A set-top box core device quality detection system and method

Abstract

The application discloses a set-top box core device quality detection system and method, the system comprises: a multi-modal data acquisition module for driving and collecting time sequence response data and state information from each core device of the set-top box in parallel and processing to generate multi-modal time sequence data; an AI inference engine module for locally and in parallel processing the multi-modal time sequence data to generate defect classification results of each core device and a whole machine health score; a dynamic test generation module for generating a test pattern sequence in real time; a multi-target decision module for executing multi-level decision logic to output a quality detection result; and a production line integrated interface module for communicating with a factory MES system, reporting the quality detection result and aggregated features after desensitization, and accepting dynamic threshold configuration and OTA update instructions of the model. The application directly deploys a lightweight multi-task AI inference engine in a test firmware of a set-top box main control SoC to realize localized, parallelized and intelligent quality detection.

Description

Technical Field

[0001] This application relates to the field of set-top box testing technology, specifically to a quality testing system and method for core components of a set-top box. Background Technology

[0002] In the set-top box manufacturing process, quality inspection of core components such as storage (eMMC / DDR), networking (WiFi / Bluetooth), audio / video (HDMI), power supply (PMIC, Power Management Integrated Circuit), and interfaces (USB) is a crucial step in ensuring product reliability. Traditional production line inspection methods mainly rely on the following technologies:

[0003] (1) Fixed Pattern Read / Write Test: The memory (eMMC / DDR) is read and written using a predetermined March algorithm or pseudo-random sequence.

[0004] (2) Connectivity and threshold judgment: Perform simple connectivity tests, RSSI measurements or EDID readings on peripheral modules (such as Wi-Fi, HDMI) and compare them with fixed thresholds.

[0005] (3) Cloud or host analysis: Upload the test data to the central server or have it analyzed and processed by an external test host.

[0006] (4) Aging test: Try to expose device defects by running a fixed test sequence for a long time (several hours).

[0007] The above technology has the following drawbacks:

[0008] (1) Insufficient detection coverage: Fixed patterns cannot simulate complex access patterns in the real world (such as mixed read and write, multi-threaded concurrency, and temperature changes), resulting in missed detection of hidden defects such as timing margin failure and sensitivity degradation. The detection of modules such as WiFi and PMIC is particularly crude.

[0009] (2) Low testing efficiency: The tests of each module are executed in sequence and rely on long aging time, which results in a long test time for a single unit, seriously affecting the production line cycle and manufacturing cost.

[0010] (3) False alarms and false alarms coexist: Using a fixed threshold for judgment (e.g., ECC greater than 10 is considered defective) cannot adapt to the individual differences of different batches and supplier devices, resulting in false judgment of good products (increasing retesting costs) or false judgment of defective products (flowing to the market).

[0011] (4) Data analysis is risky: relying on cloud or server analysis poses risks of network latency, bandwidth consumption and data privacy leakage, and cannot provide real-time feedback to guide production line decisions. Summary of the Invention

[0012] To address the technical problems of low coverage of core components (storage, network, audio / video, power supply, interfaces, etc.) in set-top box production line testing, low testing efficiency, reliance on external computing power, high false alarm rate, and inability to accurately locate defects in existing technologies, this application proposes a set-top box core component quality inspection system and method. This system directly deploys a lightweight multi-task AI inference engine into the test firmware of the set-top box main control SoC, achieving localized, parallel, and intelligent quality inspection.

[0013] This application is achieved through the following technical solution:

[0014] A quality inspection system for core components of a set-top box, comprising:

[0015] The multimodal data acquisition module is used to drive and acquire timing response data and status information from various core components of the set-top box in parallel and process them to generate multimodal timing data.

[0016] The AI ​​inference engine module is used to perform local parallel processing on the multimodal time series data to generate defect classification results for each core device and overall health score; wherein the AI ​​inference engine module is a lightweight multi-task neural network model deployed in the test firmware partition of the set-top box SoC.

[0017] The dynamic test generation module is used to generate adaptive and scenario-based test pattern sequences in real time based on the target device type and its current operating status characteristics identified by the AI ​​inference engine identification module.

[0018] The multi-objective decision module is used to receive the defect classification results and overall health score generated by the AI ​​inference engine recognition module, combine the preset hard threshold and the statistically based dynamic threshold, execute multi-level judgment logic, and output the final quality inspection result.

[0019] Additionally, a production line integration interface module is used to communicate with the factory's MES system, report the desensitized quality inspection results and aggregated features, and accept dynamic threshold configuration and OTA update instructions for the model.

[0020] In some implementations, the multimodal data acquisition module is configured as follows:

[0021] Simultaneously perform tests on multiple devices and collect test data within the same test cycle;

[0022] The collected raw test data is preprocessed and then converted into a fixed-dimensional tensor to serve as input to the AI ​​inference engine module.

[0023] In some implementations, the lightweight multi-task neural network model employs a multi-task learning architecture, including a shared feature extraction layer and a multi-task output head;

[0024] The shared feature extraction layer is composed of a one-dimensional convolutional neural network, an attention mechanism, and a bidirectional long short-term memory network, and is used to extract general feature representations related to defects of each device from the input multimodal time-series data;

[0025] The multi-task output head includes multiple classification heads and a health regression head. Each classification head corresponds to a device, is composed of a fully connected layer, and uses the Softmax function to output the probability distribution of the device belonging to various types of defects. The health regression head is used to output the overall health score of the device.

[0026] In some implementations, the training process of the lightweight multi-task neural network model includes:

[0027] Offline training: A training dataset is built, and supervised training is performed on the server using a GPU cluster. The trained model is then lightweighted and quantized, and the lightweighted and quantized model is deployed to the test firmware partition of the set-top box SoC. The loss function used during model training is a weighted sum of multi-task losses.

[0028] The model undergoes continuous OTA evolution: The production line integration interface module periodically packages and uploads the desensitized feature vector set and inference results to the cloud training platform. The cloud training platform gathers feature data from multiple production lines. When an incremental training task is triggered, the gathered feature data is used to fine-tune the existing model. After the new model has undergone accuracy and latency tests on the validation set, it is silently deployed and updated to all detection stations during idle periods of the production line through the downlink configuration interface of the production line integration interface module.

[0029] In some implementations, the training process of the lightweight multi-task neural network model further includes:

[0030] A federated learning optimization model is adopted.

[0031] In some implementations, the dynamic adjustment of the test pattern sequence is as follows:

[0032] During the test, if marginal signals are detected in real time, the dynamic test generation module will adjust the subsequent test pattern sequence in real time: add fine-grained tests in suspicious address areas; adjust the access frequency to trigger timing failures; or add stress tests for specific scenarios.

[0033] In some implementations, the multi-objective decision-making module runs on the set-top box CPU and is used to execute three-level decision logic, including:

[0034] Level 1, Hard Threshold Judgment: Check for serious out-of-specification items. If serious out-of-specification items are found, they are directly judged as defective products and the judgment is terminated. If no serious out-of-specification items are found, proceed to Level 2, Dynamic Threshold Judgment.

[0035] The second level is dynamic threshold determination: the feature value output by the AI ​​inference engine module is compared with the dynamic threshold. If the feature value exceeds the dynamic threshold, it is determined to be a marginal product and needs to be retested; otherwise, it proceeds to the third level of AI confidence determination. The dynamic threshold is based on the statistics of several good products in recent times.

[0036] The third level is AI confidence determination: if the defect probability output by the AI ​​inference engine module is greater than the first probability threshold, it is determined to be a defective product; if the defect probability output by the AI ​​inference engine module is less than the second probability threshold, it is determined to be a good product; other cases are determined by combining the dynamic threshold determination results; wherein the first probability threshold is greater than the second probability threshold.

[0037] In some implementations, the dynamic threshold is calculated as follows:

[0038] Based on recent data from several high-quality equipment units, the preset quantiles of each feature are calculated as the dynamic thresholds for that feature.

[0039] In some embodiments, the production line integration interface module includes:

[0040] The reporting interface, via REST API or message queue, reports the anonymized test results, defect location, and overall machine health score to the factory MES system, without reporting any original timing data or device serial numbers;

[0041] The downlink configuration interface is used to receive dynamic threshold parameters and AI model version update packages issued by the configuration center.

[0042] OTA updates support silent downloading and updating of AI models during production line downtime.

[0043] Data feedback involves periodically packaging and uploading the desensitized feature vectors and recognition results to a cloud training platform for federated learning or incremental training of the model.

[0044] On the other hand, this application also proposes a quality inspection method for core components of a set-top box, implemented based on the quality inspection system for core components of a set-top box described in any of the above embodiments, comprising:

[0045] Parallel acquisition of test data for each component of the set-top box under test;

[0046] Multi-task inference is performed in parallel on the NPU of the set-top box under test to generate defect probability, feature value and overall health score;

[0047] A three-level judgment logic is executed on the CPU of the set-top box under test, and the test result is output.

[0048] Based on the test results, the set-top box under test is diverted to the next corresponding workstation, and the test results are reported.

[0049] After the test is completed, wait for the next set-top box to be tested, and perform the above test process periodically to form a pipeline cycle.

[0050] This application proposes a quality inspection system and method for core components of a set-top box, which can significantly improve the inspection coverage and defect detection rate: the defect detection rate is significantly improved compared with traditional methods, and it can detect latent defects such as timing edge failure, WiFi sensitivity degradation, HDCP handshake timeout, and PMIC ripple abnormality; it can also significantly improve testing efficiency: the estimated total testing time for a single device can be significantly shortened compared with traditional methods, end-side inference (millisecond level) replaces cloud analysis (minute level), eliminates network latency, and parallel testing replaces serial testing; this application has high data privacy and security: all raw data is processed locally on the set-top box, and only the anonymized detection results and aggregated features are reported, which complies with the requirements of the General Data Protection Regulation (GDPR); this application has low false alarm rate and low retest cost: by combining AI confidence and dynamic threshold and other multi-level judgment logic, the estimated false alarm rate is reduced and unnecessary retesting is reduced; this application can also achieve accurate defect location and repair guidance: it can output specific defective devices (such as WiFi) and fine-grained defect types (such as insufficient sensitivity), improving repair efficiency;

[0051] In addition, this application does not increase hardware costs and has good production line adaptability: it can reuse the NPU of the existing SoC of the set-top box without increasing additional hardware costs, supports OTA model updates, and can be continuously optimized. Attached Figure Description

[0052] The accompanying drawings, which are included to provide a further understanding of the embodiments of this application and form part of this application, do not constitute a limitation on the embodiments of this application. In the drawings:

[0053] Figure 1 This is a schematic diagram of the overall architecture of the inspection system proposed in the embodiments of this application;

[0054] Figure 2 This is a schematic diagram of the AI ​​inference engine module architecture according to an embodiment of this application;

[0055] Figure 3 This is a schematic diagram of the three-level decision logic in an embodiment of this application;

[0056] Figure 4This is a schematic diagram of the production line integration interface module according to an embodiment of this application;

[0057] Figure 5 This is a schematic diagram of the inspection method proposed in the embodiments of this application. Detailed Implementation

[0058] In the following, the terms “comprising” or “may include” as used in the various embodiments of this application indicate the presence of a function, operation, or element of the invention and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in the various embodiments of this application, the terms “comprising,” “having,” and their cognates are intended only to indicate a specific feature, number, step, operation, element, component, or combination of the foregoing and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or adding one or more combinations of the foregoing.

[0059] In various embodiments of this application, the expression "or" or "at least one of A and / or B" includes any combination or all combinations of the words listed simultaneously. For example, the expression "A or B" or "at least one of A and / or B" may include A, may include B, or may include both A and B.

[0060] The terms used in the various embodiments of this application (such as "first," "second," etc.) may modify various constituent elements in the various embodiments, but do not limit the corresponding constituent elements. For example, the above terms do not limit the order and / or importance of the elements. The above terms are only used for the purpose of distinguishing one element from other elements. For example, a first user device and a second user device refer to different user devices, although both are user devices. For example, without departing from the scope of the various embodiments of this application, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element.

[0061] It should be noted that if a description is made of "connecting" one component to another, then the first component can be directly connected to the second component, and a third component can be "connected" between the first and second components. Conversely, when a component is "directly connected" to another component, it can be understood that there is no third component between the first and second components.

[0062] The terminology used in the various embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. The terms (such as those defined in a generally used dictionary) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0064] like Figure 1 As shown in the figure, this application proposes a quality inspection system for core components of a set-top box, including:

[0065] The multimodal data acquisition module is used to drive and acquire timing response data and status information from various core components of the set-top box in parallel and process them to generate multimodal timing data.

[0066] The AI ​​inference engine module is used to perform local parallel inference on multimodal time series data to generate defect classification results for each core component and overall system health score. The AI ​​inference engine module is a lightweight multi-task neural network model deployed in the test firmware partition of the set-top box SoC.

[0067] The dynamic test generation module is used to generate adaptive, scenario-based test pattern sequences in real time based on the target device type (device category) and its current operating status characteristics identified by the AI ​​inference engine module.

[0068] The multi-objective decision-making module receives the defect classification results and overall machine health score generated by the AI ​​inference engine module, combines preset hard thresholds and statistically based dynamic thresholds, executes multi-level judgment logic, and outputs the final quality inspection result.

[0069] Additionally, the production line integration interface module is used to communicate with the factory's MES (Manufacturing Execution System, i.e., the information system for production management and scheduling at the factory floor level), report the de-identified quality inspection results and aggregated features, and receive dynamic threshold configuration and OTA update instructions for the model. In other words, the inspection system interacts with external systems through the production line integration module.

[0070] Furthermore, the multimodal data acquisition module in this embodiment is used to drive and acquire timing response data and status data from various core components of the set-top box in parallel and process them to generate multimodal data as the data source for the AI ​​inference engine module. Specifically, the multimodal data acquisition module is configured as follows:

[0071] The parallel acquisition architecture is adopted to perform tests on multiple devices and acquire test data simultaneously within the same test cycle, avoiding the time accumulation of traditional serial tests.

[0072] After preprocessing (missing value imputation, outlier denoising, etc.), the collected raw data is standardized (normalized, formatted) into a fixed-dimensional tensor, with the output format being [time step, feature channel], for use by the subsequent AI inference engine module;

[0073] Multi-threaded / multi-task parallel data acquisition is adopted, with each device test executed independently. Data is then aggregated after all acquisitions are completed to ensure data time alignment.

[0074] Furthermore, such as Figure 2 As shown, the AI ​​inference engine module of this application embodiment adopts a multi-task learning (MTL) architecture, mainly including a shared feature extraction layer and a multi-task output head. The shared feature extraction layer is mainly composed of a one-dimensional convolutional neural network (1D-CNN, used to extract local temporal patterns), an attention mechanism (used to focus on key defect features), and a bidirectional long short-term memory network (Bi-LSTM, used to capture long-term temporal dependencies) connected in series. This connected structure can extract general feature representations related to defects in various devices from multimodal temporal data. The multi-task output head includes multiple classification heads (i.e., a first classification head, a second classification head, ...). The Nth classification head (where N is the number of devices tested) and the health regression head. Each classification head corresponds to a device (e.g., eMMC, DDR, WiFi, HDMI, PMIC, USB), and consists of fully connected layers. It uses the Softmax function (a mathematical function that converts multiple values ​​into a probability distribution) to output the probability distribution of the device belonging to each type of defect. The health regression head consists of two fully connected layers and an output layer connected in series: the first fully connected layer reduces the shared feature vector to 64 dimensions and uses the ReLU activation function, followed by a Dropout layer (dropout rate 0.3); the second fully connected layer further reduces the shared feature vector to 32 dimensions and uses the ReLU activation function; the output layer uses the Sigmoid activation function to map to the 0-100 range and outputs the overall machine health score. This overall machine health score serves as an auxiliary reference for the third-level judgment of the subsequent multi-objective decision module, and is also reported to the MES system by the production line integration interface module for batch quality trend monitoring. In addition, it provides a reference for the repair station to prioritize repairs. Meanwhile, the features in the intermediate feature vector output by the shared feature extraction layer within the AI ​​inference engine module represent the underlying statistical features extracted from multimodal time series data, such as read latency P99, ECC peak value, RSSI variance, etc., which are used for the second-level dynamic threshold determination of the subsequent multi-objective decision module. They also provide a data source for the warning triggering of the subsequent dynamic test generation module.

[0075] The model features a lightweight design: the total number of parameters is approximately 650K, and the model file is less than 200K after INT8 quantization, making it easy to integrate into test firmware. The inference latency is expected to be less than 50ms on mainstream set-top box NPUs (Neural Processing Units), meeting the real-time requirements of production lines.

[0076] The model's inputs and outputs are as follows: The input layer receives tensors generated by the multimodal data acquisition module, and the output layer outputs multiple results simultaneously, including but not limited to: the probability of each device belonging to various types of defects, such as P_eMMC = [P_normal, P_badblock, P_timing_edge], which represents the eMMC identification result, including the probability that the eMMC belongs to normal, badblock, or timing_edge; P_WiFi = [P_normal, P_sensitivity_low, P_reconnect_fail], which represents the WiFi identification result, including the probability that the WiFi belongs to normal, sensitivity_low, or reconnect_fail; and Score_health (a regression value of the overall system health from 0 to 100), which represents the overall system health regression value.

[0077] This lightweight model runs entirely on the NPU of the set-top box SoC (System on Chip), utilizing its parallel computing capabilities to achieve low-latency, low-power inference without occupying main CPU resources for complex calculations.

[0078] This application's embodiments adopt a closed-loop model of "offline centralized training and edge-side online inference optimization" to ensure the model's high accuracy and continuous refinement capabilities. The specific implementation process includes:

[0079] (1) Offline training

[0080] First, construct the training dataset.

[0081] Gold Sample Selection: Select a batch of "good" and "bad" set-top boxes from the production line that have undergone rigorous testing and long-term aging verification using traditional methods;

[0082] Defect injection and edge sample mining: On selected gold samples, controllable defects are created through laboratory methods to obtain sample data covering all defect types. For example, for eMMC / DDR chips, cold solder joints and short circuits are created through soldering processes; for PMICs, ripple interference is injected through external power supply; for Wi-Fi modules, signal attenuation is simulated and sensitivity degradation is created through shielding boxes and attenuators.

[0083] Timing edge sample mining: By adjusting the I / O voltage and clock frequency, the device is made to operate in the Pass / Fail edge state, and its critical failure data is collected;

[0084] Aging Sample Collection: Collect time-series data on the slow degradation of device performance during long-term aging experiments.

[0085] Data annotation: Accurately annotate the collected multimodal time series data. The labels include [device, defect type, health score], such as [eMMC, bad_block, 25], [WiFi, sensitivity_low, 40].

[0086] (2) Model training

[0087] Pre-training: Supervised training is performed on the collected labeled dataset using a GPU cluster on the server side. The loss function is a weighted sum of the losses from multiple tasks.

[0088] L_total = λ1 * L_classification (cross-entropy loss) + λ2 * L_regression (mean squared error loss);

[0089] Where L_classification and L_regression are the classification loss (used to optimize the accuracy of defect probability prediction for each device) and regression loss (used to optimize the accuracy of health score), respectively; λ1 and λ2 are the weights of the classification loss and regression loss, respectively; and L_total is the total loss.

[0090] Lightweighting and Quantization:

[0091] By employing network pruning techniques, low-contribution neuron connections are removed from the trained model, reducing the number of parameters to approximately 650K.

[0092] Using INT8 quantization for perceptual training, the model weights are fine-tuned to simulate the inference process under INT8 precision, and the final exported model file is less than 200K.

[0093] Edge deployment: Package the quantized model files and inference library together into the test firmware partition of the set-top box SoC.

[0094] (3) The model’s OTA continues to evolve

[0095] Data feedback: The production line integration interface module periodically feeds back the anonymized feature vector set (such as [delay P99, ECC (Error Correction Code, a mechanism in memory used to detect and correct data errors), peak value, RSSI variance, ...). The results of the inference and reasoning were packaged and uploaded to the cloud training platform.

[0096] Incremental training: The cloud training platform gathers feature data from multiple production lines around the world. When the data stream accumulates to a certain extent, or when a new defect pattern is discovered, an incremental training task is triggered to fine-tune the existing model.

[0097] Model Validation and Release: After the new model has undergone accuracy and latency testing on the validation set, it is silently released and updated to all testing stations during idle periods of the production line through the configuration release interface of the production line integration interface module.

[0098] Optionally, federated learning can also be used to optimize the model: For customers with extremely high data security requirements, a federated learning framework can be used, where each production line only uploads the model update gradient instead of the original data, completely isolating the risk of data leakage.

[0099] Furthermore, the dynamic test generation module in this embodiment generates an adaptive, scenario-based test pattern sequence in real time based on the target device type and its current operating status characteristics identified by the AI ​​inference engine module (i.e., features in the intermediate feature vector output by the shared feature extraction layer within the AI ​​inference engine module, such as read latency P99, ECC peak value, RSSI variance, etc.), replacing the traditional fixed pattern. Here, the pattern (test mode / test sequence) refers to a series of read / write operations, protocol interactions, and stress application instructions executed when testing various devices in the set-top box. This embodiment differs from the traditional fixed pattern by employing a dynamic pattern.

[0100] Pattern dynamic adjustment mechanism: During the test, if a marginal signal is detected in real time (i.e., a warning event triggered when one or more feature values ​​output by the AI ​​inference engine module enter the preset warning range, such as "ECC count is close to the threshold" or "retransmission rate increases"), the dynamic test generation module will adjust the subsequent pattern in real time: add fine-grained tests in suspicious address areas; adjust the access frequency to trigger timing failures; and add stress tests for specific scenarios.

[0101] Furthermore, the multi-objective decision-making module in this embodiment runs on the CPU and executes three-level decision logic, specifically as follows: Figure 3 As shown, it includes:

[0102] Level 1, Hard Threshold Judgment: Perform a check for serious out-of-specification items. Any serious out-of-specification items will be directly judged as defective (FAIL) to ensure that no absolute defects are missed. Serious out-of-specification items include, but are not limited to: eMMC initialization failure, complete absence of voltage rails, DDR calibration failure, WiFi MAC read failure, etc. If any of the above serious out-of-specification items are present, the product will be directly judged as defective and the judgment will be terminated. Otherwise (i.e., there are no serious out-of-specification items), the product will proceed to Level 2, Dynamic Threshold Judgment.

[0103] The second level is dynamic threshold determination: The feature values ​​output by the AI ​​inference engine module (such as read latency P99, maximum ECC value, etc.) are compared with dynamic thresholds (such as 99th percentile) based on statistics from N (e.g., 1000) good products recently. If the feature value exceeds the dynamic threshold, it is determined to be an edge product (EDGE, the test result is in the critical state between PASS and FAIL) and needs to be retested; otherwise (i.e. the feature value does not exceed the dynamic threshold), it enters the third level of AI confidence determination.

[0104] The third level is AI confidence level determination: If the defect probability output by the AI ​​inference engine module is greater than the first probability threshold (e.g., 0.7, which is configurable), it is determined to be a defective product (FAIL); if the defect probability output by the AI ​​inference engine module is less than the second probability threshold (e.g., 0.3), it is determined to be a good product (PASS); other cases (i.e., greater than or equal to the second probability threshold and less than or equal to the first probability threshold) are comprehensively judged based on the dynamic threshold determination results. The first probability threshold is greater than the second probability threshold.

[0105] Furthermore, one implementation scheme for dynamic threshold calculation is as follows: based on recent data from a number of (e.g., 1000) good-quality equipment, calculate the preset quantile (e.g., the 99th quantile) of each feature as the dynamic threshold of that feature.

[0106] Furthermore, this testing system interacts with external systems, such as configuration centers, MES systems, and cloud training platforms, through a production line integration interface module. Specifically, for example... Figure 4 As shown, the production line integration interface module includes:

[0107] The reporting interface, via REST API or message queue, reports the anonymized test results (PASS / FAIL / EDGE), defect location (e.g., insufficient WiFi sensitivity), and overall device health score to the MES system without reporting any original timing data or device serial numbers, thus protecting privacy.

[0108] The downlink configuration interface receives dynamic threshold parameters and AI model version update packages from the configuration center.

[0109] OTA updates support silent downloading and updating of AI models during production line downtime, enabling continuous optimization.

[0110] Data feedback involves periodically aggregating and uploading the desensitized feature vectors (not the original data) and recognition results to the cloud training platform for federated learning or incremental training of the model, thus forming a data loop.

[0111] In practical applications, the aforementioned detection system is installed at the detection station on the production line to perform defect detection on products and dynamically sift them according to the detection results (PASS / FAIL / EDGE). FAIL products flow to the repair station, PASS products flow to the packaging station, and EDGE products flow to the retesting station for re-inspection. The detection results are reported to the MES system via REST API. The configuration center manages the AI ​​model version and test parameters and supports OTA upgrades. The de-identified feature vectors are periodically aggregated and uploaded to the cloud training platform for federated learning or incremental training optimization.

[0112] Based on the above-mentioned detection system, this application also proposes a method for quality detection of core components of a set-top box, such as... Figure 5 As shown, it includes the following steps:

[0113] Step 1: Collect test data for each component of the product under test (set-top box) in parallel, including but not limited to storage test data (read / write latency, ECC, number of retries, etc.), network test data (RSSI sequence, retransmission rate, throughput, etc.), audio and video test data (EDID response, HDCP handshake timing, etc.), power supply test data (voltage, ripple, power-on timing, etc.), and interface test data (enumeration success rate, VBUS drop test, etc.).

[0114] Step 2: Perform multi-task inference in parallel on the set-top box NPU (using a lightweight multi-task neural network model) to generate defect probability, feature value, and overall health score.

[0115] Step 3: Execute three-level judgment logic on the set-top box CPU: the first level is hard threshold judgment, the second level is dynamic threshold judgment, and the third level is AI confidence judgment, and output the detection results.

[0116] Step 4: Based on the test results, the products to be tested are diverted to the next appropriate workstation: repair workstation, packaging workstation, or retesting workstation, and the test results are reported.

[0117] Step 5: After the test is completed, wait for the next product to be tested, and perform the above steps 1-4 periodically to form a production line cycle.

[0118] The detection system and method proposed in this application can significantly improve detection coverage and defect detection rate: the estimated defect detection rate can be significantly improved compared with traditional methods, and it can detect latent defects such as timing edge failure, WiFi sensitivity degradation, HDCP handshake timeout, and PMIC ripple abnormality; it can also significantly improve testing efficiency: the estimated total testing time for a single device can be significantly shortened compared with traditional methods, edge-side inference (millisecond level) replaces cloud analysis (minute level), eliminates network latency, and parallel testing replaces serial testing; it has high data privacy and security: all raw data is processed locally on the set-top box, and only the anonymized detection results and aggregated features are reported. It complies with General Data Protection Regulation (GDPR) requirements; it is expected to reduce false alarms and retesting costs: the embodiments of this application combine AI confidence and dynamic thresholds and other multi-level judgment logic to reduce the estimated false alarm rate and reduce unnecessary retesting; it can achieve accurate defect location and repair guidance: the embodiments of this application can output specific defective devices (such as WiFi) and fine-grained defect types (such as insufficient sensitivity), improving repair efficiency; finally, the detection system and detection method proposed in the embodiments of this application do not increase hardware costs and have good production line adaptability: the embodiments of this application can reuse the NPU of the existing SoC of the set-top box without increasing additional hardware costs, support OTA model updates, and can be continuously optimized.

[0119] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0120] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A quality inspection system for core components of a set-top box, characterized in that, include: The multimodal data acquisition module is used to drive and acquire timing response data and status information from various core components of the set-top box in parallel and process them to generate multimodal timing data. The AI ​​inference engine module is used to perform local parallel processing on the multimodal time series data to generate defect classification results for each core component and overall health score. The AI ​​inference engine module is a lightweight multi-task neural network model deployed in the test firmware partition of the set-top box SoC; The dynamic test generation module is used to generate adaptive and scenario-based test pattern sequences in real time based on the target device type and its current operating status characteristics identified by the AI ​​inference engine identification module. The multi-objective decision module is used to receive the defect classification results and overall health score generated by the AI ​​inference engine recognition module, combine the preset hard threshold and the statistically based dynamic threshold, execute multi-level judgment logic, and output the final quality inspection result. Additionally, a production line integration interface module is used to communicate with the factory's MES system, report the de-identified quality inspection results and aggregated features, and accept dynamic threshold configuration and OTA update instructions for the model.

2. The set-top box core component quality inspection system according to claim 1, characterized in that, The multimodal data acquisition module is configured as follows: Simultaneously perform tests on multiple devices and collect test data within the same test cycle; The collected raw test data is preprocessed and then converted into a fixed-dimensional tensor to serve as input to the AI ​​inference engine module.

3. The set-top box core component quality inspection system according to claim 1, characterized in that, The lightweight multi-task neural network model adopts a multi-task learning architecture, including a shared feature extraction layer and a multi-task output head; The shared feature extraction layer is composed of a one-dimensional convolutional neural network, an attention mechanism, and a bidirectional long short-term memory network, and is used to extract general feature representations related to defects of each device from the input multimodal time-series data; The multi-task output head includes multiple classification heads and a health regression head. Each classification head corresponds to a device, is composed of a fully connected layer, and uses the Softmax function to output the probability distribution of the device belonging to various types of defects. The health regression head is used to output the overall health score of the device.

4. The set-top box core component quality inspection system according to claim 3, characterized in that, The training process of the lightweight multi-task neural network model includes: Offline training: A training dataset is built, and supervised training is performed on the server using a GPU cluster. The trained model is then lightweighted and quantized, and the lightweighted and quantized model is deployed to the test firmware partition of the set-top box SoC. The loss function used during model training is a weighted sum of multi-task losses. The model undergoes continuous OTA evolution: The production line integration interface module periodically packages and uploads the desensitized feature vector set and inference results to the cloud training platform. The cloud training platform gathers feature data from multiple production lines. When an incremental training task is triggered, the gathered feature data is used to fine-tune the existing model. After the new model has undergone accuracy and latency tests on the validation set, it is silently deployed and updated to all detection stations during idle periods of the production line through the downlink configuration interface of the production line integration interface module.

5. The set-top box core component quality inspection system according to claim 4, characterized in that, The training process of the lightweight multi-task neural network model also includes: A federated learning optimization model is adopted.

6. The set-top box core component quality inspection system according to claim 1, characterized in that, The dynamic adjustment method for the test pattern sequence is as follows: During the test, if marginal signals are detected in real time, the dynamic test generation module will adjust the subsequent test pattern sequence in real time: add fine-grained tests in suspicious address areas; adjust the access frequency to trigger timing failures; or add stress tests for specific scenarios.

7. The set-top box core component quality inspection system according to claim 1, characterized in that, The multi-objective decision-making module runs on the set-top box CPU and is used to execute three-level decision logic, including: Level 1, Hard Threshold Judgment: Check for serious out-of-specification items. If serious out-of-specification items are found, they are directly judged as defective products and the judgment is terminated. If no serious out-of-specification items are found, proceed to Level 2, Dynamic Threshold Judgment. The second level is dynamic threshold determination: the feature value output by the AI ​​inference engine module is compared with the dynamic threshold. If the feature value exceeds the dynamic threshold, it is determined to be a marginal product and needs to be retested; otherwise, it proceeds to the third level of AI confidence determination. The dynamic threshold is based on the statistics of several good products in recent times. The third level is AI confidence determination: if the defect probability output by the AI ​​inference engine module is greater than the first probability threshold, it is determined to be a defective product; if the defect probability output by the AI ​​inference engine module is less than the second probability threshold, it is determined to be a good product; other cases are determined by combining the dynamic threshold determination results; wherein the first probability threshold is greater than the second probability threshold.

8. The set-top box core component quality inspection system according to claim 7, characterized in that, The dynamic threshold is calculated as follows: Based on recent data from several high-quality equipment units, the preset quantiles of each feature are calculated as the dynamic thresholds for that feature.

9. The set-top box core component quality inspection system according to claim 1, characterized in that, The production line integration interface module includes: The reporting interface, via REST API or message queue, reports the anonymized test results, defect location, and overall machine health score to the factory MES system, without reporting any original timing data or device serial numbers; The downlink configuration interface is used to receive dynamic threshold parameters and AI model version update packages issued by the configuration center. OTA updates support silent downloading and updating of AI models during production line downtime. Data feedback involves periodically packaging and uploading the desensitized feature vectors and recognition results to a cloud training platform for federated learning or incremental training of the model.

10. A method for quality inspection of core components of a set-top box, characterized in that, The set-top box core component quality inspection system based on any one of claims 1-9 includes: Parallel acquisition of test data for each component of the set-top box under test; Multi-task inference is performed in parallel on the NPU of the set-top box under test to generate defect probability, feature value and overall health score; A three-level judgment logic is executed on the CPU of the set-top box under test, and the test result is output. Based on the test results, the set-top box under test is diverted to the next corresponding workstation, and the test results are reported. After the test is completed, wait for the next set-top box to be tested, and perform the above test process periodically to form a pipeline cycle.

Images