Edge AI automatic judgment method and device for HPM damage level

By deploying edge computing devices outside the shielded cabin and utilizing a conformal fiber electric field sensor array and plastic fiber transmission links, multi-dimensional sensing data can be acquired and processed in real time. A pre-trained damage level determination model is adopted to solve the problems of low efficiency and poor consistency in HPM damage level determination, thus achieving efficient and real-time automatic damage level determination.

CN121542929AActive Publication Date: 2026-02-17NAVAL AVIATION UNIV

Patent Information

Application Number
CN202610062898.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-19
Publication Date
2026-02-17
Estimated Expiration
2046-01-19

AI Technical Summary

Technical Problem

In the existing technology, the determination efficiency of HPM damage level is low and real-time output cannot be achieved, which makes it difficult to adjust the launch parameters and test body attitude in a timely manner based on the determination results during the test, and the consistency and objectivity of the determination results are poor.

Method used

By deploying edge computing devices outside the shielded cabin, and utilizing conformal fiber electric field sensor arrays and plastic fiber transmission links, multi-dimensional sensing data can be acquired and processed in real time, and a pre-trained damage level determination model can be used for automatic determination.

Benefits of technology

It achieves real-time and efficient determination of HPM damage level, improves the consistency and objectivity of the determination results, and can conduct more than 8 test rounds per day, ensuring test efficiency and process closed-loop capability, while avoiding the influence of electromagnetic interference.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121542929A_ABST
    Figure CN121542929A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of high-power microwave effect tests and testing, and particularly provides an automatic edge AI judgment method and device for an HPM damage grade, and the method comprises the steps: synchronously obtaining electric field, temperature and FBG reference data through a conformal optical fiber sensor array and a reference sensor in a test body; and the optical fibers are reinjected to the outside of the shielding cabin in a zero-metal manner in a lossless manner through a plastic optical fiber-CAN-FD link. Voxelization, multi-modal fusion and INT8 quantization processing are carried out on the data on an outboard edge computing device, and a standard input tensor is generated. And a damage grade judgment model deployed on the equipment is used for reasoning, and a damage grade judgment result and confidence are output. Optionally, the model may also output a thermodynamic diagram identifying the critical space region. According to the method, the AI model is locally deployed through the edge computing device, multi-dimensional sensing data are fused, real-time efficient judgment of the HPM damage level is achieved, and meanwhile the consistency and objectivity of the judgment result are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high power microwave effect test and test technology, in particular to an edge AI automatic determination method and device for HPM damage level. BACKGROUND

[0002] In the field of high power microwave (HPM) effect test, accurate determination of the damage level of the test body is the core link of evaluating the effect of HPM. The determination efficiency, consistency and real-time affect the effectiveness of the test data and the closed-loop ability of the test process.

[0003] Related automatic determination schemes mostly adopt the "remote server offline processing" mode. In this mode, the electric field, temperature and other data collected by the sensor need to be sent to the remote server through a complex transmission link. Not only does the large amount of data and long transmission distance cause significant delay, but the server also needs to bear the data analysis, format conversion and calculation tasks of multiple links, resulting in a cumbersome overall processing process and a long time-consuming. The real-time output of the damage level cannot be realized, making it difficult to adjust the key conditions such as HPM emission parameters and test body posture in a timely manner during the test process, which cannot meet the core needs of closed-loop regulation of the test, and further limits the number of test rounds per day (usually ≤3 rounds), resulting in low test efficiency.

[0004] Part of the scene still relies on manual participation in data interpretation and level judgment. The operator needs to make subjective judgments on the damage level based on the original data, waveform diagram and other materials exported after the test, combined with their own experience. The time-consuming of a single evaluation is long. There are differences in the experience level and understanding of the judgment standard of different operators, resulting in low consistency of the evaluation results of the same group of test data, making it difficult to ensure the objectivity and reliability of the test data. SUMMARY

[0005] To solve the above problems, the present application provides an edge AI automatic determination method and device for HPM damage level, which realizes real-time and efficient determination of HPM damage level by locally deploying AI models on edge computing devices and fusing multi-dimensional sensor data, while significantly improving the consistency and objectivity of the determination results.

[0006] In the first aspect, the technical solution of the present application provides an edge AI automatic determination method for HPM damage level, comprising the following steps: The electric field amplitude data inside the test body is obtained by the conformal optical fiber electric field sensor array built in the test body; at the same time, the temperature field data and fiber Bragg grating reference data inside the test body are obtained; The electric field amplitude data, temperature field data and fiber Bragg grating reference data are losslessly injected back to the outside of the shielding cabin in zero metal mode through the plastic optical fiber transmission link; On the edge computing device located outside the shielded chamber, the reinjected electric field amplitude data is processed to generate electric field volume data in the form of a three-dimensional voxel mesh; and the electric field volume data, temperature field data and fiber Bragg grating reference data are fused and formatted to generate model input data that conforms to the predetermined input specifications. The pre-trained damage level determination model deployed on the edge computing device performs inference based on the model input data and outputs a multi-level damage level determination result corresponding to the current HPM effect.

[0007] Secondly, the technical solution of the present invention provides an edge AI automatic determination device for HPM damage level, which implements the above-mentioned method, including: Conformal fiber electric field sensor array: Built inside the test body, used to acquire electric field amplitude data inside the test body; Temperature and reference data acquisition unit: Deployed inside or on the surface of the test body to acquire temperature field data and fiber Bragg grating reference data inside the test body; Plastic optical fiber transmission link: Its input end is connected to the conformal optical fiber electric field sensor array and the temperature and reference data acquisition unit respectively, and its output end passes through the shielded cabin to transmit electric field amplitude data, temperature field data and fiber Bragg grating reference data back to the outside of the shielded cabin in a zero-metal lossless manner. Edge computing device: Located outside the shielded cabin, connected to the output end of the plastic fiber optic transmission link, configured to perform the following operations: receive reinjected electric field amplitude data, temperature field data, and fiber Bragg grating reference data; process the electric field amplitude data to generate electric field volume data in the form of a three-dimensional voxel mesh; fuse and format the electric field volume data, temperature field data, and fiber Bragg grating reference data to generate model input data conforming to a predetermined input specification; call the pre-trained damage level determination model deployed inside it, perform inference based on the model input data, and output a multi-level damage level determination result corresponding to the current HPM effect.

[0008] As can be seen from the above technical solutions, this application has the following advantages: (1) Edge computing devices are deployed near the shielded cabin and have a pre-trained damage level determination AI model built in. Data processing and inference are completed locally on the device, eliminating the need for long-distance data transmission. The edge device has a rapid local computing response and the determination delay can be controlled within milliseconds, enabling real-time output of damage levels. During the test, the HPM launch parameters and test subject attitude can be adjusted in a timely manner based on the determination results. The number of test rounds per day can be increased to more than 8 rounds, greatly improving test efficiency and process closure capability. (2) By integrating multi-dimensional sensor data such as electric field, temperature, and test body reference parameters, the AI ​​model trained by deep learning is used for automatic reasoning and judgment. The AI ​​model can accurately capture the multi-level damage characteristics under complex HPM field, replace subjective manual judgment and simple threshold algorithm, improve the consistency of judgment results, and significantly enhance the objectivity and comparability of data. (3) In addition, a conformal fiber electric field sensor array built into the test body is used, along with a plastic fiber transmission link, to achieve zero-metal data transmission. The fiber sensing and transmission are not affected by electromagnetic interference, ensuring the shielding effectiveness of the shielding chamber while ensuring the purity and reliability of data acquisition such as electric field amplitude and temperature field. The conformal design can fit closely to the test body, adapting to test bodies with different structures, and the plastic fiber is easy to deploy without damaging the original structure of the test body. Attached Figure Description

[0009] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an edge AI automatic determination method for HPM damage level provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic block diagram of an edge AI automatic determination device for HPM damage level provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0014] Figure 1 This is a schematic diagram of an AI-based automatic edge determination method for HPM damage level provided in an embodiment of the present invention, as shown below. Figure 1As shown, the method includes the following steps.

[0015] S1 acquires electric field amplitude data inside the test body through a conformal fiber electric field sensor array built into the test body; at the same time, it acquires temperature field data and fiber Bragg grating reference data inside the test body.

[0016] S2 transmits electric field amplitude data, temperature field data, and fiber Bragg grating reference data back to the outside of the shielded cabin in a zero-metal, lossless manner via a plastic optical fiber transmission link.

[0017] S3, on the edge computing device located outside the shielded chamber, processes the reinjected electric field amplitude data to generate electric field volume data in the form of a three-dimensional voxel mesh; and fuses and formats the electric field volume data, temperature field data and fiber Bragg grating reference data to generate model input data that conforms to the predetermined input specifications.

[0018] S4, based on the input data of the pre-trained damage level determination model deployed on the edge computing device, performs inference and outputs a multi-level damage level determination result corresponding to the current HPM effect.

[0019] As a refinement and extension of the specific implementation of the above embodiments, in order to fully explain the specific implementation process of this embodiment, the following will provide possible embodiments to describe the specific implementation of the above steps in a non-limiting manner.

[0020] In this embodiment, a conformal fiber electric field sensor array embedded within the test specimen acquires the electric field amplitude data inside the specimen. This array consists of multiple electro-optic modulation sensor units based on lithium niobate (LiNbO3) thin-film microstrip structures, tightly bonded to the inner curved surface of the test specimen in a helical chain arrangement. The sensors are mounted using peelable thermally conductive adhesive, enabling rapid assembly and disassembly with zero metal and zero opening of the compartment. The array covers key areas of the test specimen, such as the antenna coupling area of ​​the first compartment, the electronic device area of ​​the middle compartment, and the fuse area of ​​the stern compartment, with a spatial measurement point density of no less than 20 points / meter. For a typical 3-meter-long test specimen, the total number of measurement points is approximately 60, achieving distributed continuous imaging of the internal electric field, rather than traditional discrete point measurements.

[0021] Based on the linear electro-optic effect (Pockels effect), an external high-power microwave electric field penetrates the test chamber wall and acts on the LiNbO3 thin film, causing a linear change in its refractive index, which in turn modulates the phase of the probe light passing through the film. This phase change is demodulated into a light intensity signal (S) using an interferometer. 11The amplitude of the electric field is proportional to the strength of the external electric field. This sensor features high sensitivity (≥0.4mV / (V·m)), wide bandwidth (DC-18GHz), fast response (rise time <1ns), and a large range (>100kV / m). Each sensor unit outputs a real-time analog or digital signal proportional to the local electric field amplitude, i.e., "electric field amplitude data," accompanied by a unique sensor ID.

[0022] In this embodiment, to eliminate the influence of ambient temperature changes on the readings of the electric field sensor, temperature field data and fiber Bragg grating reference data are acquired simultaneously to form a dual-redundant temperature compensation.

[0023] Temperature field data is acquired through distributed or array-type temperature sensors to provide a general temperature distribution within the test specimen with a certain spatial resolution.

[0024] To obtain reference data for fiber Bragg gratings (FBGs), a fiber Bragg grating is inscribed on the same polarization-maintaining fiber (PM1550) of the conformal fiber electric field sensor array. FBGs are extremely sensitive to temperature; the central Bragg wavelength of their reflections drifts linearly with temperature (typical sensitivity approximately 10 pm / ℃). By demodulating this wavelength, the precise temperature at the sensor's location can be obtained. Because the FBG and the electric field sensor are located in the same physical location and thermal environment, the temperature reference it provides allows for the most direct and accurate decoupling correction of the electric field readings, reducing the uncertainty introduced by temperature drift by more than 60%, ensuring that the system sensitivity drift is less than 1.5% after 120 rounds of repeated testing.

[0025] All sensors for electric field, temperature, and FBG are synchronously sampled under the control of a unified White-Rabbit clock system. Each data sampling point carries a global timestamp with an accuracy better than 1 nanosecond, enabling subsequent data from different locations and types of sensors to achieve strict alignment in the time domain with sub-nanosecond accuracy, thereby improving the accuracy of multimodal data fusion and the real-time performance of analysis.

[0026] In this embodiment, step S2, data transmission overcomes the shortcomings of traditional metal cables passing through walls. It employs a rapidly deployable hybrid communication link designed for high-power microwave compact environments, enabling lossless real-time data transmission from inside the shielded enclosure to the outside. The plastic fiber optic transmission link in this embodiment is not a simple media replacement, but rather an adaptation and integration of the plastic fiber optic physical layer with the flexible data rate protocol of the controller area network.

[0027] Specifically, the physical layer uses 1mm diameter polymethyl methacrylate (PMMA) plastic optical fiber. Compared to glass fiber (125μm), it has higher bending resistance and mechanical strength, can pass through the narrow center aperture of the compact field turntable without precise alignment, and is inexpensive. The operating wavelength is 650nm, using a standard low-cost HFBR series optoelectronic converter.

[0028] The protocol layer adopts the CAN-FD protocol, which conforms to the ISO11898-1:2015 standard for CAN-FD bus protocols. This protocol, inherited from the automotive electronics field, features high reliability, real-time performance, and strong electromagnetic interference resistance. It supports higher data rates and longer data frames, sufficient to carry multi-channel sensor data. In this embodiment, the link bandwidth is configured to 1 Mbps, stably supporting 48 sensor channels operating simultaneously at a 1 kHz sampling rate.

[0029] The critical connection point where the link passes through the shielded enclosure uses an all-plastic SMA-905 fiber optic connector. This connector, with an outer diameter not exceeding 3mm, is fixed to a pre-fabricated penetration hole in the shielded enclosure using a rubber O-ring compression seal and a grounding spring (used only for grounding the connector housing, not for signal conduction). The entire penetration path contains no metal conductors, physically eliminating any path for induced current introduction. Electro-optical / optical-electrical conversion modules are deployed on both the inner and outer sides of the shielded enclosure. The electrical signal generated by the sensor is converted into a 650nm optical pulse inside the enclosure, transmitted through the plastic optical fiber to the outside, and then converted back into an electrical signal for use by the edge computing device. This process achieves complete electrical isolation, with an isolation voltage exceeding 2.5kV, ensuring no breakdown even under extremely strong HPM fields (>100kV / m). Actual measurements show that the link's bit error rate is less than 10% over a 50-meter transmission distance. -9 This meets laboratory-level data integrity requirements. The transmission delay of optical signals in optical fibers is extremely small; combined with protocol processing, the end-to-end system transmission delay is strictly controlled to within 200 milliseconds, meeting real-time requirements.

[0030] The plastic fiber-CAN-FD hybrid link used in this embodiment achieves zero-metal, high-isolation, and lossless data reinjection, providing reliable source data for subsequent AI model inference and improving the accuracy and reliability of automatic HPM damage level determination.

[0031] This embodiment converts the raw, discrete sensor data injected via a plastic fiber optic link into a structured, dense three-dimensional spatial representation in real time on an edge computing device outside the shielded cabin, providing a standardized input format for subsequent AI inference.

[0032] Edge computing devices are preferably based on NVIDIA Jetson Orin series embedded AI platforms. This platform integrates a multi-core ARM CPU and a GPU with 2048 CUDA cores, providing parallel computing capabilities with a power consumption of less than 30W. Its unified memory architecture allows the CPU and GPU to share the same physical memory, facilitating zero-copy data transfer.

[0033] In step S3, the injected electric field amplitude data is processed to generate electric field volume data in the form of a three-dimensional voxel grid, specifically including the following steps S3.1 to S3.3.

[0034] S3.1, Data reception and parsing.

[0035] Edge computing devices receive serial data streams from plastic fiber optic links via their integrated CAN-FD controller. Each data frame contains a sensor ID, a timestamp (nanosecond accuracy), and the raw electric field amplitude. The electric field amplitude is either voltage or a digital quantity converted from photoelectric values. First, based on preset sensor calibration parameters, the raw amplitude data is linearly converted into a spatial point electric field strength value in V / m.

[0036] S3.2, Coordinate Mapping and Point Cloud Generation.

[0037] A sensor location lookup table is maintained, obtained through calibration before the experiment, recording the three-dimensional Cartesian coordinates (x, y, z) of each sensor ID within the experimental body. For each frame of data, the electric field intensity value is placed onto the corresponding three-dimensional coordinates according to the sensor ID, thereby generating a sparse three-dimensional point cloud in real time. The density of this point cloud depends on the layout of the sensor array, and is no less than 20 points / meter.

[0038] S3.3, Parallel voxelization processing.

[0039] This step converts the sparse point cloud into a dense voxel mesh, which is then executed on the GPU by calling the CUDA parallel computing kernel function.

[0040] A 256×256×256 resolution three-dimensional array is pre-allocated in memory, corresponding to a cubic grid that uniformly divides the entire interior space of the experimental subject. Each grid cell is called a "voxel". The CUDA kernel function starts several threads to work in parallel. Each thread is responsible for processing one sensor point, calculating which voxel grid it falls into based on its (x,y,z) coordinates, i.e., determining the voxel index (i,j,k). Then, the electric field intensity value of that point is assigned to the corresponding voxel. For multiple points falling into the same voxel, strategies such as averaging or taking the maximum value can be used. For voxels without sensor points, trilinear interpolation can be performed based on the values ​​of its neighboring voxels to generate a spatially continuous and dense three-dimensional electric field intensity distribution field, i.e., electric field volume data.

[0041] This embodiment deploys a high-performance edge AI computing device outside the shielded cabin and utilizes a CUDA-accelerated parallel voxelization algorithm to reconstruct the reinjected discrete sensor data into a high-resolution three-dimensional spatial electric field distribution in real time. This enables the conversion of physical measurement data into standardized features that can be understood by deep learning models and contain spatial structural information.

[0042] In step S3, the electric field data, temperature field data, and fiber Bragg grating reference data are fused and formatted to generate model input data that conforms to the predetermined input specifications. Specifically, this includes the following steps S3.4 to S3.8.

[0043] S3.4, based on a unified time reference, align the electric field data, temperature field data, and fiber Bragg grating reference data with timestamps and register their spatial coordinates.

[0044] At the moment of sampling, the analog-to-digital converter at the sensor end directly reads a 64-bit time counter (CycleCounter) from the White-Rabbit Slave hardware counter directly connected to it, which serves as the native hardware timestamp for that data point. The synchronization error between this timestamp and the White-Rabbit master clock is less than 1 ns.

[0045] Edge computing devices maintain a global time base, also a WR slave clock. After receiving data, based on the hardware timestamp attached to each data packet, a nearest neighbor matching algorithm is used: within a predefined time fault-tolerant window, all sensor data are sorted by timestamp, and the set of electric field, temperature, and FBG data with the closest timestamps is determined to be the dataset of "the same moment", packaged and entered into the subsequent processing flow to achieve time alignment.

[0046] Before system deployment, the three-dimensional coordinates (X,Y,Z) of each conformal fiber electric field sensor, temperature sensor, and FBG sensing unit in the test body coordinate system are measured using a 3D laser scanner or photogrammetry, generating a "sensor ID-spatial coordinate" lookup table, which is stored in the memory of the edge computing device.

[0047] When processing data, the lookup table is queried in real time based on the sensor ID in the data packet to obtain its corresponding 3D coordinates. For electric field point cloud data, these coordinates are directly used for voxelization. For temperature data and FBG data, these coordinates are used to determine which voxel indices in the 256³ voxel grid they should be associated with in the subsequent channelization step.

[0048] S3.5 Upsamples the temperature field data to the same spatial resolution as the electric field volume data to generate a temperature data channel.

[0049] The temperature field data comes from a sparse temperature sensor array, with each data point containing a sensor ID (corresponding spatial coordinates) and a temperature value. Typically, the spatial density of temperature sensors is much lower than that of electric field sensors. Therefore, a trilinear interpolation algorithm is used. First, the sparse temperature point cloud is placed into a low-resolution 3D grid (e.g., 32) according to its coordinates. 3 Then, for the high-resolution target mesh (256) 3 For each voxel in the data, the coordinates of its eight nearest neighboring grid points in the low-resolution grid are calculated. Then, based on the temperature values ​​of these eight grid points, a linear weighted average is applied according to distance to calculate the temperature value of the high-resolution voxel. Finally, a high-resolution voxel with the same resolution (256) as the electric field volume data is generated. 3 (a dense temperature data channel).

[0050] S3.6, Based on the predetermined mapping relationship between the sensing position and the test body space, the one-dimensional fiber Bragg grating reference data is assigned to all corresponding voxels in the three-dimensional voxel grid to generate a reference data channel with the same spatial resolution as the electric field volume data.

[0051] The fiber Bragg grating reference data is a one-dimensional scalar data sequence from multiple FBG sensors. During the calibration phase, each FBG sensor not only had its coordinates measured, but its sensitive region was also experimentally determined to correspond to the voxel index range in the voxel grid. For example, the voxel index range corresponding to FBG_1 is x:[100,150], y:[50,100], z:[80,120].

[0052] A zero-matrix of the same dimension as the electric field volume data is created as the reference data channel. For each FBG sensor, its current measurement value is directly assigned to all elements within its predefined voxel index range in the reference data channel matrix. This operation is executed in parallel on the GPU using CUDA kernel functions, with each thread handling one voxel index range. Ultimately, a spatially uniformly partitioned reference data channel is generated.

[0053] S3.7, the electric field data, temperature data channel and reference data channel are spliced ​​together in the channel dimension to form multi-channel fused data.

[0054] The three processed data points—electric field data, temperature data channel, and baseline data channel—are treated as three independent layers. They are then combined along the channel dimension to form a single, three-channel four-dimensional tensor. This tensor has dimensions [C,D,H,W]=[3,256,256,256], where: Channel 0: Stores the normalized electric field intensity distribution; Channel 1: Stores the temperature distribution after upsampling; Channel 2: Stores the space compensation baseline values ​​derived from the FBG baseline.

[0055] S3.8 performs INT8 fixed-point quantization on the multi-channel fused data to generate a multi-channel three-dimensional data tensor that conforms to the predetermined input specifications. This tensor is the model input data.

[0056] This step converts the fused data from high-precision floating-point numbers (such as FP32) to low-precision integers (INT8), significantly reducing computational and memory access overhead to achieve low latency at the edge.

[0057] The INT8 fixed-point quantization process uses quantization parameters determined through a calibration dataset during the training phase of the damage severity assessment model. The calibration dataset consists of real data from 120 rounds of HPM trials. The quantization formula is:

[0058] in, The value in the FP32 tensor. The scaling factor is calculated based on the statistical distribution of the calibration data. The zero point is used, and `clamp` is the truncation function to ensure that the result is within the INT8 range. This operation is completed by traversing the tensor's CUDA kernel function once, and the output is the final INT8 type multi-channel 3D data tensor, which is used as the model input.

[0059] In this embodiment, the damage level determination model is a lightweight neural network quantized with INT8, used to perform real-time, automatic determination of high-power microwave damage levels on edge computing devices. Its overall architecture includes an input layer, a backbone network, a feature fusion layer, a classification head, and an output layer.

[0060] Input layer: Directly receives model input data, or generates model input data by executing steps S3.4 to S3.8.

[0061] Backbone Network: Employs a hybrid architecture combining 3D Convolutional Neural Network (3D-CNN) and Time-Series Transformer to output high-dimensional feature maps that integrate spatiotemporal characteristics. By modeling the spatial distribution and temporal evolution of the HPM effect, the comprehensiveness and accuracy of the judgment are improved.

[0062] The 3D-CNN part is responsible for processing from 256 3 Three-dimensional spatial features are extracted from voxel electric field data. By performing multi-layer three-dimensional convolution and pooling operations, the input data is downsampled by a factor of 8 to gradually abstract the spatial distribution pattern, gradient changes, and local hotspot features of the electric field within the experimental body.

[0063] The Time-Series Transformer is responsible for modeling time-series features. It processes feature sequences from consecutive sampling times and captures the dynamic process and temporal dependencies of damage evolution through a self-attention mechanism, thereby achieving a temporal understanding of the damage process.

[0064] Feature fusion layer: This layer includes a global average pooling layer and a fully connected layer. The spatially dimensional feature maps extracted by the backbone network are compressed into fixed-length one-dimensional feature vectors through global average pooling. Subsequently, the fully connected layer performs non-linear transformation and dimensionality reduction on these feature vectors, further fusing and refining the information.

[0065] Classification Head: Contains a Softmax classifier. It maps the feature vectors output by the feature fusion layer to a 5-dimensional vector, with each dimension corresponding to a predefined damage level (Class 0 to Class 4) probability.

[0066] Output layer: Outputs a comprehensive assessment result, including damage level and confidence level. The output damage level is a 5-level classification result from Class 0 (no damage) to Class 4 (physical damage), presented in the form of a probability distribution. The output confidence level is a value ranging from 0% to 100%, representing the model's degree of confidence in the current damage level assessment result.

[0067] Based on the above damage level determination model, step S4 involves reasoning based on the model input data, specifically including the following steps S4.1 to S4.5.

[0068] S4.1, receive model input data directly through the input layer, or have the input layer generate model input data by executing steps S3.4 to S3.8.

[0069] The input layer is the interface between the model and external data, and it has two adaptation modes.

[0070] Direct receiving mode: When preprocessing is done by a separate program, the input layer only serves as a data entry point, receiving the prepared INT8 type multi-channel three-dimensional data tensor [3,256,256,256].

[0071] Integrated processing mode: The input layer can also be designed to include a preprocessing unit. In this mode, the input layer directly receives the raw data that has been time-aligned but not fully fused, namely electric field data, temperature data, and FBG baseline data, and has a built-in lightweight CUDA kernel function to sequentially perform channel splicing (step S3.7) and INT8 quantization (step S3.8) to dynamically generate standard model input tensors in memory.

[0072] S4.2 Extracts the spatial distribution features of the model input data through the 3D convolutional neural network in the backbone network, and extracts the temporal dynamic features of the model input data at continuous time steps through the time series transformer. Then, the spatial distribution features and the temporal dynamic features are fused to obtain an intermediate feature map with fused spatiotemporal features.

[0073] Specifically, the spatial feature extraction branch consists of multiple 3D convolutional layers, 3D batch normalization layers, and ReLU activation functions stacked alternately, interspersed with 3D max pooling layers for downsampling. This branch primarily processes the 0th channel of the input tensor (electric field volume data), as the electric field data contains the richest spatial structure information. Simultaneously, the 1st and 2nd channels (temperature, FBG) serve as auxiliary inputs, implicitly fused through multi-channel convolutional kernels of 3D convolution. Their information is used to modulate the electric field feature extraction process, such as enhancing or suppressing certain spatial patterns. This branch outputs a high-dimensional spatial feature map, encoding key spatial patterns such as the three-dimensional distribution, gradient, and local hotspots of the electric field within the experimental volume.

[0074] The temporal dynamic feature extraction branch processes a time series segment. This segment is formed by stacking the model input tensors of the current time step and the previous N consecutive time steps in the time dimension, forming a five-dimensional sequence [N,3,256,256,256].

[0075] First, a lightweight 3D CNN is used to initially encode the spatiotemporal features of the data at each time step, compressing its spatial dimension. Then, the sequence is input into a Transformer encoder. The Transformer, through its self-attention mechanism, calculates the correlation between features at different time steps in the sequence, thereby modeling the dynamic evolution of the damage effect, such as the establishment, decay, and resonant frequency drift of the field strength, and outputs a feature vector that incorporates temporal context information.

[0076] The outputs of the two branches are concatenated along the feature dimension. Specifically, the spatial feature map output by the 3D-CNN branch is flattened into a feature vector, and concatenated with the temporal feature vector output by the Transformer branch along dimension 1 to form a joint feature vector that integrates spatial structure and temporal evolution, which serves as the input for subsequent processing. This fusion is performed at the end of the backbone network.

[0077] S4.3 compresses the intermediate feature map that integrates spatiotemporal features into a one-dimensional feature vector through a global average pooling layer, and then performs nonlinear transformation and dimensionality reduction through a fully connected layer to generate a feature vector that represents global information.

[0078] Global average pooling receives the fused feature map output from the backbone network. This layer calculates the average of the feature values ​​at all spatial locations in each channel of the feature map, thereby compressing a feature map of arbitrary spatial size into a fixed-length one-dimensional global feature vector.

[0079] The global feature vector is input into one or more fully connected layers. Each fully connected layer consists of a linear transformation (multiplication of weight matrices) and a nonlinear activation function to perform nonlinear transformation of high-dimensional features, further information fusion and dimensionality reduction, and finally output a global feature representation as the input of the classification head.

[0080] S4.4 Input the feature vector representing global information into the classification head. The classifier outputs a multi-dimensional vector, where each dimension corresponds to the probability of a predefined damage level.

[0081] Specifically, the classifier is a Softmax classifier that receives a one-dimensional global feature vector from the feature fusion layer. First, a linear layer maps the feature vector to a 5-dimensional logistic vector. Then, the Softmax function is applied to normalize the logistic vector, as shown below:

[0082] in, For logical values, The output probabilities are given. The Softmax function ensures that the sum of all output probabilities is 1, and that each... The model represents the current sample belonging to the first... The posterior probability of each damage level.

[0083] S4.5 The output layer takes the level corresponding to the dimension with the highest probability value as the damage level determination result, and outputs the highest probability value as the determination confidence level.

[0084] The output layer outputs a 5-dimensional probability vector from the Softmax layer. Perform the argmax operation to find the index with the highest probability value; this index represents the final damage level determined by the model. The value of (i.e., the maximum probability value) is directly used as the confidence level output for this decision. This value ranges from 0 to 1 (or is expressed as a percentage), intuitively reflecting the model's confidence in the decision result. The output layer outputs (damage level, confidence level) as the decision result.

[0085] In some optional implementations, the damage level determination model is optimized during the training phase using a composite loss function of adaptive weighted focal loss, which is a linear combination of a weighted focal loss term and a critical interval penalty term. The total loss function is expressed as:

[0086] in, These are hyperparameters used to balance the contributions of the two losses. For the weighted focus loss item, This is a penalty term for the critical interval.

[0087] The weighted focus loss term addresses the imbalance in the number of samples with different damage levels in the training data and enhances the model's ability to distinguish levels with low historical classification accuracy, i.e., levels that are difficult to differentiate, through dynamic weights. The critical interval penalty term is designed to meet the safety requirements of HPM damage level determination. It applies an exponentially increasing penalty to cases where the predicted level deviates significantly from the actual level, effectively suppressing severe misclassifications such as classifying severe damage as no damage. By jointly optimizing the above loss functions, the model can significantly improve the discrimination of critical damage levels while maintaining high overall classification accuracy and reducing the probability of catastrophic misclassifications, thus meeting the requirements of high consistency and high reliability of determination results in laboratory HPM effect experiments.

[0088] Weighted focus loss item Based on the standard focus loss, a ranking difficulty weight is introduced to address the problem of imbalanced samples and different classification difficulties in HPM experiments, expressed as:

[0089] in, Damage level, Number of damage levels For the true label of the sample, To predict the level of a sample in the model The probability, The class balancing weights are inversely proportional to the number of training samples for each level, used to alleviate class imbalance, such as when there are far more "no damage" samples than "physical damage" samples. This is a focusing parameter, which is greater than or equal to 0, used to reduce the number of easily classified samples. The loss contribution is close to 1), allowing the model to focus more on samples that are difficult to classify.

[0090] For the indicator function, when the real label Compared to the current level The indicator function is 1 if the values ​​are the same, and 0 otherwise. The indicator function ensures that only the values ​​corresponding to the true class appear in the entire loss summation term. That item will be calculated and accumulated, the others... The terms are ignored because the indicator function is 0.

[0091] For level The difficulty weights can be dynamically updated. For classes with low accuracy in this validation, the difficulty weight for that class will be increased in the next training round. This means that for classes with consistently low classification accuracy—classes that the model struggles to classify, such as the boundary between Class 1 soft failure and Class 2 hard failure—a higher difficulty weight will be assigned. This value forces the model to pay more attention to these difficulty levels during training.

[0092] Specifically, at the start of training, the difficulty weights for all levels are... Setting it to 1 means the model uses the current weights. Perform a training cycle, and at the end of the training cycle, evaluate the model on an independent validation set using the latest model weights, calculating the model at each damage level. Classification accuracy Then, based on the results of this validation, the difficulty weights for the next round of training are updated according to the predetermined strategy. The strategy is expressed as:

[0093] in, Is the model at the level The latest verified accuracy rate, It is a smoothing factor used to control the update magnitude. It is a very small number, to prevent division by zero.

[0094] This update strategy allows for adjusting the weights of levels with lower accuracy in the current validation round during the next training iteration. It will increase.

[0095] Critical interval penalty term Designed for HPM testing scenarios, this approach penalizes severe misjudgments that "leapfrog" while tolerating minor deviations at the "adjacent" level, conforming to the tolerance logic of expert evaluation. It is represented as follows:

[0096] in, For prediction level Compared to the actual level absolute distance, The distance penalty coefficient is a piecewise increasing function, which can be expressed as:

[0097] This loss term not only penalizes prediction errors but also amplifies them non-linearly according to the severity of the error. For example, misclassifying Class 4 (physical damage) as Class 0 (no damage) is extremely costly. The cost of misclassifying Class2 as Class1 is much lower. This directly guides the model learning to produce neighbor bias, and the judgment criterion for avoiding qualitative errors is consistent with the requirements for safety and reliability of conclusions in HPM effect experiments.

[0098] The damage rating model was trained using specially acquired high-quality HPM effect test data. The data came from over 120 rounds of laboratory HPM effect tests. In each round, a conformal fiber optic sensor array, temperature sensor, and FBG simultaneously acquired complete physical field data inside the test specimen. After each round, at least three domain experts independently interpreted the data based on post-test inspection results and complete multimodal data records to jointly determine the corresponding damage rating (Class 0 to Class 4). The final label was determined through consensus by an expert review panel, ensuring the label's authority and accuracy. This resulted in a model containing >10 4 A labeled dataset of valid samples. The total dataset is randomly divided into training, validation, and test sets in an approximately 7:2:1 ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and preventing overfitting, and the test set is used to finally evaluate the model's generalization performance.

[0099] To accelerate convergence and improve performance, the model employs a transfer learning strategy for initialization. The weights of the 3D-CNN backbone are initialized using a model pre-trained on large-scale public 3D medical image datasets or point cloud datasets, enabling the model to inherit general 3D feature extraction capabilities from massive datasets. The weights of the remaining parts, such as the Time-Series Transformer module, feature fusion layer, and classification head, are randomly initialized using standard methods such as Xavier or Kaiming.

[0100] The model's training objective is guided by a composite loss function designed for this task, which is minimized through an optimization algorithm. A linear combination of the aforementioned adaptive weighted focus loss and critical interval penalty loss is used as the total loss. The AdamW optimizer is employed, with an initial learning rate set to 1e-4, and a cosine annealing learning rate scheduling strategy is used to smoothly reduce the learning rate during training, helping the model converge to a better local minimum.

[0101] Training was conducted on a high-performance computing server equipped with multiple NVIDIA GPUs.

[0102] Forward propagation: The data (preprocessed INT8 tensors) of each training batch are input into the model, passing through the input layer, backbone network, feature fusion layer, and classification head in sequence, and finally outputting a 5-dimensional probability distribution.

[0103] Loss Calculation and Backpropagation: The total loss is calculated based on the probability distribution of the model output and the true labels of the samples. The gradient of the loss with respect to all trainable parameters of the model is calculated using the backpropagation algorithm.

[0104] Parameter update: The optimizer uses the calculated gradient to update all weight parameters of the model according to its internal rules.

[0105] Batch normalization and dropout: Batch normalization layers are used in 3D-CNNs to accelerate training and improve stability. Dropout is used in fully connected layers and other locations to randomly disable a subset of neurons, thus reducing overfitting.

[0106] Dynamic weight adjustment: During training, the model's classification accuracy for each level is periodically evaluated on the validation set. For levels with consistently low average accuracy, their difficulty weight in the weighted focus loss term is automatically increased in subsequent training epochs. This forces the model to devote more attention to learning these difficult examples, thus enabling adaptive learning.

[0107] To meet the low latency and low power consumption requirements of edge devices, the trained model needs to be quantized. After the floating-point model training converges, quantization is performed using TensorRT's INT8 quantization tool. The quantization process requires a small calibration dataset, which can be randomly sampled from the training set and not used in training. This dataset is used to statistically analyze the dynamic range of activation values ​​in each layer of the model to determine the optimal quantization parameters, including scaling factors and zeros.

[0108] The quantized model is then compiled and optimized using TensorRT, including layer fusion, automatic kernel tuning, and memory optimization, to generate an inference engine. This engine can be directly deployed to edge computing devices such as NVIDIA Jetson Orin.

[0109] The final quantized model was rigorously evaluated using an independent test set. Key performance indicators included: overall classification accuracy; classification precision, recall, and F1 score for each damage level; model inference latency, tested on the Jetson Orin platform, which must be <80ms; and expert consistency: the model's judgment results on the test set were compared with the results of blind expert reviews, and the consistency percentage was calculated, which must be >95%.

[0110] In some optional implementations, the damage severity assessment model also outputs heatmap data; the classification head of the damage severity assessment model also includes a heatmap branch.

[0111] Correspondingly, the pre-trained damage level determination model performs inference based on the model input data, and also includes: inputting the intermediate feature map into the heatmap branch in the classification head, the heatmap branch processing the intermediate feature map to generate a three-dimensional spatial attention weight map that is consistent with the spatial dimension of the electric field volume data in the model input data; the output layer outputs the three-dimensional spatial attention weight map as heatmap data, where the weight values ​​are used to identify the contribution of different three-dimensional spatial locations in the test body to the damage level determination result.

[0112] Specifically, the classification head consists of the original Softmax classification branch and a newly added heatmap generation branch, forming a dual-task learning architecture. The two branches share input features from the backbone network and the feature fusion layer, and perform different decoding tasks.

[0113] The input to the heatmap branch is an intermediate feature map output from the backbone network that has not undergone global average pooling compression. This feature map retains complete spatial structure information. Specifically, the intermediate feature map is a high-dimensional spatial feature map output by the 3D-CNN branch before the last pooling layer in step S4.2. Its spatial resolution has typically been downsampled to 1 / 8 or 1 / 16 of the original input.

[0114] The heatmap branch processes the input intermediate feature map to generate a three-dimensional attention map that is spatially aligned with the original electric field data. Specifically, it includes the following steps.

[0115] a) Feature decoding and channel compression.

[0116] The heatmap branch first restores the spatial resolution of the downsampled intermediate feature map through one or more deconvolutional or upsampling layers, gradually upsampling it to the same spatial resolution (256, 256, 256) as the input electric field volume data. Simultaneously, channel compression is performed through a 1×1×1 3D convolutional layer, ultimately outputting a single-channel 3D feature map.

[0117] b) Spatial attention weight activation and normalization.

[0118] For each spatial location (voxel) of the single-channel 3D feature map, an element-wise sigmoid activation function is applied, mapping it to the [0,1] interval. This value represents the raw attention weight of that spatial location for the final classification decision. Subsequently, the entire 3D map is typically min-max normalized or normalized by dividing by the maximum value to generate the final 3D spatial attention weight map. The closer the weight value is to 1, the higher the attention given to that spatial region when the model makes a damage level determination, and the greater its contribution to the determination result.

[0119] It should be noted that the generated heatmap is a three-dimensional saliency map registered with the internal space of the original experimental body. Each voxel value quantifies the degree of attention the model pays to the corresponding physical location within the experimental body when making the final grading determination, corresponding to key areas with severe electric field distortion, concentrated energy deposition, or potential physical structural failure.

[0120] To achieve the output of the heatmap, a heatmap supervision loss term is added to the total loss function described above. The total loss function is then expressed as:

[0121] in, This is a hyperparameter.

[0122] Heatmap monitoring of loss items Weakly supervised signals are generated using domain knowledge of the HPM effect. For example, it is assumed that, in most cases, the region with the highest electric field intensity is most likely to be the critical damage area, and the heatmap supervises the loss term. The guided model generates a physically interpretable heatmap, represented as follows:

[0123] in, For the model to the first Heatmaps predicted for each sample, total One sample, For the corresponding input electric field volume data, This is the normalization function. This loss term utilizes the prior physical knowledge that "field strength is positively correlated with energy deposition" in the high-power microwave effect to constrain the spatial distribution of the predicted thermal map to match the electric field energy density distribution, thereby ensuring that the thermal map can reliably indicate potential damage areas of interest and improving the credibility and practicality of the visualization results.

[0124] To enable the model to generate physically meaningful heatmaps simultaneously, targeted adjustments need to be made to the training objectives, data preparation, and model structure. To supervise the learning of the heatmap branch, the training data needs to provide weak spatial supervision signals. This embodiment employs a physically-based, automated annotation method, eliminating the need for additional manual annotation.

[0125] For each sample in the training set, using its own electric field volume data, a distribution map of the square of its electric field intensity is calculated. This physical quantity approximately characterizes the microwave energy deposition density in space. Subsequently, this distribution map is normalized and thresholded to generate a binarized or continuous three-dimensional saliency map, which serves as a weakly supervised label for the heatmap branch training of that sample. This method directly injects domain physics knowledge into the training process.

[0126] The overall loss function adds a heatmap supervision loss term to the original classification loss. Simultaneously, since the learning difficulty and gradient scale of the classification branch and the heatmap branch may differ, gradient normalization or dynamic loss weight adjustment strategies are employed during training to prevent one branch from dominating training and inhibiting the learning of the other. The weights of the heatmap branch are randomly initialized, and its final upsampling layer can be initialized using a bilinear interpolation kernel to ensure a smooth and reasonable spatial attention map is output in the initial stage.

[0127] During training, each batch of data is used to optimize both branches simultaneously. The gradients calculated by backpropagation simultaneously update the shared weights (backbone network) of the classification branch and the heatmap branch, as well as their respective unique weights. Multi-task training enables the backbone network to extract features that are both conducive to accurate classification and capable of decoding key spatial regions, thus achieving the reuse and enhancement of feature representation.

[0128] The foregoing has described in detail an embodiment of an automatic edge AI determination method for HPM damage level. Based on the automatic edge AI determination method for HPM damage level described in the above embodiment, this invention also provides an automatic edge AI determination device for HPM damage level corresponding to the method.

[0129] Figure 2 This is a schematic block diagram of an edge AI automatic determination device for HPM damage level provided in an embodiment of the present invention. This device implements the edge AI automatic determination method for HPM damage level described in the above embodiment, such as... Figure 2 As shown, the device includes the following components.

[0130] Conformal fiber electric field sensor array: Built inside the test body, used to acquire electric field amplitude data inside the test body.

[0131] Temperature and reference data acquisition unit: Deployed inside or on the surface of the test body to acquire temperature field data and fiber Bragg grating reference data inside the test body.

[0132] Plastic optical fiber transmission link: Its input end is connected to the conformal optical fiber electric field sensor array and the temperature and reference data acquisition unit respectively, and its output end passes through the shielded cabin to transmit electric field amplitude data, temperature field data and fiber Bragg grating reference data back to the outside of the shielded cabin in a zero-metal lossless manner.

[0133] Edge computing device: Located outside the shielded cabin, connected to the output end of the plastic fiber optic transmission link, configured to perform the following operations: receive reinjected electric field amplitude data, temperature field data, and fiber Bragg grating reference data; process the electric field amplitude data to generate electric field volume data in the form of a three-dimensional voxel mesh; fuse and format the electric field volume data, temperature field data, and fiber Bragg grating reference data to generate model input data conforming to a predetermined input specification; call the pre-trained damage level determination model deployed inside it, perform inference based on the model input data, and output a multi-level damage level determination result corresponding to the current HPM effect.

[0134] The edge AI automatic determination device for HPM damage level in this embodiment is used to implement the aforementioned edge AI automatic determination method for HPM damage level. Therefore, the specific implementation of this device can be found in the embodiment section of the edge AI automatic determination method for HPM damage level above. Thus, its specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0135] Furthermore, since the edge AI automatic determination device for HPM damage level in this embodiment is used to implement the aforementioned edge AI automatic determination method for HPM damage level, its function corresponds to the function of the above method, and will not be repeated here.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An edge AI automatic determination method of HPM damage level, characterized in that, The method comprises the following steps: Obtain the electric field amplitude data inside the test body through the conformal optical fiber electric field sensor array embedded in the test body; at the same time, obtain the temperature field data inside the test body and the fiber Bragg grating reference data; Through the plastic optical fiber transmission link, the electric field amplitude data, the temperature field data and the fiber Bragg grating reference data are losslessly injected back to the outside of the shielding cabin in a zero-metal manner; On the edge computing device located outside the shielding cabin, the injected electric field amplitude data is processed to generate electric field body data in the form of a three-dimensional voxel grid; And the electric field body data, the temperature field data and the fiber Bragg grating reference data are fused and formatted to generate model input data conforming to the predetermined input specification; The pre-trained damage level judgment model deployed on the edge computing device performs inference based on the model input data and outputs the multi-level damage level judgment result corresponding to the current HPM effect.

2. The edge AI automatic determination method of HPM damage level according to claim 1, characterized in that, The electric field body data, the temperature field data and the fiber Bragg grating reference data are fused and formatted to generate model input data conforming to the predetermined input specification, which specifically includes: According to a unified time reference, the electric field body data, the temperature field data and the fiber Bragg grating reference data are time stamped and aligned and spatially coordinated; The temperature field data is upsampled to the same spatial resolution as the electric field body data to generate a temperature data channel; According to the mapping relationship between the predetermined sensing position and the test body space, the one-dimensional fiber Bragg grating reference data is assigned to all corresponding voxels in the three-dimensional voxel grid to generate a reference data channel with the same spatial resolution as the electric field body data; The electric field body data, the temperature data channel and the reference data channel are spliced in the channel dimension to form multi-channel fusion data; The multi-channel fusion data is subjected to INT8 fixed-point quantization operation to generate a multi-channel three-dimensional data tensor conforming to the predetermined input specification, which is the model input data.

3. The edge AI automatic determination method of HPM damage level according to claim 2, characterized in that, The damage level judgment model includes an input layer, a backbone network, a feature fusion layer, a classification head and an output layer; the backbone network adopts a hybrid architecture of 3D convolutional neural network and time series transformer; the feature fusion layer includes a global average pooling layer and a fully connected layer; the classification head includes a Softmax classifier.

4. The edge AI automatic determination method of HPM damage level according to claim 3, characterized in that, The pre-trained damage level judgment model performs inference based on the model input data, specifically including: The model input data is directly received by the input layer, or the input layer performs the step of generating the model input data; The spatial distribution features of the model input data are extracted by the 3D convolutional neural network in the backbone network, and the time series dynamic features of the model input data at consecutive time points are extracted by the time series transformer, and the spatial distribution features and the time series dynamic features are fused to obtain an intermediate feature map of fused space-time features; The intermediate feature map of fused space-time features is compressed into a one-dimensional feature vector through the global average pooling layer, and then subjected to nonlinear transformation and dimension reduction through the fully connected layer to generate a feature vector representing global information; The feature vector representing global information is input into the classification head, and the classifier outputs a multi-dimensional vector, where each dimension corresponds to the probability of a predefined damage level; The output layer determines the grade corresponding to the dimension with the maximum probability value as the damage grade determination result, and outputs the maximum probability value as the determination confidence.

5. The edge AI automatic determination method of HPM damage level according to claim 4, characterized in that, The loss function of the damage grade determination model in the training stage is linearly combined by a weighted focal loss term and a critical interval penalty term, and the total loss function is represented as: wherein, is a hyperparameter, is a weighted focal loss term, is a critical interval penalty term.

6. The edge AI automatic determination method of HPM damage level according to claim 5, characterized in that, Weighted focal loss term is represented as: wherein, is the damage level, is the number of damage levels, is the true label of the sample, is the probability that the model predicts the sample belongs to level , is the class balance weight, inversely proportional to the number of training samples of each level, is the focus parameter; For level The difficulty weight is adjusted so that for levels with low accuracy in this validation, the difficulty weight of that level is increased in the next round of training. is the indicator function, which is 1 when the true label is the same as the current class and 0 otherwise.

7. The edge AI automatic determination method of HPM damage level according to claim 5, wherein, Critical interval penalty term is represented as: wherein, is the predicted grade is the absolute distance to the true grade is a distance penalty coefficient, which is a piecewise increasing function, is the probability that the model predicts the sample to belong to grade is the number of damage grades.​​ 8. The edge AI automatic judgment method of HPM damage level according to any one of claims 5-7, characterized in that, The damage grade determination model also outputs heat map data; the classification head of the damage grade determination model also includes a heat map branch; The pre-trained damage grade determination model performs inference based on the model input data, which also includes: inputting the intermediate feature map into the heat map branch in the classification head, processing the intermediate feature map in the heat map branch to generate a three-dimensional spatial attention weight map consistent with the spatial dimension of the electric field body data in the model input data; the output layer outputs the three-dimensional spatial attention weight map as heat map data, wherein the weight value is used to identify the contribution of different three-dimensional spatial positions in the test body to the damage grade determination result.

9. The edge AI automatic determination method of HPM damage level according to claim 8, wherein, The loss function of the damage grade determination model in the training stage also includes a heat map supervision loss term, and the total loss function is represented as: wherein is a hyperparameter; Heatmap supervision loss term for guiding the model to generate heatmaps with physical interpretability, denoted as: wherein, is a heat map predicted by the model for the i-th sample, is a heat map predicted by the model for the i-th sample, is a heat map predicted by the model for the i-th sample, is the corresponding input electric field data, is a normalization function.

10. An edge AI automatic determination device of HPM damage level, characterized in that, The method of any one of claims 1 to 9, comprising: a conformal optical fiber electric field sensor array built into the test body to obtain electric field amplitude data inside the test body; a temperature and reference data acquisition unit deployed inside or on the surface of the test body to obtain temperature field data and fiber Bragg grating reference data inside the test body; a plastic optical fiber transmission link with its input end connected to the conformal optical fiber electric field sensor array and the temperature and reference data acquisition unit, and its output end passing through the shielded cabin, for lossless injection of the electric field amplitude data, temperature field data and fiber Bragg grating reference data to the outside of the shielded cabin in a zero-metal manner; an edge computing device disposed outside the shielded cabin and connected to the output end of the plastic optical fiber transmission link, configured to perform the following operations: receive the injected electric field amplitude data, temperature field data and fiber Bragg grating reference data; process the electric field amplitude data to generate electric field body data in the form of a three-dimensional voxel grid; fuse and format process the electric field body data, temperature field data and fiber Bragg grating reference data to generate model input data conforming to the predetermined input specification; call the pre-trained damage grade determination model deployed in its internal to perform inference based on the model input data, and output the multi-level damage grade determination result corresponding to the current HPM effect.

Citation Information

Patent Citations

  • Battlefield target three-dimensional reconstruction and damage evaluation method based on multi-source information fusion

    CN117475080A

  • Mine filling body performance detection and stability evaluation method and system

    CN120506275A

  • Industrial equipment defect detection method based on multi-modal feature fusion and dynamic optimization

    CN121073978A

  • Attention mechanism-fused laryngeal cancer primary screening and lesion area positioning method and device

    CN121213992A

  • Visual real-time monitoring system and method for electric field

    CN121347911A

Cited By

  • Bridge crack image segmentation system and method thereof

    CN121982045A