform, enabling it to achieve fast response and stable recognition under limited computing power conditions, thereby avoiding reliance on cloud computing and improving the system's real-time performance and independent operation capability.

[0048] Through the aforementioned hardware and software co-design, compared to existing portable detection devices that only have event indication functions, this system can not only achieve real-time capture of transient signals, but also identify signal types on the embedded side. Furthermore, by introducing adaptive waveform processing and a lightweight model deployment mechanism, the detection accuracy and environmental adaptability are effectively improved while ensuring system portability and low cost.

Claims

1. A method for detecting transient electrostatic discharge signals based on software-defined radio, characterized in that, Includes the following steps: Step 1: Acquisition, digitization, and buffering of electrostatic discharge transient signals: Electromagnetic radiation signals in the 5MHz-25MHz frequency band are acquired, and through down-conversion and analog-to-digital conversion, a complex baseband signal is generated. Its discrete expression is as follows: (1), where and They represent the in-phase component and the quadrature component, respectively, and the continuously sampled data is stored in real time based on a ring buffer structure; Step 2: Preprocess the acquired continuous signal, including first performing amplitude square operation on the I / Q signal in the buffer to obtain the energy sequence signal characterizing the change of electromagnetic field intensity. In order to suppress random noise fluctuations, the energy sequence signal is subjected to moving average filtering. A threshold decision mechanism is used to complete the detection trigger. That is, by setting a threshold by preset or adaptive setting, when the threshold is met, it is determined that there is a potential transient event at that position, and the initial trigger position is recorded. Step 3: Perform waveform window truncation and standardization on burst signals that meet the conditions: After a burst signal is detected, extract a data window of fixed length centered on the trigger position, align the labeled data, and uniformly truncate the length of the output data to ensure that the length of the burst signal segment output each time is consistent, thus forming standardized data; Step 4: Feed the standardized data into the dual-branch recognition model for recognition processing: The first branch is the I / Q raw signal feature extraction branch, which adopts a progressive convolutional modeling method to extract local mutations and global structural features from the temporal dimension; the second branch is the statistical feature learning branch, which is used to calculate the rising edge slope, multi-scale energy proportion and spectral centroid, and performs nonlinear mapping through fully connected layers to achieve feature compression and recombination, and obtain a feature representation with strong discriminative ability; Step 5: Concatenate the depth feature vector output from the first branch with the statistical feature vector output from the second branch. After processing by the fusion discriminant layer, output the probability value of the signal belonging to the electrostatic discharge category through the Sigmoid activation function to complete signal recognition.

2. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, In step 2, the amplitude squared operation is performed on the I / Q signal to characterize the change in electromagnetic field intensity, and its expression is: (2) To reduce the impact of noise fluctuations on the detection results, a moving average filter is introduced to smooth the energy sequence. The calculation method is as follows: （3）， in, Indicates the first n The smoothed energy value at each location after moving average. Indicates the current point n Counting forward k The energy value of each sampling point Indicates the length of the sliding window. .

3. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, In step 3, during waveform truncation, the raw data window containing the trigger region is first extracted from the annular buffer based on the output of the burst detection module. The initial window length is set to 100 sampling points to cover the complete transient response process. An adaptive trigger point relocation mechanism based on amplitude constraints is introduced within the window. After relocation, 64 points are symmetrically truncated around the relocated trigger point as the model input. The amplitude of the I / Q data within the window is then calculated. （4）， And according to the preset threshold Select the set of valid sample points that meet the criteria. : （5）， Based on this, the geometric center of the set is used as the new reference trigger position. : （6）， in Indicates the size of the set. This indicates index summation, enabling adaptive correction of the initial trigger point and eliminating time offset issues caused by threshold detection errors or noise fluctuations.

4. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, The standardization process described in step 3 specifically involves: performing amplitude normalization on the truncated waveform data, assuming the original waveform sequence is... , in and Let represent the minimum and maximum values ​​in the sequence, respectively. Then the normalized data... It can be represented as: （7）。 5. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, In step 4 of the dual-branch recognition model, the I / Q original signal branches include: Local transient feature extraction unit: A lightweight one-dimensional convolutional structure with a kernel size of 5 and 32 channels is used to model the signal locally. A depthwise separable convolution method is used to capture amplitude abrupt changes and short time scale features of rising edges. Multi-scale temporal feature extraction unit: It adopts a convolutional structure with a kernel size of 5 and a channel number of 64. It extracts signal features in different time ranges through multi-scale convolutional kernels and uses a depth-separable convolutional structure to achieve multi-scale modeling, which improves the ability to characterize complex transient patterns while maintaining temporal resolution. High-dimensional semantic feature extraction unit: It adopts a convolutional structure with a kernel size of 3 and 128 channels to perform high-level abstraction and compression of the previous features. It generates discriminative feature embeddings by combining depthwise separable convolution with global feature aggregation, thereby realizing the expression of high-level semantic information.

6. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, In step 4 of the dual-branch recognition model, the statistical feature learning branch introduces a multi-dimensional feature modeling method based on the ESD discharge mechanism, which extracts the pulse envelope. To pinpoint the key time points within the 10% to 90% range of its peak value. and And calculate the slope of the rising edge accordingly. : （8）， in: The amplitude at which the signal amplitude reaches 10% of its peak value. The amplitude at which the signal amplitude reaches 90% of its peak value. The point in time when the amplitude first reaches a 10% peak value during the signal's ascent. The point in time when the amplitude first reaches 90% of its peak value during the signal's ascent. To characterize the distribution characteristics of signal energy across time scales, a multi-scale energy ratio method is used to calculate the energy proportion of the signal within different time windows. : (9), of which The complete transient electromagnetic signal throughout the entire time window. Signals within a sub-window at a certain time. Signal energy Sum all sampling points within the time window. The spectral centroid is used to characterize the central tendency of the spectral energy distribution of an electrostatic discharge signal. Obtained from the following formula: (10), where: :frequency, Corresponding frequency f The spectral amplitude or power spectral density at that location. Weighted frequency Total spectral energy.

7. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, In step 5, after completing the two-branch feature extraction, the deep representation features extracted by the temporal structure branch and the physical attribute features constructed by the statistical feature branch are first jointly modeled. The discrimination process of the model is represented as follows: （11）， in: This represents the deep feature vector extracted from the I / Q raw signal branch. This represents the feature representation extracted by the statistical feature branch. This indicates a feature concatenation operation. This is the weight matrix of the fusion layer. For bias terms, This represents the Sigmoid activation function. The feature concatenation operation is used to fuse multi-source information, combining the detailed structural features of the signal extracted by the waveform branch with the statistical features of the signal extracted by the statistical feature branch to form a complete feature representation containing both temporal structural information and physical statistical information, thereby achieving high-precision recognition. The two types of features are complementary in information expression. The I / Q branch focuses on characterizing the transient waveform structural features, while the statistical feature branch focuses on describing energy evolution and spectral distribution characteristics. The combination of the two enhances the ability to distinguish signals under complex interference environments.

8. The method for detecting transient electrostatic discharge signals based on software radio according to claim 1, characterized in that, Before training the dual-branch recognition model, a training dataset was constructed: electrostatic discharge signal samples were collected using an electrostatic discharge generator conforming to the IEC 61000-4-2 standard under different voltage levels, test distances and environmental conditions; interference signal samples generated by power strip switches, light switches or lighter switches were also collected. Data augmentation is performed on the I / Q dual-channel signal, including at least one of amplitude scaling, phase rotation, noise perturbation, and time jitter, to generate augmented samples.

9. A device for detecting transient electrostatic discharge signals based on software-defined radio, characterized in that, Used to perform the software-defined radio-based electrostatic discharge transient signal detection method as described in any one of claims 1-8; It is equipped with a signal acquisition module, a data processing and recognition module, and a result output module. The signal acquisition module is equipped with a broadband antenna and an SDR receiver. The signal acquisition module sends the acquired I / Q data to the data processing and recognition module. The data processing and recognition module is equipped with an embedded processing platform, a burst detection and waveform truncation processing unit, a dual-branch recognition model feature extraction model, and an ESD recognition model.

10. The electrostatic discharge transient signal detection device based on software radio according to claim 9, characterized in that, The signal acquisition module is used to receive electromagnetic radiation signals in the 5 MHz to 25 MHz frequency band through a broadband antenna, and then down-convert them into complex baseband I / Q signals through a software-defined wireless device. The embedded processing platform is connected to the software-defined wireless device via a USB interface and runs the above-mentioned detection method to achieve burst signal triggering, waveform standardization, feature extraction, and signal recognition. It also has a display module connected to the embedded processing module to display the detection results and recognition probability in real time. The embedded processing module adopts a Raspberry Pi platform, and the dual-branch recognition model is deployed locally in a lightweight form, without the need for cloud computing.

Images