Intelligent batch display card display test method
By integrating 2D convolutional neural networks and ConvLSTM image recognition models, the problems of unmanned and efficient recognition in batch testing of graphics cards were solved, achieving efficient and accurate graphics card display testing, reducing costs and improving the traceability of test results.
Patent Information
- Application Number
- CN202511975076.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-15
AI Technical Summary
Existing batch testing methods for graphics cards cannot achieve continuous monitoring or unmanned testing. They are costly, inefficient, and difficult to efficiently identify complex and transient display anomalies.
An image recognition model that integrates 2D convolutional neural networks and ConvLSTM convolutional long short-term memory networks, combined with multi-scale feature extraction and temporal dynamic modeling, is used to automatically identify whether the content displayed on the graphics card is normal and generate test records.
It enables efficient batch testing, improves test throughput, reduces labor costs, enhances identification accuracy and test result traceability, and supports flexible expansion and optimization.
Smart Images
Figure CN122044965A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of graphics card testing technology, and in particular relates to an intelligent batch graphics card display testing method. Background Technology
[0002] As the core display output component of a computer, the quality and stability of the graphics card are of paramount importance. Rigorous display function testing is required during the production, repair, and quality inspection processes.
[0003] The core objective of mass production graphics card testing is to filter out products that fail to meet performance standards or have functional defects, while ensuring the consistency of each batch of graphics cards. The entire process is divided into two main modules: pre-shipment testing and production line sampling verification. Batch testing on the production line is the main testing process in the mass production stage. For each graphics card leaving the factory, rapid testing is completed using standardized fixtures and automated equipment to avoid errors.
[0004] The testing process involves inserting the graphics card into a mass production test fixture (compatible with different PCIe slots), connecting it to a stable power supply and motherboard signals, and then starting the test program via an automated script. Traditional testing methods involve simultaneously displaying the output of multiple graphics cards on a large video wall. Testers visually inspect the displayed content, quickly checking for issues such as dead pixels, color casts, black screens, and screen distortion; simultaneously verifying that the resolution and refresh rate meet preset standards (e.g., 4K@60Hz, 2K@144Hz). Traditional testing methods rely heavily on manual operation and visual observation. Testers need to install the graphics card on a test machine, run the test program, and then observe the screen for abnormalities such as screen distortion, stripes, stuttering, and misalignment.
[0005] This method firstly requires manual visual identification of the display, which is inefficient and can only test one or a small number of graphics cards at a time, failing to meet the needs of large-scale production; secondly, it is highly dependent on the subjective experience and attention of the testers, and long working hours can easily lead to misjudgment or omission due to fatigue, resulting in poor consistency of test results; thirdly, the testing process is difficult to standardize and quantify, and the test results are not fully recorded, which is not conducive to problem tracing and analysis.
[0006] With the development of artificial intelligence and computer vision technologies, some solutions for automated detection using image recognition have emerged. However, existing solutions are mostly designed for single-card testing, or they are insufficient in the accuracy of recognition algorithms and the ability to detect temporal anomalies (such as intermittent stuttering and flickering), making them difficult to apply directly to efficient and reliable batch graphics card testing scenarios. Summary of the Invention
[0007] In view of the above problems, the purpose of this invention is to provide an intelligent batch graphics card display testing method and system, which aims to solve the technical problems of existing batch graphics card testing being unable to achieve continuous monitoring and unmanned testing, and having high testing costs and low efficiency.
[0008] The present invention adopts the following technical solution:
[0009] On one hand, the testing method includes the following steps:
[0010] Step S1: Based on historical data of video clips from graphics card tests, train an image recognition model that integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays.
[0011] Step S2: After starting the test, issue test instructions according to the test program script to control the execution of the test program;
[0012] Step S3: Collect the generated timing frame buffer data and runtime logs;
[0013] Step S4: Using the trained image recognition model, perform real-time analysis on the received time frame buffer data, determine whether the displayed content is normal, and output the judgment result.
[0014] Furthermore, the test method also includes:
[0015] Step S5: Based on the judgment result and the running log, associate the image display anomaly with the underlying log events to generate a test record; the test record includes whether a certain display anomaly is accompanied by a system error or warning.
[0016] On the other hand, the testing system includes multiple test bases, into which the graphics card to be tested is inserted, and also includes a test control unit, an image processing unit, and a data collection unit;
[0017] The test control unit is used to issue test instructions according to the test program script and control the test base to execute the test program;
[0018] The test base is used to mount the graphics card under test and run the display test program to generate frame buffer data and running logs.
[0019] The data collection unit is used to collect the generated time-series frame buffer data and runtime logs;
[0020] The image processing unit is used to train an image recognition model based on historical data of video clips from graphics card testing. The image recognition model integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays.
[0021] It also utilizes a trained image recognition model to perform real-time analysis on the received time-series frame buffer data, determine whether the displayed content is normal, and output the judgment result.
[0022] Furthermore, the testing system also includes:
[0023] The result analysis unit is used to associate and correspond image display anomalies with underlying log events based on the judgment results and operation logs, and generate test records; the test records include whether a certain display anomaly is accompanied by system errors or warnings.
[0024] The beneficial effects of this invention are:
[0025] I. High-efficiency batch testing: This method supports testing dozens or even hundreds of graphics cards simultaneously, greatly improving the testing throughput and meeting the cycle time requirements of modern production lines.
[0026] II. Automated and Intelligent Testing: The entire process requires no manual intervention for interpretation. Advanced AI algorithms are used to automatically identify anomalies, freeing up manpower and reducing labor costs.
[0027] III. High accuracy and reliability: The image recognition model, which combines spatial and temporal analysis, has a higher recognition accuracy for complex and transient display anomalies (such as short-term lag and flickering), avoiding human error in detection.
[0028] IV. Test Results Traceability: The generated test records deeply correlate image anomalies with system logs, providing strong data support for subsequent quality analysis, fault location, and product improvement.
[0029] V. Flexibility and Scalability: The system architecture supports easily increasing or decreasing the number of test bases, and the image recognition model can be continuously optimized and adapted to new anomaly types by adding new training samples. Attached Figure Description
[0030] Figure 1 This is a flowchart of the intelligent batch graphics card display testing method provided by the present invention.
[0031] Figure 2 This is a schematic diagram of the intelligent batch graphics card display testing system provided by the present invention. Detailed Implementation
[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.
[0033] To illustrate the technical solution described in this invention, specific embodiments are described below.
[0034] For ease of explanation, only the parts relevant to the embodiments of the present invention are shown.
[0035] Example 1:
[0036] like Figure 1 As shown, the intelligent batch graphics card display testing method provided in this embodiment includes the following steps:
[0037] Step S1: Based on historical data of video clips from graphics card tests, train an image recognition model that integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays.
[0038] This step is the model training phase. A standard display test procedure is selected, and historical video clip data is prepared, including an image / video dataset containing various normal and abnormal display cases, to train the image recognition model. The input to this step is the raw frame data stream of the historical data, and the output is the judgment result. The specific process of this step is as follows:
[0039] S11. Perform multi-channel parallel data preprocessing and quality enhancement on the historical data of the video segments to obtain a standardized frame sequence queue. Here, data preprocessing adopts spatiotemporal alignment, and quality enhancement adopts noise suppression and enhancement. These will be explained below.
[0040] The spatiotemporal alignment process is as follows: the collected video clip historical data is synchronized using a timestamp mechanism to ensure the temporal consistency of the output frames from multiple graphics cards; then, the frame rate is adaptively adjusted to handle the differences in output frame rates from different graphics cards; finally, spatial normalization is performed to unify the different resolution outputs to a standard analysis resolution, such as 2560×1440.
[0041] The noise suppression and enhancement process is as follows: an adaptive median filter is applied to eliminate single-pixel noise points; and a noise reduction algorithm based on wavelet transform is used to preserve image edge features.
[0042] Finally, this step outputs a queue of standardized frame sequences that have been timestamped and noise-reduced.
[0043] S12. Using the U-Net++ network structure of a 2D convolutional neural network, end-to-end multi-scale spatial feature extraction is performed on the standardized frame sequence queue to obtain a set of multi-scale feature maps and heat maps of abnormal regions.
[0044] This step employs a modified U-Net++ network structure to achieve end-to-end multi-scale spatial feature extraction. The specific process is as follows:
[0045] 121. Use the basic feature layer to extract low-level features of edges, corners, and textures. The basic feature layer uses a 3×3 convolution kernel with a stride of 1 and a ReLU activation function.
[0046] A 3×3 convolutional kernel is the optimal small convolutional kernel for extracting low-level visual features. It can capture local relationships between pixels (adapting to the local characteristics of edges and corners) while reducing the number of parameters and ensuring computational efficiency. A stride of 1 ensures that the feature map size is not excessively reduced, fully preserving the detailed information of the input image and avoiding the loss of low-level features. The ReLU activation function is used to introduce a non-linear transformation, enhancing the model's ability to express complex features, while also solving the gradient vanishing problem and accelerating training convergence.
[0047] Edge features capture abrupt changes in pixel grayscale values (horizontal, vertical, and diagonal edges) through convolutional kernels; corner features capture abrupt changes in pixel grayscale values in multiple directions through a combination of responses from multi-directional convolutional kernels; texture features capture repetitive patterns and arrangement rules of local pixels through stacked convolutional layers or multi-channel convolution.
[0048] 122. Set up multi-scale parallel feature extraction towers, including a high-frequency feature tower for detecting screen distortion, a mid-frequency feature tower for detecting color anomalies, and a low-frequency feature tower for detecting structural anomalies; the high-frequency feature tower extracts high-resolution features from the image, i.e., maintains the original image resolution, the mid-frequency feature tower extracts medium-resolution features from the image, i.e., maintains 50% of the original image resolution, and the low-frequency feature tower extracts low-resolution features from the image, i.e., maintains 25% of the original image resolution.
[0049] The high-frequency feature tower processes images at high resolution (maintaining 75% of the original resolution) and is sensitive to checkerboard noise and mosaic anomalies. The mid-frequency feature tower processes images at medium resolution (50% of the original resolution) and detects color casts, color blocks, and stripes. The low-frequency feature tower processes images at low resolution (25% of the original resolution) and detects image tearing, misalignment, and missing parts.
[0050] 123. An adaptive attention mechanism is introduced to generate a multi-scale feature map set and establish an anomaly feature heatmap to quantify the anomaly confidence. The adaptive attention mechanism includes a spatial attention mechanism and a channel attention mechanism. The spatial attention mechanism enables dynamic focusing on suspicious regions, while the channel attention mechanism enhances important feature dimensions.
[0051] Adaptive attention achieves dynamic focusing on key information through the complementary collaboration of spatial attention and channel attention. Spatial attention focuses on the spatial dimension of the image, dynamically highlighting suspicious regions (such as the location of abnormal targets) and suppressing redundant background information by learning pixel-level weights. Channel attention focuses on the channel dimension of features, strengthening feature dimensions crucial for anomaly detection (such as low-level feature channels like edges and textures) and weakening ineffective channels by learning channel-level weights.
[0052] Anomaly feature heatmaps are created by weighted fusion of attention weights and feature maps, visually representing the spatial distribution of anomalous regions. Global pooling and normalization are applied to the heatmap, outputting a confidence score in the 0-1 range to quantify the anomaly confidence. This quantifies the probability of an anomaly existing in the image (a higher score indicates a greater likelihood of an anomaly).
[0053] S13. Temporal dynamic modeling is performed using bidirectional ConvLSTM and an anomaly temporal pattern recognition library to obtain a fused image recognition model. The specific process of this step is as follows:
[0054] 131. Establish a temporal feature coding layer: Bidirectional ConvLSTM processes multi-frame temporal windows, such as 30 frames, where the forward LSTM learns the evolution of anomalies from the past to the present; the backward LSTM learns the impact of anomalies on subsequent frames.
[0055] Traditional LSTM can only process unidirectional temporal information, while ConvLSTM, while retaining temporal modeling capabilities, preserves spatial feature information through convolution operations, making it more suitable for image sequence processing. A fixed-length temporal window ensures the integrity of temporal features, avoiding both the loss of key temporal information due to an excessively short window and the increased computational cost of an excessively long window. By fusing the output features of bidirectional ConvLSTM, a complete temporal feature containing both "past evolution" and "future influence" is obtained, providing support for subsequent tasks such as anomaly detection.
[0056] 132. Set up an abnormal timing pattern recognition library, including a stuttering pattern detector, a flickering pattern detector, and a progressive deterioration detector.
[0057] 1. Lag Mode Detector: Used to accurately detect phenomena such as stagnant or severely discontinuous screen updates.
[0058] The process for setting up the stuttering mode detector is as follows:
[0059] Set up the primary perception layer: Based on structural similarity, use the multi-scale structural similarity index to evaluate perception quality. When the multi-scale structural similarity MS-SSIM of consecutive frames > 0.985 and the duration > 83ms, a primary alarm is triggered.
[0060] Set up a motion analysis layer: Based on optical flow, use optical flow algorithm to track the motion trajectory of feature points and calculate the global average motion vector amplitude. In a test scene with continuous motion (such as a rotating 3D model), if the global average motion vector amplitude is detected to be <0.05 pixels / frame and coincides with the primary alarm in time, it is confirmed as "motion freeze type stutter".
[0061] Set up a system resource verification layer: Monitor the GPU usage log in the operation log. If a hardware signal such as "GPU usage drops to <5% instantly" or "VRAM access bandwidth drops abnormally" appears during a suspected stuttering period, it is determined to be a hardware-level or driver-level stuttering with high confidence. If there is no such signal, it is determined to be "abnormal rendering content but the GPU is still working", which needs to be analyzed in conjunction with other detectors.
[0062] 2. Flicker pattern detector, used to identify abnormally rapid changes in brightness, color or pattern of the screen, whether periodic or non-periodic, and captured through dual analysis in the time and frequency domains.
[0063] The flashing mode detector setup process is as follows:
[0064] Frequency domain feature extraction: For periodic flicker, perform fast Fourier transform on the brightness time sequence signal of a specific region of interest (ROI), analyze the spectrum, and look for significant energy peaks that appear at unexpected frequencies (such as within the range of 1-30Hz, not screen refresh rate harmonics). For example, if a strong peak at 15Hz is detected, it is determined that there is "low-frequency periodic flicker".
[0065] Temporal statistical diagnosis: For non-periodic flicker, calculate the inter-frame difference variance of brightness or chroma between consecutive frames. If the variance value shows intermittent, irregular, and severe spikes during normal rendering periods when there is no scene switching, it is judged as "random flicker".
[0066] State machine modeling: For abnormal switching, the state of each frame is marked as "normal" or "abnormal", forming a binary timing sequence. The switching frequency of "normal" or "abnormal" state per unit time (e.g., per second) is counted. If the switching frequency is >10Hz and there is no clear periodicity, it is judged as "high-frequency jitter flicker".
[0067] 3. Progressive deterioration detectors are used to identify latent defects that accumulate slowly and may eventually lead to complete failure.
[0068] The gradual degradation detector setup process is as follows:
[0069] Time-series trend quantification: Key abnormal indicators such as "screen distortion area ratio", "average color error value", "cumulative stuttering time ratio" and "video memory error pixel ratio" are quantified according to time windows (e.g., every 10 seconds). Then, linear regression and sliding window averaging analysis are performed on the abnormal indicator sequence to calculate its regression slope.
[0070] Deterioration pattern identification: If the regression slope of the "percentage of faulty pixels in memory" indicator is > 0.001 / minute and continues to increase, it indicates that there may be slow failure of memory units. The abnormal indicator is correlated with the GPU temperature time series curve. If it is found that the abnormal indicator rises exponentially after the temperature exceeds the set threshold, it indicates that there is a thermal stability problem in the heat dissipation system or chip.
[0071] Prediction and early warning: Based on the autoregressive integral moving average model ARIMA or simple exponential smoothing, the severity of anomalies in the next short period (the next 60-second window) is predicted. When the predicted value exceeds the dynamic early warning line, an "early warning of performance degradation" is issued and it is recommended to focus on monitoring or terminate the test in advance to avoid complete failure.
[0072] 133. Integrate spatiotemporal features and establish a spatiotemporal correlation map, outputting the judgment result: {anomaly type, start time, duration, spatial location, severity score}.
[0073] This step employs a dual-path fusion design, combining convolutional processing with a spatiotemporal correlation graph, to ultimately output a standardized anomaly information tuple {anomaly type, start time, duration, spatial location, severity score}. This approach combines the robustness of feature extraction with the interpretability of spatiotemporal correlation. The judgment result includes anomaly type, temporal information, and location information.
[0074] Step S2: After starting the test, issue test instructions according to the test program script to control the execution of the test program.
[0075] This step is the test startup phase. The graphics card under test is installed on multiple test docks, the system is connected, and then the test programs on all docks are launched in batches with a single click. The specific execution process involves each test dock executing its internally stored test scripts according to the configured requirements. The system also allows for starting, pausing, and terminating the test program, and enables the unified distribution and updating of test scripts and software to ensure a consistent testing environment.
[0076] Step S3: Collect the generated timing frame buffer data and runtime logs.
[0077] This step is the data collection and transmission phase. During the test, each test base collects frame buffer data and operation logs in real time, and then continuously captures frame buffer data and operation logs from each test base and streams the frame data.
[0078] Step S4: Using the trained image recognition model, perform real-time analysis on the received time frame buffer data, determine whether the displayed content is normal, and output the judgment result.
[0079] This step is the image analysis and judgment stage. It uses an image recognition model to analyze the real-time input frame sequence, judging the display status frame by frame or sequence by sequence, identifying display anomalies, and immediately outputting judgment results with detailed information. The image recognition model can identify the type of anomaly, the time of its occurrence, and its specific location on the screen.
[0080] Further preferred options include:
[0081] Step S5: Based on the judgment result and the running log, associate the image display anomaly with the underlying log events to generate a test record; the test record includes whether a certain display anomaly is accompanied by a system error or warning.
[0082] This step is the results synthesis and output stage. After the test is completed, all information is integrated to generate a final test report for each graphics card. The report clearly lists all abnormal events and their corresponding system logs and context information.
[0083] By employing the aforementioned testing methods, this invention can reduce testing costs, including labor costs and monitor procurement costs; simultaneously, it can improve detection accuracy and testing efficiency. Specifically, it accurately identifies and classifies anomalies through computer vision algorithms; it continuously monitors and records anomalies and their occurrence times in a timely manner; and it improves testing efficiency by using a main control computer to control testing machines in batches.
[0084] Example 2:
[0085] Based on the testing method shown in Embodiment 1, this embodiment provides an intelligent batch graphics card display testing system based on the above testing method, such as... Figure 2 As shown, it includes multiple test bases, into which the graphics card to be tested is inserted. It also includes a test control unit, an image processing unit, and a data collection unit.
[0086] The test control unit is used to issue test instructions according to the test program script and control the test base to execute the test program;
[0087] The test base is used to mount the graphics card under test and run the display test program to generate frame buffer data and running logs.
[0088] The data collection unit is used to collect the generated time-series frame buffer data and runtime logs;
[0089] The image processing unit is used to train an image recognition model based on historical data of video clips from graphics card testing. The image recognition model integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays.
[0090] It also utilizes a trained image recognition model to perform real-time analysis on the received time-series frame buffer data, determine whether the displayed content is normal, and output the judgment result.
[0091] In addition, the testing system also includes:
[0092] The result analysis unit is used to associate and correspond image display anomalies with underlying log events based on the judgment results and operation logs, and generate test records; the test records include whether a certain display anomaly is accompanied by system errors or warnings.
[0093] Further preferred options may include:
[0094] The display unit, which is connected to the data collection unit, is used to display the content of multiple test bases on one or more screens in real time for manual monitoring and reference.
[0095] Each functional unit in this testing system implements the corresponding testing steps in Embodiment 1. Multiple external test docks are equipped with the graphics card under test and run display test programs, generating frame buffer data and runtime logs. A test control unit is connected to the external test docks and is used to batch control the running, pausing, and terminating of the display test programs on the test docks, as well as the distribution and updating of test program scripts. The test control unit can simultaneously control and manage multiple test docks, enabling batch distribution and synchronous execution of test tasks. A data collection unit is connected to each external test dock and collects the frame buffer data and runtime logs generated by each test dock. An image processing unit is connected to the data collection unit and internally deploys a trained image recognition model. It receives frame buffer data, identifies display anomalies, and outputs a judgment result containing anomaly type, time information, and location information; the image recognition model is a deep learning model fusing 2D convolutional neural networks and convolutional long short-term memory networks. A result analysis unit is connected to both the image processing unit and the data collection unit, receives the judgment results and the runtime logs, associates display anomalies with log information, and generates the final test record. The image recognition model in the image processing unit is trained under supervision using pre-input samples of normal and abnormal display content to establish a criterion for judging whether the display content is normal or not. Abnormal display content includes one or more of the following: screen tearing, stuttering, and display misalignment.
[0096] In this implementation, the external test dock is a desktop computer without a graphics card, used to provide the hardware platform. Its function is to mount the graphics card under test, run the specified display test program, and output the frame buffer data (i.e., the raw data of the screen display content) and system operation logs generated during the test.
[0097] The test control unit is directly connected to all test bases, serving as the central control hub. It can issue batch commands to multiple test bases, controlling the start, pause, and termination of test programs, and can uniformly distribute and update test scripts to ensure a consistent test environment.
[0098] The data collection unit is responsible for collecting frame buffer data and operation logs from each test base in real time, and summarizing and transmitting them to the subsequent processing unit, serving as the hub for data flow.
[0099] The image processing unit receives frame buffer data from the data collection unit and performs intelligent analysis using its internally deployed image recognition model. This model combines the advantages of 2D convolutional neural networks in spatial feature extraction with the capabilities of ConvLSTM in temporal sequence modeling, effectively processing video stream data to accurately identify temporal display anomalies such as screen tearing, stuttering, and display misalignment. Before testing, the model undergoes training by inputting a large number of labeled normal and abnormal display samples to learn the discrimination criteria. After analysis, the unit not only outputs a binary result of "normal" or "abnormal," but also provides detailed information such as the specific time point of the anomaly, its location on the screen, and the type of anomaly.
[0100] The results analysis unit receives judgment results from the image processing unit and operation logs from the data collection unit. The results analysis unit correlates image-level anomalies with underlying system log events to generate a detailed test record. This record clearly indicates whether a particular display anomaly is accompanied by a specific system error or warning, greatly facilitating fault diagnosis and root cause analysis.
[0101] The display unit is connected to the data collection unit and can project the display screens of all or part of the test bases onto one or more large screens in real time, allowing testers to perform auxiliary visual monitoring and confirmation without interfering with the automated testing process.
[0102] The test control unit, data collection unit, image processing unit, and result analysis unit can be deployed on one or more high-performance servers. External test docks consist of numerous desktop computers without graphics cards, connected to the server cluster via a local area network or dedicated data cables. The display unit can be a large monitor with a multi-screen splitter.
[0103] In practical deployments, the test control unit, data collection unit, image processing unit, and result analysis unit can be deployed on one or more high-performance servers. External test docks consist of numerous desktop computers without graphics cards, connected to the server cluster via a local area network or dedicated data cables. The display unit can be a large monitor with a multi-screen splitter.
[0104] The following is a specific example to illustrate this.
[0105] This testing system was deployed in the quality control workshop of a graphics card manufacturing plant. 100 testing stations (i.e., 100 external testing docks) were set up.
[0106] First, model training is performed: using historical data accumulated within the factory, including tens of thousands of normally displayed screenshots and thousands of video clips containing abnormalities such as screen tearing, stripes, color blocks, and stuttering, the image recognition model (2DCNN + ConvLSTM model) in the image processing unit is trained until the model achieves a recognition accuracy of over 99% on the validation set.
[0107] Then, proceed with the routine testing process:
[0108] 1. Workers insert 100 graphics cards to be tested into 100 test sockets.
[0109] 2. On the management interface of the test control unit, the engineer selects the standard 3DMark stress test script and clicks "Batch Execution". The test control unit distributes the script to 100 docks via the network and starts the test simultaneously.
[0110] 3. The data collection unit captures frame buffer data and system / application logs at 60 frames per second in real time through an agent program installed on each test base, and transmits them to the server via a high-speed network.
[0111] 4. The image processing unit launches 100 parallel analysis threads, each corresponding to the video stream of one test base. The model analyzes each frame of the image in real time and combines the information from previous and subsequent frames to determine the temporal continuity. For example, when analyzing the video stream of base number 47, the model detects that at 3 minutes and 25 seconds after the start of the test, snow-like noise appears in the upper right corner of the screen for 15 consecutive frames, which is then marked as a "screen distortion" anomaly, and the timestamp and coordinate information are recorded.
[0112] 5. The results analysis unit received an anomaly alert from the image processing unit regarding stand #47. Simultaneously, it discovered a graphics card driver error record for that stand around the same time in the logs. The results analysis unit correlated these two pieces of information and generated a high-priority fault item in the test record of graphics card #47: "Time: 03:25, Anomaly Type: Screen Flickering, Location: (80%, 10%)-(95%, 20%), Corresponding Log: GPU Driver Timeout Error".
[0113] 6. At the same time, the display unit displays thumbnails of 100 images on the monitoring screen in real time, allowing inspection personnel to intuitively see the overall test situation and quickly confirm any abnormalities in the system alarms.
[0114] 7. After the test, the system automatically generates a comprehensive report containing the test results of all 100 graphics cards. Qualified graphics cards proceed to the next stage, while unqualified graphics cards (such as card number 47) are sent for repair and analysis. The entire process is efficient, automated, and data-driven, significantly improving quality inspection efficiency and reliability.
[0115] Therefore, this embodiment tests a system comprising several external test bases, a test control unit, a data collection unit, an image processing unit, a result analysis unit, and a display unit. The test control unit controls multiple test bases without graphics cards to run the display test program; the data collection unit collects frame buffer data and operation logs from each base; the image processing unit uses an image recognition model based on a combination of 2D convolutional neural networks and ConvLSTM to process the time-series display data and automatically identify display anomalies and their details; the result analysis unit integrates the image judgment results and operation logs to generate corresponding test records; and the display unit is used for optional manual monitoring. This invention achieves parallel and automated testing of multiple graphics cards, improving testing efficiency and accuracy while reducing labor costs.
[0116] In summary, this invention provides an intelligent batch graphics card display testing solution for parallel and automated display function testing of multiple graphics cards, with the following specific features: Unified management: After connecting all graphics cards under test to the test bench, they are all connected to a single main control computer via network cables and a switch. The main control computer issues test-related commands. Batch automated testing: The main control computer distributes and runs batch test scripts to the test machines. Continuous monitoring: The main control computer monitors the display status, kernel status, and other relevant statuses of the test bench, and records operational status, abnormal situations, and problems. Unmanned testing: The testing process requires only a small amount of manual intervention to install graphics cards, confirm test scripts, input example information, and start the test; the rest is handled by the system. Intelligent recognition: The main control computer automatically processes data using image recognition algorithms, eliminating the need for manual monitoring. Reduced costs: By directly reading display pixel information from the test bench's display buffer, a large number of display devices are unnecessary. If display is required, a monitor can also be connected. Scalability: Multiple graphics cards can be tested simultaneously, and the number can be freely adjusted.
[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the protection scope of the present invention.
Claims
1. An intelligent batch graphics card display testing method, characterized in that: The testing method includes the following steps: Step S1: Based on historical data of video clips from graphics card tests, train an image recognition model that integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays. Step S2: After starting the test, issue test instructions according to the test program script to control the execution of the test program; Step S3: Collect the generated timing frame buffer data and runtime logs; Step S4: Using the trained image recognition model, perform real-time analysis on the received time frame buffer data, determine whether the displayed content is normal, and output the judgment result.
2. The intelligent batch display card display test method according to claim 1, wherein: The testing method also includes: Step S5: Based on the judgment result and the running log, associate the image display anomaly with the underlying log events to generate a test record; the test record includes whether a certain display anomaly is accompanied by a system error or warning.
3. The intelligent batch display card display test method according to claim 2, characterized in that: In step S1, the specific process of training the image recognition model based on historical data of video clips from graphics card testing is as follows: S11. Perform multi-channel parallel data preprocessing and quality enhancement on historical data of video segments to obtain a standardized frame sequence queue; S12. Using the U-Net++ network structure of a 2D convolutional neural network, end-to-end multi-scale spatial feature extraction is performed on the standardized frame sequence queue to obtain a set of multi-scale feature maps and heat maps of abnormal regions. S13. A fusion image recognition model is obtained by using bidirectional ConvLSTM and an abnormal temporal pattern recognition library for temporal dynamic modeling.
4. The intelligent batch display card display test method according to claim 3, characterized in that: In step S11, data preprocessing employs spatiotemporal alignment, and quality enhancement employs noise suppression and enhancement. The spatiotemporal alignment process is as follows: the collected video clip historical data is synchronized using a timestamp mechanism to ensure the temporal consistency of the output frames from multiple graphics cards; then, the frame rate is adaptively adjusted to handle the differences in output frame rates from different graphics cards; finally, spatial normalization is performed to unify the different resolution outputs to the standard analysis resolution. The noise suppression and enhancement process is as follows: an adaptive median filter is applied to eliminate single-pixel noise points; and a noise reduction algorithm based on wavelet transform is used to preserve image edge features.
5. The intelligent batch display card display test method according to claim 4, characterized in that: In step S12, the multi-scale spatial feature extraction process is as follows:
121. Use the basic feature layer to extract low-level features of edges, corners, and textures. The basic feature layer uses a 3×3 convolution kernel with a stride of 1 and a ReLU activation function.
122. Set up multi-scale parallel feature extraction towers, including a high-frequency feature tower for screen distortion detection, a mid-frequency feature tower for color anomaly detection, and a low-frequency feature tower for structural anomalies. The high-frequency feature tower extracts high-resolution features from the image, i.e., it maintains the original image resolution; the mid-frequency feature tower extracts medium-resolution features from the image, i.e., it maintains 50% of the original image resolution; and the low-frequency feature tower extracts low-resolution features from the image, i.e., it maintains 25% of the original image resolution.
123. Introduce an adaptive attention mechanism to generate a multi-scale feature map set and establish an anomaly feature heatmap to quantify anomaly confidence. The adaptive attention mechanism includes spatial attention and channel attention. Spatial attention enables dynamic focusing on suspicious regions, while channel attention enhances important feature dimensions.
6. The intelligent batch display card display test method according to claim 5, characterized in that: The specific process of step S13 is as follows:
131. Establish a temporal feature coding layer: Bidirectional ConvLSTM processes multi-frame temporal windows, where forward LSTM learns the abnormal evolution from the past to the present; The impact of backward LSTM learning of anomalies on subsequent frames; 132. Set up an abnormal timing pattern recognition library, including a stuttering pattern detector, a flickering pattern detector, and a progressive degradation detector; The stuttering mode detector is used to accurately capture phenomena such as stagnant or severely choppy screen updates. A flicker pattern detector is used to identify abnormally rapid changes in brightness, color, or pattern in a screen, whether periodic or non-periodic, and is captured through dual analysis in the time and frequency domains. A progressive degradation detector is used to identify latent defects that accumulate slowly and may eventually lead to complete failure.
133. Integrate spatiotemporal features and establish a spatiotemporal correlation map, outputting the judgment result: {anomaly type, start time, duration, spatial location, severity score}.
7. The intelligent batch display card display test method according to claim 6, characterized in that: In the abnormal time sequence pattern recognition library a. The process for setting up the stuttering mode detector is as follows: Set up the primary perception layer: Based on structural similarity, use the multi-scale structural similarity index to evaluate perception quality. When the multi-scale structural similarity MS-SSIM of consecutive frames > 0.985 and the duration > 83ms, a primary alarm is triggered. Set up a motion analysis layer: Based on optical flow, use the optical flow algorithm to track the motion trajectory of feature points and calculate the global average motion vector amplitude. If the global average motion vector amplitude is detected to be <0.05 pixels / frame and coincides with the primary alarm in time, it is confirmed as "motion freeze type stutter". Set up a system resource verification layer: Monitor the GPU usage log in the operation log. If a hardware signal such as "GPU usage drops to <5% instantly" or "Video memory access bandwidth drops abnormally" appears during a suspected stuttering period, it is determined to be a hardware-level or driver-level stuttering with high confidence. If there is no such signal, it is determined to be "abnormal rendering content but the GPU is still working", which needs to be analyzed in conjunction with other detectors. b. The flashing mode detector setup process is as follows: Frequency domain feature extraction: For periodic flicker, perform fast Fourier transform on the brightness time-series signal of a specific region of interest (ROI), analyze the spectrum, and look for significant energy peaks that appear at unexpected frequencies to determine the presence of "low-frequency periodic flicker". Temporal statistical diagnosis: For non-periodic flicker, calculate the inter-frame difference variance of brightness or chroma between consecutive frames. If the variance value shows intermittent, irregular, and severe spikes during normal rendering periods when there is no scene switching, it is judged as "random flicker". State machine modeling: For abnormal switching, the state of each frame is marked as "normal" or "abnormal", forming a binary timing sequence. The switching frequency of the state "normal or abnormal" per unit time is counted. If the switching frequency is > 10Hz and there is no clear periodicity, it is judged as "high-frequency jitter flicker". c. The process for setting up the progressive degradation detector is as follows: Quantification of time-series trends: The key abnormal indicators such as "percentage of screen distortion area", "average color error value", "percentage of cumulative stuttering time" and "percentage of memory error pixels" are quantified according to time windows. Then, linear regression and sliding window averaging analysis are performed on the abnormal indicator sequence to calculate its regression slope. Deterioration pattern identification: If the regression slope of the "percentage of faulty pixels in memory" indicator is > 0.001 / minute and continues to increase, it indicates that there may be slow failure of memory units. The abnormal indicator is correlated with the GPU temperature time series curve. If it is found that the abnormal indicator rises exponentially after the temperature exceeds the set threshold, it indicates that there is a thermal stability problem in the heat dissipation system or chip. Prediction and early warning: Based on the autoregressive integral moving average model ARIMA or simple exponential smoothing, the severity of anomalies in the short term is predicted. When the predicted value exceeds the dynamic early warning line, an "early warning of performance degradation" is issued and it is recommended to focus on monitoring or terminate the test in advance to avoid complete failure.
8. An intelligent batch display card display test system, characterized in that, The testing system includes multiple test bases, into which the graphics card to be tested is inserted, and also includes a test control unit, an image processing unit, and a data collection unit. The test control unit is used to issue test instructions according to the test program script and control the test base to execute the test program; The test base is used to mount the graphics card under test and run the display test program to generate frame buffer data and running logs. The data collection unit is used to collect the generated time-series frame buffer data and runtime logs; The image processing unit is used to train an image recognition model based on historical data of video clips from graphics card testing. The image recognition model integrates a 2D convolutional neural network and a convolutional long short-term memory network (ConvLSTM). The historical data includes example data of normal and abnormal displays. It also utilizes a trained image recognition model to perform real-time analysis on the received time-series frame buffer data, determine whether the displayed content is normal, and output the judgment result.
9. The intelligent batch display card display test system of claim 8, wherein, The testing system also includes: The result analysis unit is used to associate and correspond image display anomalies with underlying log events based on the judgment results and operation logs, and generate test records; the test records include whether a certain display anomaly is accompanied by system errors or warnings.
10. The intelligent batch graphics card display testing system as described in claim 9, characterized in that, The test control unit is also used to control the start, pause, and termination of the test program, and to uniformly distribute and update the test program scripts.