Chip radiation source positioning method based on near-field scanning and artificial intelligence

Abstract

The application provides a chip radiation source positioning method based on near-field scanning and artificial intelligence. The method comprises the following steps: performing three-dimensional space scanning on a to-be-tested chip in a working state to collect original near-field electromagnetic field distribution data; performing preprocessing and feature extraction on the original data to generate a multi-dimensional feature data set; inputting the multi-dimensional feature data set into a pre-trained deep learning discrimination model for comparison and analysis, and outputting chip radiation source positioning results and internal circuit module discrimination results; and fusing the positioning and discrimination results with chip-related drawings to generate a visual diagnosis report. The near-field scanning technology and artificial intelligence analysis are combined to realize automatic positioning of the chip electromagnetic radiation source, intensity evaluation, and traceability and discrimination of the internal circuit module. The method does not need to rely on manual diagnosis of the radiation source exceeding the standard, improves the efficiency, and reduces the time cost and diagnosis accuracy.

Description

Technical Field

[0001] This application relates to the field of electromagnetic compatibility testing, specifically to a chip radiation source localization method based on near-field scanning and artificial intelligence. Background Technology

[0002] As electronic devices develop towards higher frequencies, higher speeds, and higher integration, chip-level electromagnetic compatibility (EMC) issues are becoming increasingly prominent. Traditional chip EMC testing mainly relies on far-field radiated emissions (RE) and conducted emissions (CE) tests. While these methods meet standards, they can only determine whether emissions exceed limits at the device level and cannot accurately pinpoint the specific radiation source within the chip, let alone identify which circuit module is causing the problem.

[0003] Near-field scanning technology is currently the mainstream method for chip-level EMC diagnostics, locating radiation sources by measuring the near-field magnetic (H) or electric (E) field distribution on the chip surface. However, existing technologies have significant limitations. They require experienced engineers to visually observe and interpret the scan patterns, which is inefficient and highly subjective; they can only provide the spatial distribution of radiation intensity, making it difficult to directly correlate surface hotspots with specific circuit modules deep within the chip; and they cannot automatically identify the nature of the radiation source, making manual analysis almost powerless for complex modulation signals or situations involving multiple coupled radiation sources. Therefore, the industry urgently needs an automated and intelligent solution for chip EMI fault diagnosis, location, and tracing to significantly shorten product development cycles and reduce rectification costs. Summary of the Invention

[0004] To address the problem that current chip-level EMC diagnostics cannot automatically identify the nature of radiation sources in practical applications, this application provides a chip radiation source localization method based on near-field scanning and artificial intelligence.

[0005] The first aspect of this application provides a chip radiation source localization method based on near-field scanning and artificial intelligence, the method comprising: Perform a three-dimensional spatial scan on the chip under test in the working state to collect raw data on the near-field electromagnetic field distribution; The raw data is preprocessed and features are extracted to generate a multidimensional feature dataset; The multidimensional feature dataset is input into a pre-trained deep learning discrimination model for comparison and analysis, and the chip radiation source localization result and internal circuit module discrimination result are output. The system integrates the location and identification results with chip-related drawings to generate a visual diagnostic report.

[0006] Optionally, the raw data of the near-field electromagnetic field distribution collected includes: The chip under test is placed on a 3D scanning platform in a shielded environment; Connect the near-field probe to the spectrum analyzer and configure the scanning parameters and target frequency point or band; The chip surface is scanned across the entire area according to the preset scanning path, and the spatial coordinates, field strength amplitude and phase information are recorded simultaneously. The scan data are compiled to form a raw dataset of near-field electromagnetic field distribution.

[0007] Optionally, the preprocessing and feature extraction of the original data includes: The raw data is calibrated and normalized to remove outliers; Integrate spatial coordinates, frequency, field strength amplitude, and phase information to construct a multidimensional feature dataset; Key features such as radiation intensity distribution, hotspot clustering patterns, and field strength variation trends were extracted from the dataset. The key features are dimensionality reduced to form a feature vector that is suitable for the model input.

[0008] Optionally, the process of training a deep learning discrimination model includes: Near-field scanning data of chips with various known internal layouts and electromagnetic properties under different operating modes were collected. Mark the radiation hotspots in the scanned data to clarify their spatial location and corresponding internal circuit modules; Select a deep learning model adapted for spatial data processing and import the labeled dataset for training; The model performance is evaluated using a validation set, and the model parameters are optimized until the preset accuracy requirements are met.

[0009] Optionally, the step of selecting a deep learning model adapted for spatial data processing and training it by importing the labeled dataset includes: Choose a convolutional neural network as the basic architecture based on the characteristics of the data; Divide the model into training, validation, and test sets, and set the number of iterations and optimization objectives for model training; The model weights are adjusted by backpropagation algorithm, enabling the model to learn the mapping relationship between the near-field distribution pattern and the internal radiation source; The model's localization and module identification accuracy were verified using the test set, and the model training was completed.

[0010] Optionally, the comparison analysis includes the following steps: The processed feature vectors are then input into the trained deep learning discrimination model. The model performs hierarchical analysis of feature vectors to identify the spatial distribution characteristics of radiation hotspots; By combining the mapping relationships learned during model training, the corresponding internal circuit modules are matched. Calculate the contribution probability of each suspected module to form preliminary analysis results.

[0011] Optionally, the output positioning and identification results include the following steps: Extract the accurate spatial coordinates of the radiation source and the radiation intensity level from the model analysis results; Summarize the names of suspected internal functional circuit modules and their corresponding contribution probabilities; Perform a reasonableness check on the results and eliminate any modules that appear to have logical inconsistencies; Output the verified radiation source location results and module identification list.

[0012] Optionally, the fusion of localization and identification results with chip-related drawings to generate a visual diagnostic report includes: Obtain the layout diagram or functional block diagram of the chip under test; Overlay the radiation source coordinates, intensity level, and suspected module information onto the corresponding positions on the chip drawing; The contribution probability of each suspected module is labeled, and an intuitive visualization layer is generated; Integrate layers and analysis descriptions to form a standardized diagnostic report.

[0013] Optionally, the method further includes: Collect simulation data from the chip design phase or near-field scan data from early samples; Generate a feature dataset using the same preprocessing procedure and input it into the model; Obtain the potential radiation risk locations and suspected risk modules from the model output; Based on the prediction results, circuit design optimization suggestions are proposed to avoid EMI risks in advance.

[0014] A second aspect of this application provides a chip radiation source localization system based on near-field scanning and artificial intelligence, the system comprising: The acquisition module is used to perform three-dimensional spatial scanning on the chip under test in the working state and acquire raw data of the near-field electromagnetic field distribution. The feature extraction module is used to preprocess and extract features from the raw data to generate a multidimensional feature dataset; The prediction module is used to input the multidimensional feature dataset into the pre-trained deep learning discrimination model for comparison and analysis, and output the chip radiation source localization result and the internal circuit module discrimination result. The visualization module is used to integrate the positioning and identification results with chip-related drawings to generate a visual diagnostic report.

[0015] As can be seen from the above technical solution, this application proposes a precise and efficient diagnostic method and system for electromagnetic radiation sources of chip-level components. By combining near-field scanning technology with artificial intelligence analysis, it achieves automatic location, intensity assessment, and source identification of electromagnetic radiation sources within the chip, as well as the tracing and identification of their internal circuit modules. It eliminates the need for manual diagnosis of radiation source exceedances, improving efficiency, reducing time costs, and enhancing diagnostic accuracy. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a chip radiation source localization method based on near-field scanning and artificial intelligence, as described in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a chip radiation source localization system based on near-field scanning and artificial intelligence, as described in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Based on this, this application provides an implementation method for a chip radiation source localization method based on near-field scanning and artificial intelligence, such as... Figure 1 As shown, the method includes: S101 performs a three-dimensional spatial scan on the chip under test in the working state and collects the raw data of the near-field electromagnetic field distribution. S102, preprocess and extract features from the raw data to generate a multidimensional feature dataset; S103 inputs the multidimensional feature dataset into the pre-trained deep learning discrimination model for comparison and analysis, and outputs the chip radiation source localization result and the internal circuit module discrimination result. S104 integrates the positioning and identification results with chip-related drawings to generate a visual diagnostic report.

[0021] It should be noted that, to avoid interference from external electromagnetic signals and ensure the authenticity and accuracy of near-field scanning data, the chip under test (DUT) is first deployed on a 3D scanning platform inside the anechoic chamber. For example, the DUT is preprocessed by soldering it onto a matching test board and connecting it to a stable power supply and signal source to ensure the chip can enter a normal working state consistent with the actual application scenario. The test scenario in this application is a 5G signal reception state under mobile phone standby, corresponding to a 900MHz frequency. The test board carrying the DUT is horizontally fixed on the stage of the 3D scanning platform and accurately positioned using fixtures to ensure no displacement or shaking of the chip during scanning, with positioning errors controlled within ±0.01mm. The initial position of the 3D scanning platform is adjusted so that the center of the DUT's package surface is aligned with the initial scanning position of the near-field probe, ensuring that subsequent scanning covers the entire package surface of the chip without any blind spots. The door of the anechoic chamber is closed to ensure it is completely sealed, shielding external electromagnetic radiation signals and preventing leakage of internal electromagnetic signals during scanning, thus ensuring the electromagnetic purity of the scanning environment. The 3D scanning platform is driven by a high-precision servo motor and supports movement in the X, Y, and Z directions. It can meet the high-precision spatial positioning requirements of near-field scanning of chips and provide hardware support for the subsequent accurate acquisition of electromagnetic field data at different spatial locations.

[0022] This application relies on a near-field scanning system and scanning control software. The near-field scanning system mainly consists of a near-field probe and a spectrum analyzer. First, the hardware connection and software configuration of the scanning system are completed. For example, a ring-shaped near-field magnetic field probe adapted to a 900MHz frequency is selected for the hardware connection, as it has high sensitivity to magnetic field signals and good frequency response characteristics. The output end of the near-field probe is connected to the RF input port of the spectrum analyzer via a low-loss RF cable, ensuring a secure and good contact to reduce signal loss and interference during transmission. The near-field probe is then fixed on the probe bracket of the 3D scanning platform, and the probe's posture is adjusted. The probe is positioned so that its detection surface is perpendicular to the package surface of the chip under test, and the initial distance between the probe and the chip surface is controlled at 0.3 mm. This distance is the optimal detection distance for near-field scanning, which ensures the strength of the detection signal while avoiding collision between the probe and the chip surface. Subsequently, the scanning control software is started and initialized on the scanning control host to complete the communication initialization between the software and the 3D scanning platform and the spectrum analyzer. This ensures that the scanning control software can accurately send commands to control the movement of the platform and the operation of the spectrum analyzer, while also receiving and recording the platform's position data and the spectrum analyzer's detection data in real time.

[0023] This application can automatically, quickly, and accurately locate electromagnetic radiation hotspots on the chip surface and further identify the internal functional circuit modules that cause the hotspots.

[0024] In one embodiment that can be implemented in this application, the acquisition of raw near-field electromagnetic field distribution data includes: S201, Place the chip under test on a 3D scanning platform in a shielded environment; S202 connects the near-field probe to the spectrum analyzer, and configures the scanning parameters and target frequency points or bands; S203 performs a full-area scan of the chip surface according to a preset scan path, and simultaneously records spatial coordinates, field strength amplitude and phase information; S204, summarize the scan data to form a raw data set of near-field electromagnetic field distribution.

[0025] For example, the core scanning parameters are configured through the scanning control software. The chip package size of the scanning area is 15mm×15mm, and the scanning area is set to 20mm×20mm, completely covering the chip package surface and a small surrounding area to avoid scanning blind spots. The scanning stepping accuracy is set to 0.5mm in the X and Y directions, taking into account the chip's integration and positioning accuracy requirements. This accuracy ensures scanning efficiency while accurately capturing the details of the electromagnetic field distribution on the chip surface. The Z direction maintains a fixed height of 0.3mm without movement. A raster-type scanning path is used, where the probe scans line by line along the X direction, moves one step value along the Y direction after completing one line, and then scans the next line. This path ensures uniform scanning coverage. Uniformity and integrity; the scanning speed is set to 1 mm / s for the 3D scanning platform to ensure the spectrum analyzer has sufficient time to complete signal detection and data acquisition at each scanning point; target frequency and detection parameter configuration: the detection parameters are configured in the spectrum analyzer, the target frequency is set to 900 MHz, consistent with the frequency point where the chip radiation exceeds the standard, the frequency resolution is set to 1 kHz, the detection method is quasi-peak detection, and the detection range is set to -80 dBm to 20 dBm to ensure complete capture of near-field electromagnetic signals of different intensities; at the same time, a synchronous triggering mechanism for the spectrum analyzer and the scanning platform is set in the scanning control software to ensure that the spectrum analyzer can synchronously complete signal detection and record data when the platform moves to each scanning point.

[0026] After completing the deployment of the chip under test, connection of the scanning system, and parameter configuration, the automatic scanning process is initiated through the scanning control software. The 3D scanning platform drives the near-field probe to perform a full-area 3D scan of the chip package surface according to the preset grating scanning path. During the scanning process, spatial coordinates, field strength amplitude, and phase information are simultaneously acquired and recorded. The specific implementation process is as follows: For example, the scanning control software sends a start command, and the 3D scanning platform moves the near-field probe from its initial position along the X direction, stopping after each 0.5mm step. The platform then sends a trigger signal to the spectrum analyzer. Upon receiving the trigger signal, the spectrum analyzer immediately detects the 900MHz electromagnetic signal at the current scanning point, acquiring the signal's field strength amplitude and phase data, and transmitting the detection data to the scanning control software in real time. The scanning control software simultaneously records the 3D spatial coordinates of the current scanning point, specifically the X and Y coordinates (fixed values, Z=0.3mm), and the field strength amplitude and phase information detected by the spectrum analyzer. This set of coordinate-amplitude-phase data is bound into the raw data of a scanning point and stored in the local database. A single scan... After the data acquisition of the plotted points is completed, the scanning control software sends a command, and the 3D scanning platform continues to move along the X direction by the next step value, repeating the above data acquisition and recording process until one line of scanning is completed. After one line of scanning is completed, the 3D scanning platform moves 0.5mm along the Y direction, and then moves along the opposite X direction to scan the next line, moving according to the step value and completing the data acquisition and recording of each scanning point. This process continues until the entire scanning area is scanned. During the scanning process, the scanning control software displays the scanning progress and the field strength amplitude data of the current scanning point in real time. If the field strength amplitude exceeds the preset range, the probe displacement is abnormal, or the communication is interrupted, the system automatically stops scanning and issues an alarm prompt, allowing staff to promptly troubleshoot the problem and ensure the smooth progress of the scanning process. For example, in this embodiment, the scanning area is 20mm × 20mm, the step accuracy is 0.5mm, and a total of 1600 scanning points are formed. Each scanning point acquires the corresponding spatial coordinates, field strength amplitude, and phase information, completely recording the near-field electromagnetic field distribution characteristics of the 900MHz frequency point on the surface of the chip under test.

[0027] For example, after the full-domain 3D scanning is completed, the scanning control software summarizes and organizes the collected data from all scanning points to form a raw data set of near-field electromagnetic field distribution. The specific implementation process is as follows: The scanning control software extracts the raw data of all scanning points from the local database, sorts it according to the scanning path order, and forms a structured dataset. Each record in the dataset corresponds to a scanning point and includes fields: X coordinate, Y coordinate, Z coordinate, field strength amplitude, phase, and scanning time. The dataset is then checked for completeness, examining for missing scanning point data or data with empty fields. If any are found, the missing locations are marked, and the area can be rescanned to ensure the integrity of the raw data set. The verified structured dataset is stored in a standard data format, and a data acquisition log is generated, recording information such as scanning time, scanning parameters, equipment status, and environmental parameters. This log is bound to the raw data set and stored to form a complete raw data file of near-field electromagnetic field distribution. This file will serve as input data for subsequent data preprocessing and feature extraction. It should be noted that the standard data format in this embodiment is CSV format.

[0028] The raw data collected in this step fully includes core information such as the spatial distribution, field strength, and phase characteristics of the near-field electromagnetic field at a frequency of 900MHz on the surface of the chip under test. This provides real and comprehensive basic data for subsequent data analysis and radiation source location, avoiding errors caused by manual operation and ensuring the objectivity and accuracy of the data.

[0029] In one embodiment of this application, the preprocessing and feature extraction of the raw data includes: S301, calibrate and normalize the raw data, and remove outlier data; S302 integrates spatial coordinates, frequency, field strength amplitude, and phase information to construct a multi-dimensional feature dataset; S303 extracts key features such as radiation intensity distribution, hotspot clustering patterns, and field strength variation trends from the dataset; S304 performs dimensionality reduction on key features to form a feature vector that fits the model input.

[0030] It should be noted that this application performs standardized preprocessing on the collected raw near-field electromagnetic field distribution data, removing invalid data, integrating multi-dimensional information, and extracting key information that reflects the characteristics of the radiation source to generate a multi-dimensional feature dataset and feature vector adapted to the input of the deep learning discrimination model. For example, the raw dataset may contain abnormal data due to factors such as equipment system errors, electromagnetic interference, and scanning displacement deviations. Furthermore, the field strength amplitude data at different scanning points may have inconsistent dimensions, and the phase data may have shifts. Therefore, the raw data first needs to be calibrated and normalized, and abnormal data removed to ensure data validity and consistency. Specifically, based on the calibration parameters of the near-field scanning system, system error compensation is performed on the raw data, including frequency response calibration of the near-field probe, gain calibration of the spectrum analyzer, and loss calibration of the RF cable. The calibration factor is applied to the field strength amplitude data at each scanning point to eliminate the equipment's own system errors and obtain the true field strength amplitude. Using the reference point on the surface of the chip under test package as the phase reference, the phase data of all scanning points is calibrated to eliminate the initial shift during phase measurement, ensuring that the phase data truly reflects the phase distribution characteristics of the electromagnetic field. It should be noted that this application... The reference point for the surface of the chip package to be tested is the ground terminal of the chip pin. The 3σ criterion is used to identify outliers in the calibrated field strength amplitude data. The mean μ and standard deviation σ of the field strength amplitude data are calculated, and data exceeding the range [μ-3σ, μ+3σ] are identified as outliers. Simultaneously, considering spatial distribution characteristics, isolated scan point data that differs significantly from surrounding data are also identified as outliers. For identified outliers, the first step is to check if there is a data acquisition error. If so, interpolation is used to supplement the data with surrounding normal data. If supplementation is not possible, the data is directly discarded to ensure the validity of the remaining data. To eliminate the influence of data dimensions and ensure that the values ​​of different features are on the same order of magnitude, facilitating subsequent AI model training and analysis, the calibrated field strength amplitude data is normalized using a minimum-maximum normalization algorithm to map the field strength amplitude data to the [0,1] interval. The normalization calculation formula is: ,in, This is the normalized field strength amplitude. This is the original, calibrated field strength amplitude. This represents the maximum value of the electric field amplitude. The minimum value of the field strength amplitude is used. The phase data is normalized, mapping the phase values ​​(0°~360°) to the [0,1] interval, using the same minimum-maximum normalization algorithm to ensure dimensional consistency between the phase data and the field strength amplitude data. After completing the calibration, normalization, and outlier removal of the original data, the processed data undergoes multi-dimensional information integration to construct a multi-dimensional feature dataset containing space, frequency, field strength, and phase. Specifically, information dimension supplementation: frequency dimension information is added to the processed dataset. In this embodiment, the target frequency point for scanning is 900MHz, so a frequency field is added to the records of all scanning points, with a fixed value of 900MHz. If it is a frequency band scan, the corresponding value is added according to the actual scanning frequency point. Multi-dimensional information integration uses the scanning point as the basic unit, integrating the four dimensions of spatial coordinate dimension, frequency dimension, field strength feature dimension, and phase feature dimension of each scanning point to form... The multidimensional feature vectors for each scan point are as follows: the frequency dimension represents the target frequency / band, the field strength dimension represents the normalized field strength amplitude, and the phase dimension represents the normalized phase. The multidimensional feature dataset is constructed by arranging the multidimensional feature vectors of all scan points in spatial coordinate order, forming a two-dimensional dataset. The rows of this dataset correspond to the spatial location of the scan points, and the columns correspond to different feature dimensions, specifically X, Y, and Z coordinates, frequency, field strength amplitude, and phase information. The dataset is stored in matrix form, with each element corresponding to a specific feature dimension value for a scan point. This constructed multidimensional feature dataset integrates core information such as the spatial distribution, frequency characteristics, field strength, and phase characteristics of the near-field electromagnetic field, realizing the transformation from single scan data to multidimensional feature information and providing a structured data foundation for subsequent key feature extraction.

[0031] The multidimensional feature dataset contains a large amount of basic feature information. It is necessary to extract key features that reflect the characteristics of electromagnetic radiation sources. These key features are the core basis for AI models to identify radiation sources and distinguish internal circuit modules. In this implementation, the extracted key features include three main categories: radiation intensity distribution features, hotspot aggregation pattern features, and field strength variation trend features. The specific implementation process is as follows: Radiation intensity distribution feature extraction: Based on the field strength amplitude features in the multidimensional feature dataset, the radiation intensity distribution features on the chip surface are extracted, including the spatial distribution matrix of the field strength amplitude, the maximum field strength amplitude, the average field strength amplitude, the variance and standard deviation of the field strength amplitude, and the area of ​​the high field strength region. Spatial location, etc.; Based on the spatial distribution matrix of field strength amplitude, a two-dimensional heat map matrix of radiation intensity is generated. The values ​​of the heat map matrix reflect the radiation intensity at different spatial locations. This matrix, as the core feature of radiation intensity distribution, can intuitively reflect the spatial distribution law of radiation intensity on the chip surface; Hot spot aggregation pattern feature extraction: First, the radiation hot spots on the chip surface are identified. Scanning points with field strength amplitude greater than a preset threshold are defined as radiation hot spots. The spatial coordinates of all radiation hot spots are extracted to form a set of hot spot coordinates. It should be noted that the preset threshold in this application is 0.8 × the maximum field strength amplitude. This preset threshold is adjusted based on the actual scenario, and this application does not impose any restrictions on it. Cluster analysis was performed on the hotspot coordinate set using the K-means clustering algorithm. Based on the spatial distribution of hotspots, they were divided into one or more hotspot clusters. Each hotspot cluster corresponds to a potential radiation source. The core features of each hotspot cluster were extracted, including the center coordinates, radius, number of hotspots within the cluster, and shape of the cluster. These features constitute the core features of the hotspot clustering pattern. If multiple hotspot clusters exist, features such as spatial distance and field strength amplitude differences between the clusters were also extracted to reflect the coupling characteristics of multiple radiation sources. Field strength variation trend features are extracted along the X and Y scanning directions to extract the field strength amplitude variation curves, reflecting the field strength variation trend on the chip surface in different directions. Features such as the slope, curvature, and number of extreme points of the field strength variation curves are calculated to extract the gradient distribution matrix of the field strength amplitude, reflecting the rate and direction of field strength change in space. The larger the gradient value, the more drastic the field strength change, and the closer it is to the core location of the radiation source. Spatial variation trend features of the phase are extracted, including the phase gradient distribution and the location of phase abrupt change points. The core location of an electromagnetic radiation source often has a phase abrupt change, so the location of the phase abrupt change point is an important feature for radiation source localization.

[0032] The key features extracted above reflect the distribution pattern of the near-field electromagnetic field of the chip under test and the characteristics of the radiation source from different perspectives. They break through the surface information of the original data and uncover the feature information that can reflect the essence of the radiation source, providing a core basis for the accurate analysis of the subsequent AI model.

[0033] The extracted key features contain information across multiple dimensions, and some features exhibit redundancy. Directly inputting these features into an AI model would increase computational load, reduce analytical efficiency, and even affect localization accuracy. Therefore, dimensionality reduction of key features is necessary to eliminate redundant features and retain core effective features, forming a feature vector adapted to the input of a deep learning identification model. The specific implementation process is as follows: Key feature normalization: Various key features such as the mean and variance of field strength amplitude, the center coordinates and radius of hotspot clusters, and field strength gradient values ​​are normalized. The min-max normalization algorithm is used to map all key feature values ​​to the [0,1] interval, eliminating dimensional differences between different features and ensuring feature comparability; Feature correlation analysis: The Pearson correlation coefficient method is used to calculate the correlation coefficient between each key feature, analyze the linear correlation between features, and determine features with an absolute correlation coefficient greater than 0.9 as highly correlated features, indicating significant redundancy; Feature dimensionality reduction: Principal component analysis is used. The PCA algorithm performs dimensionality reduction on the normalized key features, transforming the high-dimensional key feature set into a low-dimensional principal component feature set. The specific process is as follows: A feature matrix of key features is constructed, with each row corresponding to a feature sample of a scanned region and each column corresponding to a key feature; the covariance matrix of the feature matrix is ​​calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved; for example, based on the magnitude of the eigenvalues, the top N principal components with a cumulative contribution rate of over 95% are selected, and these principal components can contain over 95% of the information of the original key features; the original key feature matrix is ​​projected onto the eigenvector space corresponding to the top N principal components to obtain the dimensionality-reduced principal component feature matrix; the eigenvector generation process involves vectorizing the dimensionality-reduced principal component feature matrix row by row to form a one-dimensional feature vector. This feature vector has a dimension of n, contains the core feature information of the near-field electromagnetic field distribution of the chip under test, and is free of redundancy, with a suitable dimension, fully adaptable to the input requirements of subsequent deep learning discrimination models, serving as input data for AI model analysis.

[0034] This step, through standardized data preprocessing and feature extraction, transforms the raw near-field scanning data into low-dimensional feature vectors adapted to the AI ​​model. This ensures the integrity of the feature information and improves the efficiency and accuracy of subsequent AI model analysis, making it a key step in achieving intelligent radiation source localization.

[0035] In one embodiment that can be implemented in this application, the process of training a deep learning discrimination model includes: S401 collects near-field scan data of chips with various known internal layouts and electromagnetic characteristics under different operating modes; S402, mark the radiation hotspots in the scan data, and clarify their spatial location and corresponding internal circuit modules; S403: Select a deep learning model adapted for spatial data processing and import the labeled dataset for training. S404 evaluates model performance using a validation set and optimizes model parameters until the preset accuracy requirements are met.

[0036] It should be noted that the deep learning discrimination model in this application is an end-to-end intelligent analysis model. It eliminates the need for manually designing complex feature matching rules. Through training with a large amount of sample data, it can automatically learn the precise mapping relationship between the near-field electromagnetic field distribution pattern and the internal radiation source of the chip. Furthermore, it possesses strong generalization capabilities, enabling it to adapt to the testing needs of chips of different types, packages, and operating modes. After model training, it is deployed on an AI analysis server, enabling real-time, automated analysis of the feature vectors of the chip under test without manual intervention.

[0037] In one embodiment of this application, the step of selecting a deep learning model adapted for spatial data processing and training it by importing the labeled dataset includes: S501 selects a convolutional neural network as the basic architecture based on the characteristics of the data. S502, divide the model into training, validation and test sets, and set the number of iterations and optimization objectives for model training; S503 adjusts the model weights through the backpropagation algorithm, enabling the model to learn the mapping relationship between the near-field distribution pattern and the internal radiation source; S504 uses the test set to verify the model's localization and module identification accuracy, thus completing model training.

[0038] For example, in this embodiment of the application, the labeled chip near-field scanning sample dataset is divided according to a preset ratio. At the same time, combined with the architectural characteristics of the CNN model and the task requirements of chip radiation source diagnosis, reasonable training parameters and optimization objectives are set to ensure the stability and effectiveness of the model training process.

[0039] It should be noted that, in order to avoid uneven sample distribution caused by random partitioning, this application adopts a hierarchical random partitioning method to divide the labeled sample dataset into training set, validation set and test set, with a partitioning ratio of 7:2:1. Moreover, the chip type, working mode and test frequency distribution in each layer of the dataset are consistent with the original dataset, ensuring that the sample characteristics of each dataset are representative.

[0040] For example, this application refines the training parameters of the CNN model based on deep learning training frameworks such as TensorFlow and PyTorchd. Core parameters include the optimizer, learning rate, batch size, number of iterations, and loss function. Each parameter is optimized to meet the dual-task requirements of chip radiation source diagnosis. Specifically, the Adam optimizer is selected, which combines the advantages of momentum and adaptive learning rate methods, enabling adaptive adjustment of the learning rate of each parameter to improve the convergence speed and stability of model training. The initial learning rate is set to 0.001, and the weight decay coefficient is set to 0.0001 to prevent overfitting. The batch size is set according to the hardware performance of the analysis server and the dimensionality of the sample data, typically 32, 64, or 128. In this example, the initial value is set to 64 to ensure both training efficiency and feature diversity of samples within a batch. The initial number of iterations is set to 100 rounds, and an early stopping mechanism is introduced: if the loss function value on the validation set does not decrease for 10 consecutive rounds, model training is automatically stopped to avoid overtraining. For the model's dual-output task, a weighted combined loss function is used: total loss = regression loss × 0.6 + classification loss × 0.4. The regression loss uses mean squared error loss to measure the error between the model's predicted radiation source coordinates and the true coordinates, while the classification loss uses cross-entropy loss to measure the difference between the model's predicted module contribution probability and the true label. Through weight allocation, the accuracy of coordinate regression and module classification are balanced. It should be noted that the above data is for illustrative purposes only and is not intended to limit the scope of this application.

[0041] For example, the model training in this application employs supervised learning combined with backpropagation. The feature vectors of the training set are input into the initialized CNN model. Forward propagation yields the model's prediction results, and the loss value between the prediction results and the true labels is calculated. Then, backpropagation propagates the loss value layer by layer from the output layer to the input layer, dynamically adjusting the weights and bias parameters of each layer. This allows the model to continuously learn the mapping relationship between near-field distribution patterns and internal radiation sources, gradually improving prediction accuracy. The specific implementation process consists of four core steps: forward propagation, loss calculation, backpropagation, and parameter update, which are executed iteratively until the model converges or reaches the preset number of iterations. Forward propagation: The batch sample feature vectors from the training set are input into the CNN model. Through layer-by-layer calculations in the input layer, feature extraction layer, and feature fusion layer, the forward propagation result of the model is obtained, namely the predicted values ​​of the radiation source coordinates and the predicted contribution probability values ​​of each internal circuit module. During forward propagation, the outputs of each layer are processed according to preset activation functions and calculation rules to ensure effective feature transfer and extraction. Loss Calculation: Based on a pre-defined weighted combined loss function, the regression loss (MSE between predicted and true coordinates) and the classification loss (cross-entropy loss between predicted probability and true label) are calculated separately. Then, the total loss value of the model is calculated according to the weight ratios. The magnitude of the total loss value directly reflects the current prediction accuracy of the model; the smaller the loss value, the closer the model's prediction result is to the true value. Backpropagation uses the chain rule, differentiating the total loss value with respect to the weights and bias parameters of each layer of the model to obtain the gradient values ​​of each parameter. The direction of the gradient value indicates the direction in which the parameter needs adjustment, and the magnitude of the gradient value indicates the magnitude of the parameter adjustment. Starting from the output layer of the model, gradient information is passed layer by layer to the input layer until the parameter gradients of all convolutional and fully connected layers are obtained. Parameter updates in this application utilize the Adam optimizer. Based on the gradient values ​​of each parameter and the pre-defined learning rate, the model's weights and bias parameters are adaptively updated. The update formula follows the parameter update rules of the Adam optimizer, adjusting the model's parameters in a direction that reduces the total loss value. Meanwhile, to prevent gradient explosion, the parameter gradients are pruned, and a maximum norm for the gradient is set. If the gradient norm exceeds a threshold, the gradient is normalized. During model training, mini-batch stochastic gradient descent is used, selecting a batch of samples for training and parameter updates each time, rather than the entire training set. This ensures the randomness of model training, improves the model's generalization ability, and reduces hardware resource consumption, thus improving training efficiency. Furthermore, after each training round, the model's performance is evaluated using a validation set, recording the validation set loss value, radiation source localization accuracy, and module identification accuracy, providing a basis for subsequent hyperparameter tuning.

[0042] In one embodiment that can be implemented in this application, the comparison analysis includes the following steps: S601, input the processed feature vector into the trained deep learning discrimination model; S602, the model performs hierarchical analysis of feature vectors to identify the spatial distribution characteristics of radiation hotspots; S603, combined with the mapping relationship learned by model training, matches the corresponding internal circuit module; S604, calculate the contribution probability of each suspected module to form preliminary analysis results.

[0043] It should be noted that the comparative analysis in this application includes, but is not limited to, single-sample real-time analysis and multi-sample batch analysis. Single-sample real-time analysis is suitable for single-chip diagnostic needs in industrial settings, where feature vectors are input and analyzed individually, and results are output in real time. Multi-sample batch analysis is suitable for batch testing needs in the chip R&D stage, where multiple feature vectors are input in batches, and the model is analyzed in parallel, greatly improving the efficiency of batch diagnostics.

[0044] For example, after the feature vector is input into the model, the model identifies the spatial distribution characteristics of the radiation hotspots of the chip under test from the feature vector through multi-level feature extraction and parsing. Relying on the spatial feature extraction capabilities learned during the training phase, the model parses the radiation features layer by layer from shallow to deep, ultimately obtaining core features that can accurately characterize the spatial attributes of the radiation hotspots. During shallow feature parsing, the model performs shallow spatial feature parsing on the input feature vector through the first convolutional layer and pooling layer, extracting the basic spatial features of the radiation hotspots, including: the edge contour of the radiation hotspots, the spatial location range, the basic gradient change of the field strength, and the difference in field strength between the hotspots and the surrounding areas of the chip. Shallow features are an intuitive spatial representation of the radiation hotspots, laying the foundation for subsequent deep feature parsing. During deep feature parsing, the model performs deep spatial feature parsing on the shallow features through subsequent multi-layer convolutional layers, pooling layers, and feature fusion layers, extracting refined and differentiated spatial features of the radiation hotspots, including: the precise center location of the radiation hotspots, the fine gradient distribution of the field strength, the field strength amplitude level of the hotspots, the spatial aggregation pattern of the hotspots, and the correlation features between field strength and phase. Deep features are the core basis for distinguishing the radiation characteristics of different internal circuit modules. During core feature integration, the model integrates and enhances the parsed shallow and deep spatial features through a feature fusion layer and attention mechanism, eliminating irrelevant background features and strengthening core features related to radiation source localization and module identification. This generates a core feature vector that can comprehensively and accurately characterize the spatial distribution of radiation hotspots, preparing for subsequent feature matching. The core feature vector retains all the key spatial attributes of the radiation hotspots and is the direct basis for the model's feature matching. During the feature hierarchical parsing process, the parsing results of each layer of the model are recorded in real time by the AI ​​analysis server. If anomalies occur in subsequent feature matching, the parsing results are used for source analysis to locate the problem and ensure the traceability of the comparison and analysis process.

[0045] It should be noted that after the model extracts the core feature vectors of the radiation hotspots, it uses the near-field distribution pattern-internal radiation source mapping relationship learned during training to match the core feature vectors with the radiation feature libraries of various internal circuit modules stored in the model. This initially filters out suspected internal circuit modules that are highly similar to the radiation hotspot features of the chip under test. This process is the core step in tracing the internal circuit modules of the chip, and the accuracy of the matching directly determines the accuracy of subsequent module identification. Specifically, during the model training phase, the model learns and stores the radiation features of different internal circuit modules to construct a standardized internal circuit module radiation feature library. The feature library is stored in the local database of the AI ​​analysis server, supporting dynamic updates and expansions. Subsequently, the model employs a multi-dimensional cosine similarity matching algorithm to compare the core feature vector of the radiation hotspot of the chip under test with the features of each circuit module in the radiation feature library one by one. The cosine similarity between the feature vector of the chip under test and the features of each module is calculated. The cosine similarity value ranges from -1 to 1; the closer the value is to 1, the higher the feature similarity, and the greater the likelihood that the module is a radiation source. Finally, the model performs preliminary screening of the matching results based on a cosine similarity threshold. Circuit modules with cosine similarity higher than a preset threshold are listed as suspected internal circuit modules, sorted in descending order of cosine similarity to form a preliminary list of suspected modules. If the module with the highest cosine similarity is lower than the preset threshold, the model automatically issues an alarm, indicating that the radiation feature of the chip under test is a new feature not included in the current radiation feature library, requiring an update and expansion of the feature library. It should be noted that this application is not limited to this. For chips under test with multiple radiation sources coupled together, the model can identify the core feature vectors of multiple radiation hotspots, match them with the radiation feature library respectively, and screen out the suspected modules corresponding to each radiation hotspot, so as to achieve accurate matching of multiple radiation sources and multiple modules.

[0046] It should be noted that after the model completes the initial screening of suspected internal circuit modules, based on the similarity results of feature matching, and combined with factors such as the electromagnetic radiation characteristics, operating frequency matching degree, and chip internal layout correlation of each circuit module, the model calculates the probability of each suspected module's contribution to the radiation hotspot through a preset probability calculation model. Finally, the spatial characteristics of the radiation source, the list of suspected modules, and the probability of module contribution are integrated to form the preliminary results of the comparative analysis.

[0047] In one embodiment that can be implemented in this application, the output positioning and identification result includes the following steps: S701, Extract the accurate spatial coordinates of the radiation source and the radiation intensity level from the model analysis results; S702, summarize the names of suspected internal functional circuit modules and their corresponding contribution probabilities; S703 performs a reasonableness check on the results and eliminates modules with suspected logical contradictions. S704 outputs the verified radiation source location results and module identification list.

[0048] It should be noted that the AI ​​analysis server first extracts the core information of radiation source location and internal circuit module identification from the preliminary comparative analysis results, removes redundant intermediate analysis data, retains key information directly related to chip EMI failure and rectification, and classifies and organizes the extracted information according to radiation source location information and internal circuit module identification information to form a structured core information set, which is prepared for subsequent rationality verification.

[0049] It should be noted that in this embodiment of the application, the AI ​​analysis server summarizes and sorts the extracted suspected internal circuit module information to form a standardized list of suspected modules. The list is arranged in descending order of contribution probability, and the core association information of each module is marked, so that technicians can quickly identify the core suspected modules and provide a basis for subsequent rectification priority.

[0050] It should be noted that rationality verification is the core step to ensure the accuracy and effectiveness of the output results. The AI ​​analysis server relies on the chip EMC test knowledge base, the chip internal circuit layout library, and the electromagnetic radiation propagation law library to conduct a comprehensive and rigorous rationality verification of the core information set and the list of suspected modules from four dimensions: frequency point matching, layout correlation, electromagnetic characteristics, and logical consistency. Invalid suspected modules with logical contradictions or inconsistent with objective laws are eliminated to ensure that the retained results are consistent with the actual working principle of the chip and the electromagnetic radiation law.

[0051] It should be noted that after multi-dimensional rationality verification, the AI ​​analysis server integrates the retained core information with the optimized list of suspected modules to generate standardized radiation source location results and internal circuit module identification lists. According to the application needs of industrial sites, the results are output and stored in multiple formats and through multiple channels to ensure that the results can be quickly obtained, viewed and used by technical personnel.

[0052] In one embodiment of this application, the fusion of positioning and identification results with chip-related drawings to generate a visual diagnostic report includes: S801, obtain the layout diagram or functional block diagram of the chip under test; S802 overlays the radiation source coordinates, intensity level, and suspected module information onto the corresponding position on the chip drawing. S803, label the contribution probability of each suspected module, and generate an intuitive visualization layer; S804 integrates layers and analysis descriptions to form a standardized diagnostic report.

[0053] It should be noted that, in this embodiment, the generation of the visual diagnostic report is predicated on obtaining standardized related drawings of the chip under test and preprocessing the drawings to ensure that the coordinate system, scale, and accuracy of the drawings are completely matched with the spatial coordinate system of the radiation source positioning results, thus laying the foundation for accurate information overlay in the future. The related drawings of the chip under test are professional drawings for chip design and testing, serving as the carrier for visual overlay, and mainly include chip layout diagrams, functional block diagrams, and in some scenarios, circuit schematics.

[0054] It should be noted that the visual interactive terminal overlays the radiation source location results, which have undergone reasonable verification, and the internal circuit module identification results onto the corresponding positions of the pre-processed chip-related drawings according to the principles of coordinate matching and hierarchical correspondence. This achieves the organic integration of abstract analysis results and physical chip drawings, enabling technicians to intuitively see the physical location of the radiation source on the chip, as well as the corresponding suspected internal circuit modules.

[0055] In one embodiment that can be implemented in this application, the method further includes: S901 collects simulation data during the chip design phase or near-field scan data of early samples; S902, Generate the feature dataset according to the same preprocessing procedure and input it into the model; S903, module for obtaining potential radiation risk locations and suspected risk areas from the model output; S904 provides circuit design optimization suggestions based on prediction results, thus mitigating EMI risks in advance.

[0056] It should be noted that, in the embodiments of this application, the basis for forward-looking prediction is the collection of forward-looking data related to electromagnetic radiation during the chip design stage. The data types are divided into two categories: chip electromagnetic radiation simulation data and near-field scanning data of early chip samples. Data collection must follow the same standards and specifications as actual testing to ensure the validity, comparability and accuracy of the data, so that the model can accurately identify potential radiation risks based on the data.

[0057] It should be noted that the preprocessing and feature extraction of prospective data are completed automatically by the control and data processing host. Technicians only need to select the corresponding data processing template in the system, and the system can automatically process the data according to the preset process without manual intervention, thus improving data processing efficiency.

[0058] In this embodiment, based on the potential radiation risk list output by the deep learning identification model and relying on the chip EMC design knowledge base, electromagnetic radiation suppression technology library, and industry design specifications, targeted, feasible, and engineering-oriented circuit design optimization suggestions are provided to chip designers. Optimizations are made in aspects such as chip circuit topology, layout and routing, component selection, and shielding and grounding to proactively avoid potential EMI exceedance risks. The proposed design optimization suggestions follow the principles of risk level matching, technical feasibility, and cost control, ensuring that the optimization solution can address radiation risks while considering chip design performance and R&D costs.

[0059] The second aspect of this application proposes a chip radiation source localization system based on near-field scanning and artificial intelligence, such as... Figure 2 As shown, the system includes: The acquisition module 1001 is used to perform three-dimensional spatial scanning on the chip under test in the working state and acquire raw data of the near-field electromagnetic field distribution. The feature extraction module 1002 is used to preprocess the raw data and extract features to generate a multidimensional feature dataset; Prediction module 1003 is used to input the multidimensional feature dataset into the pre-trained deep learning discrimination model for comparison and analysis, and output the chip radiation source localization result and the internal circuit module discrimination result. The visualization module 1004 is used to fuse the positioning and identification results with chip-related drawings to generate a visual diagnostic report.

[0060] For example, taking the rectification of excessive radiation at the 900MHz frequency point of a mobile phone main chip as an example, firstly, a system is set up. The SoC board under test is placed in a shielded anechoic chamber and mounted on a high-precision three-dimensional displacement platform. A near-field probe with a suitable operating frequency is connected to a spectrum analyzer. The scanning area, step accuracy, and target frequency point are set on the visualization terminal. The scan is started, and the system automatically completes data acquisition. Subsequently, the system calibrates and normalizes the acquired amplitude and phase data to form a two-dimensional matrix representing the field strength distribution at 900MHz. This matrix data is then input into a pre-trained CNN model. This model has been trained on a large amount of known chip data and can identify the radiation fingerprint characteristics of different circuit modules. Finally, the model output results show that the strongest radiation hotspot is found at coordinates (x1, y1). This hotspot has an 85% probability of being caused by the power amplifier drive circuit in the 5G Modem module and a 12% probability of being related to the resonance of the power distribution network. Based on this diagnostic result, the power filter circuit of the 5G Modem module is directly optimized, or the integrity of the PA's drive signal is checked, thereby efficiently and accurately solving the problem.

Claims

1. A chip radiation source localization method based on near-field scanning and artificial intelligence, characterized in that, The method includes: Perform a three-dimensional spatial scan on the chip under test in the working state to collect raw data on the near-field electromagnetic field distribution; The raw data is preprocessed and features are extracted to generate a multidimensional feature dataset; The multidimensional feature dataset is input into a pre-trained deep learning discrimination model for comparison and analysis, and the chip radiation source localization result and internal circuit module discrimination result are output. The system integrates the location and identification results with chip-related drawings to generate a visual diagnostic report.

2. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The raw data of the near-field electromagnetic field distribution collected includes: The chip under test is placed on a 3D scanning platform in a shielded environment; Connect the near-field probe to the spectrum analyzer and configure the scanning parameters and target frequency point or band; The chip surface is scanned across the entire area according to the preset scanning path, and the spatial coordinates, field strength amplitude and phase information are recorded simultaneously. The scan data are compiled to form a raw dataset of near-field electromagnetic field distribution.

3. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The preprocessing and feature extraction of the raw data includes: The raw data is calibrated and normalized to remove outliers; Integrate spatial coordinates, frequency, field strength amplitude, and phase information to construct a multidimensional feature dataset; Key features such as radiation intensity distribution, hotspot clustering patterns, and field strength variation trends were extracted from the dataset. The key features are dimensionality reduced to form a feature vector that is suitable for the model input.

4. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The process of training a deep learning discrimination model includes: Near-field scanning data of chips with various known internal layouts and electromagnetic properties under different operating modes were collected. Mark the radiation hotspots in the scanned data to clarify their spatial location and corresponding internal circuit modules; Select a deep learning model adapted for spatial data processing and import the labeled dataset for training; The model performance is evaluated using a validation set, and the model parameters are optimized until the preset accuracy requirements are met.

5. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 4, characterized in that, The selection of a deep learning model adapted for spatial data processing, and the import of the labeled dataset for training, includes: Choose a convolutional neural network as the basic architecture based on the characteristics of the data; Divide the model into training, validation, and test sets, and set the number of iterations and optimization objectives for model training; The model weights are adjusted by backpropagation algorithm, enabling the model to learn the mapping relationship between the near-field distribution pattern and the internal radiation source; The model's localization and module identification accuracy were verified using the test set, and the model training was completed.

6. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The comparative analysis includes the following steps: The processed feature vectors are then input into the trained deep learning discrimination model. The model performs hierarchical analysis of feature vectors to identify the spatial distribution characteristics of radiation hotspots; By combining the mapping relationships learned during model training, the corresponding internal circuit modules are matched. Calculate the contribution probability of each suspected module to form preliminary analysis results.

7. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The output positioning and identification results include the following steps: Extract the accurate spatial coordinates of the radiation source and the radiation intensity level from the model analysis results; Summarize the names of suspected internal functional circuit modules and their corresponding contribution probabilities; Perform a reasonableness check on the results and eliminate any modules that appear to have logical inconsistencies; Output the verified radiation source location results and module identification list.

8. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The fused localization and identification results, along with chip-related drawings, generate a visual diagnostic report including: Obtain the layout diagram or functional block diagram of the chip under test; Overlay the radiation source coordinates, intensity level, and suspected module information onto the corresponding positions on the chip drawing; The contribution probability of each suspected module is labeled, and an intuitive visualization layer is generated; Integrate layers and analysis descriptions to form a standardized diagnostic report.

9. The chip radiation source localization method based on near-field scanning and artificial intelligence according to claim 1, characterized in that, The method further includes: Collect simulation data from the chip design phase or near-field scan data from early samples; Generate a feature dataset using the same preprocessing procedure and input it into the model; Obtain the potential radiation risk locations and suspected risk modules from the model output; Based on the prediction results, circuit design optimization suggestions are proposed to avoid EMI risks in advance.

10. A chip radiation source localization system based on near-field scanning and artificial intelligence, characterized in that, The system includes: The acquisition module is used to perform three-dimensional spatial scanning on the chip under test in the working state and acquire raw data of the near-field electromagnetic field distribution. The feature extraction module is used to preprocess and extract features from the raw data to generate a multidimensional feature dataset; The prediction module is used to input the multidimensional feature dataset into the pre-trained deep learning discrimination model for comparison and analysis, and output the chip radiation source localization result and the internal circuit module discrimination result. The visualization module is used to integrate the positioning and identification results with chip-related drawings to generate a visual diagnostic report.

Images