Memory chip power consumption classification method based on thermal imaging recognition and system and device thereof

Abstract

This invention provides a method, system, and apparatus for classifying the power consumption of memory chips based on thermal imaging recognition, belonging to the field of memory testing technology. The power consumption classification method includes: acquiring thermal imaging data of the memory chip during the execution of a power consumption testing algorithm, including multiple frames of thermal radiation images acquired in a time series; performing temporal and spatial decomposition on the thermal imaging data to extract multi-dimensional features, including temporal and spatial features; inputting the extracted multi-dimensional features into a trained gradient boosting decision tree classification model, from which the model outputs the power consumption level of the memory chip; classifying the power consumption of the memory chip according to the power consumption level and outputting the power consumption classification result. This invention achieves non-contact automatic classification of memory chip power consumption levels with high classification accuracy, low latency, no circuit interference, and high decision robustness.

Description

Technical Field

[0001] This invention relates to the technical field of memory testing, and in particular to a method, system, and apparatus for classifying the power consumption of memory chips based on thermal imaging identification. Background Technology

[0002] Dynamic Random-Access Memory (DRAM) is a core storage device widely used in electronic devices. Power consumption monitoring is a critical aspect of storage testing systems, directly impacting the energy efficiency assessment and stability design of electronic devices. Currently, the industry standard for DRAM chip power consumption monitoring involves connecting a sampling resistor in series with the hardware power supply circuit connected to the DRAM chip. An analog-to-digital converter (ADC) is used to acquire the voltage difference across the sampling resistor. Based on Ohm's law, the line current is calculated, and then the voltage value is combined to calculate the real-time power consumption of the DRAM chip. For example, a sensing resistor is connected in series between the power supply and the chip under test, with its two ends connected to two ADC ports of an MCU. The MCU measures the voltage across the sensing resistor through the ADC ports and calculates the real-time current and power consumption. Alternatively, some solutions obtain the power supply line connected to the chip under test via a power management integrated circuit (PMIC) to determine the resistance value and voltage value of the series sampling resistor to calculate the chip's power consumption data.

[0003] However, the aforementioned existing technical solutions have technical drawbacks. First, connecting a sampling resistor in series on the power supply line of the DRAM chip introduces additional parasitic parameters and impedance, altering the original operating circuit environment of the DRAM chip and increasing system instability and interference. Second, the ADC acquisition link itself suffers from quantization errors and noise interference, affecting the accuracy of power consumption detection. Furthermore, existing solutions typically only reflect coarse power consumption levels under specific operating modes, failing to cover detailed power consumption characteristics across multiple operating modes, thus making it difficult to meet the engineering requirements of high-precision energy efficiency analysis. In addition, the series resistor solution requires configuring an independent sampling resistor and ADC acquisition channel for each power supply line, resulting in high hardware costs. Moreover, as the number of power supply lines on the DRAM chip increases, the board space and design complexity also rise.

[0004] In view of this, a novel method, system, and apparatus for classifying memory chip power consumption based on thermal imaging recognition are proposed to solve all or part of the above problems. Summary of the Invention

[0005] To address at least one of the aforementioned problems and deficiencies in the existing technology, embodiments of the present invention provide a method, system, and apparatus for classifying memory chip power consumption based on thermal imaging recognition. This method non-contactly acquires thermal imaging data of the memory chip during the execution of a power consumption testing algorithm using a thermal imaging device. Multi-dimensional features are extracted using a dual decomposition strategy combining time-domain multi-response stage windowing and spatial-domain quadrant partitioning. A trained gradient boosting decision tree classification model is then used for automatic power consumption level classification. This avoids the interference introduced by the traditional series resistor plus ADC scheme to the memory chip's operating circuitry, achieving non-contact, high-precision, and low-latency power consumption detection. After 5-fold cross-validation and independent test set validation, the model achieves an accuracy of 99%, F1-scores of over 98% across various categories, and single-prediction time in milliseconds. Furthermore, the features exhibit dispersed importance and good decision robustness, meeting the comprehensive requirements of memory chip testing production lines for real-time performance, accuracy, and stability. The technical solution is as follows:

[0006] According to one aspect of the present invention, a method for classifying the power consumption of memory chips based on thermal imaging identification is provided. The method includes:

[0007] Acquire thermal imaging data of the memory chip during the execution of a power consumption test algorithm. This thermal imaging data includes multiple frames of thermal radiation images acquired in a time sequence.

[0008] Thermal imaging data is decomposed in both the temporal and spatial domains to extract multi-dimensional features, which include both temporal and spatial features.

[0009] The extracted multi-dimensional features are input into a trained gradient boosting decision tree classification model, which outputs the power consumption level of the memory chip.

[0010] The power consumption of memory chips is classified according to their power consumption level, and the power consumption classification results are output.

[0011] In some embodiments, performing temporal and spatial decomposition on thermal imaging data to extract multi-dimensional features includes:

[0012] The multi-frame thermal radiation images acquired in time series are divided into multiple time windows, and each time window corresponds to a response phase of the power consumption test algorithm of the memory chip.

[0013] Spatial domain division is performed on the thermal radiation image within each time window, and each frame of thermal radiation image is divided into multiple image blocks;

[0014] Multi-dimensional features are extracted based on time windows and image patches.

[0015] In some embodiments, the multidimensional feature includes at least one or any combination of the following features:

[0016] Window-level global statistical features include the average temperature, maximum temperature, and standard deviation of all pixels within each time window;

[0017] Spatial fine-grained statistical features, including the average temperature, maximum temperature, and standard deviation of each image patch in each time window;

[0018] The temperature mean and temperature standard deviation of the overall temperature series consisting of all time windows;

[0019] Overall temperature sequence characteristics, including the average temperature, standard deviation of temperature, maximum temperature, and minimum temperature of all frames of the thermal radiation image;

[0020] The temperature rise rate of the overall temperature series and the slope of the linear fit of temperature change over time;

[0021] Average frame temperature for each time window;

[0022] The difference between the average frame temperatures of adjacent time windows.

[0023] According to another aspect of the present invention, a memory chip power consumption classification system based on thermal imaging identification is provided. The memory chip power consumption classification system includes:

[0024] A thermal imaging device is configured to acquire thermal imaging data of a memory chip during the execution of a power consumption test algorithm, the thermal imaging data including multiple frames of thermal radiation images acquired in a time series.

[0025] A processor, communicatively connected to the thermal imaging device, is configured to perform the following steps:

[0026] Acquire thermal imaging data collected by thermal imaging equipment;

[0027] Thermal imaging data is decomposed in both the temporal and spatial domains to extract multi-dimensional features, which include both temporal and spatial features.

[0028] The extracted multi-dimensional features are input into a trained gradient boosting decision tree classification model, which outputs the power consumption level of the memory chip.

[0029] The power consumption of memory chips is classified according to their power consumption level, and the power consumption classification results are output.

[0030] According to another aspect of the present invention, a memory chip power consumption classification device based on thermal imaging identification is provided. The memory chip power consumption classification device includes:

[0031] The thermal imaging data receiving module is configured to acquire thermal imaging data of the memory chip during the execution of a power consumption test algorithm, the thermal imaging data including multiple frames of thermal radiation images acquired in a time sequence.

[0032] The feature segmentation module is configured to perform temporal and spatial decomposition on thermal imaging data and extract multi-dimensional features, which include temporal and spatial features.

[0033] The model inference module has a built-in trained gradient boosting decision tree classification model. The model inference module receives multi-dimensional features extracted by the feature subdivision module and outputs the power consumption level of the memory chip by the gradient boosting decision tree classification model.

[0034] The power consumption classification output module is configured to classify the power consumption of memory chips according to their power consumption levels and output the power consumption classification results.

[0035] The memory chip power consumption classification method, system, and apparatus based on thermal imaging identification provided in the embodiments of the present invention have at least one or a portion of the following advantages:

[0036] (1) By acquiring thermal imaging data of the memory chip during the execution of the power consumption test algorithm and performing temporal and spatial decomposition, and then inputting it into the gradient boosting decision tree classification model, the power consumption level of the memory chip can be automatically classified in a non-contact manner, avoiding the parasitic parameters and interference introduced by the series sampling resistor on the power supply line, and improving the stability and reliability of the detection system.

[0037] (2) By dividing thermal imaging data into multiple time windows (e.g., early response window, rising response window, mid-term response window and stable response window) according to time series, and dividing the thermal radiation image in each time window into spatial quadrants, multi-dimensional features including global temperature statistics of each window, temperature statistics of each image block, overall temperature sequence statistics, temperature rise amplitude and linear fitting slope can be extracted to comprehensively characterize the heating characteristics of memory chips in different response stages. The importance of features is dispersed (the highest proportion is about 22%), does not depend on a few features, and has good decision robustness.

[0038] (3) By controlling the acquisition of thermal imaging data and the execution of power consumption test algorithm to start synchronously in time, the frames of thermal imaging data and the response stages of memory chip are aligned in time, ensuring the time correspondence of subsequent feature extraction and the accuracy of power consumption classification results.

[0039] (4) The gradient boosting decision tree classification model uses the softmax function as the output layer activation function and cross-entropy loss as the loss function. The hyperparameters are optimized by grid search (GridSearchCV) combined with 5-fold hierarchical cross-validation. The average accuracy of 5-fold cross-validation reaches 0.972 (±0.018), the accuracy of the test set reaches 0.990, and the F1-score of each type reaches above 0.98. The model has high prediction accuracy and strong generalization ability.

[0040] (5) The memory chip power consumption classification system of the present invention achieves full automation from thermal imaging data acquisition and feature extraction to model inference through the collaborative work of thermal imaging equipment, processor and memory chip test engine. The single prediction time is in the millisecond level (<10ms), which meets the real-time requirements of chip test production line.

[0041] (6) The memory chip power consumption classification device of the present invention realizes a complete software processing pipeline of “acquisition-feature extraction-reasoning-classification” through the modular design of thermal imaging data receiving module, feature subdivision module, model inference module and classification output module, which is convenient for integration and deployment in hardware devices and software maintenance. Attached Figure Description

[0042] These and / or other aspects and advantages of the present invention will become apparent and readily understood from the following description of preferred embodiments taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 This is a flowchart of the steps of a memory chip power consumption classification method based on thermal imaging recognition according to an embodiment of the present invention;

[0044] Figure 2 According to Figure 1 The diagram illustrates a time window division of an embodiment of the power consumption classification method shown.

[0045] Figure 3 This is a schematic diagram of the principle structure of a memory chip power consumption classification system according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the principle structure of a memory chip power consumption classification device according to an embodiment of the present invention;

[0047] Figure 5 This is the execution result of a script for classifying the heat level of memory chips based on the XGBoost classification model, as an example. Detailed Implementation

[0048] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings. In this specification, the same or similar reference numerals indicate the same or similar components. The following description of the embodiments of the present invention with reference to the accompanying drawings is intended to explain the overall inventive concept of the present invention and should not be construed as a limitation thereof.

[0049] In various embodiments of the present invention, the power consumption classification method and system are mainly targeted at Dynamic Random-Access Memory (DRAM) chips.

[0050] Furthermore, in various embodiments of the present invention, an Extreme Gradient Boosting (XGBoost) decision tree model is used, which is an ensemble learning algorithm based on gradient boosting decision trees. It also involves a core algorithm: GridSearchCV, an algorithm that selects the optimal hyperparameters by traversing preset parameter combinations and using cross-validation.

[0051] See Figure 1 The flowchart illustrates the steps of a memory chip power consumption classification method based on thermal imaging recognition in one embodiment.

[0052] The memory chip power consumption classification method based on thermal imaging identification provided by this invention mainly includes the following steps:

[0053] Step S110: Acquire thermal imaging data, that is, acquire thermal imaging data of the memory chip during the execution of the power consumption test algorithm. The thermal imaging data includes multiple frames of thermal radiation images acquired in a time series.

[0054] Specifically, during the execution of a preset power consumption test algorithm by the memory chip, thermal radiation images of the memory chip surface are continuously acquired using a thermal imaging device such as a thermal imager, forming thermal imaging data frames in a time sequence. For example, the thermal imager has a resolution of 640×512 pixels, a sampling frame rate of 60 frames / second, and an acquisition time of 0.8 seconds, acquiring a total of 48 frames of thermal radiation images (this example will be used as a specific illustration in the following text).

[0055] Step S120: Extract multi-dimensional features, that is, perform temporal and spatial decomposition on the thermal imaging data obtained in step S110 to extract multi-dimensional features, which include temporal features and spatial features.

[0056] Specifically, the multi-frame thermal radiation images acquired in a time series are divided into multiple time windows in the time domain, with each time window corresponding to a response phase of the power consumption test algorithm executed by the memory chip. Then, within each time window, the thermal radiation images are spatially divided, and each frame is divided into multiple image blocks (e.g., 16 quadrants). Finally, temperature statistical features are extracted based on the time windows and image blocks, including the global temperature average, maximum, and standard deviation for each time window; the temperature average, maximum, and standard deviation for each image block; the average, standard deviation, maximum, and minimum values ​​of the overall temperature series; the temperature rise rate and linear fitting slope of the overall temperature series; the frame temperature average for each time window; and the difference between the frame temperature averages of adjacent windows.

[0057] Step S130: Model inference, which involves inputting the multi-dimensional features extracted in step S120 into the trained gradient boosting decision tree (XGBoost) classification model, and the XGBoost classification model outputs the power consumption level of the memory chip.

[0058] Specifically, the XGBoost classification model uses the softmax function as the activation function of the output layer and the cross-entropy loss as the loss function. It outputs the probability distribution of the memory chip belonging to each preset power consumption level (e.g., three preset levels: "low heat", "medium heat" and "high heat"), and uses the power consumption level with the highest probability as the prediction result.

[0059] Step S140: Classification output, that is, classify the memory chips according to the power consumption level output by the XGBoost classification model in step S130, and classify the memory chips into the corresponding power consumption level category.

[0060] In a preferred embodiment, the specific implementation of extracting multi-dimensional features in step S120 is as follows:

[0061] Step S121: Divide the time window (time domain decomposition).

[0062] See Figure 2 The diagram illustrates a time window division of one embodiment.

[0063] like Figure 2 As shown, the 48 frames of thermal radiation images acquired in a time series are divided into multiple time windows, for example... Figure 2 The example divides the time window into four segments, each containing 12 frames of images, which correspond to the various response stages of the power consumption test algorithm executed by the memory chip.

[0064] In some embodiments of the present invention, the number of time windows can be set according to the specific stage division requirements of the power consumption test algorithm. As an example, the number of multiple time windows can be 3, 4, or 5, etc.

[0065] When divided into 3 time windows, these 3 time windows can correspond to each response stage and may include: early response window, mid-term response window, and late response window;

[0066] When divided into 4 time windows, these 4 time windows can correspond to each response stage and may include: early response window, rising response window, mid-term response window, and stable response window;

[0067] When divided into 5 time windows, these 5 time windows can correspond to each response stage and may include: early response window, rising response window, mid-term response window, stable response window, and late response window.

[0068] The aforementioned time window division is synchronized with the timing of the power consumption test algorithm executed by the memory chip, ensuring that the thermal imaging data within each window is precisely aligned with the corresponding response stage in time. Those skilled in the art should understand that the specific time window division method (number of divisions) described above is merely illustrative; in practical applications, the number of time windows and the corresponding response stage for each window can be adaptively adjusted according to the specific number of stages in the power consumption test algorithm.

[0069] The following is Figure 2 The following explanation uses the four time windows shown as examples:

[0070] Window 1 (Early Response Window): Images from frame 1 to frame 12, corresponding to the early response phase after the memory chip power consumption test algorithm is started;

[0071] Window 2 (Rising Response Window): Images from frames 13 to 24, corresponding to the rising power consumption response phase of the memory chip;

[0072] Window 3 (Interim Response Window): Images from frames 25 to 36, corresponding to the intermediate response phase of the memory chip;

[0073] Window 4 (Stable Period Response Window): Images from frame 37 to frame 48, corresponding to the stable period response stage of the memory chip.

[0074] The division into four time windows is synchronized with the timing of the power consumption test algorithm executed by the memory chip, ensuring that the thermal imaging data in each window is precisely aligned with the corresponding response stage in time.

[0075] Step S122: Spatial decomposition.

[0076] Each frame of thermal radiation image within each time window is spatially quadranted, and each frame of 640×512 resolution thermal radiation image is divided into 16 image blocks (quadrants), with each image block being 40×32 pixels in size.

[0077] Step S123: Extract multi-dimensional features.

[0078] Based on the time window defined in step S121 and the spatial image patch defined in step S122, the following multi-dimensional features are extracted, totaling 217 dimensions:

[0079] (1) Window-level global statistical features (12 dimensions): The average temperature, maximum temperature and standard deviation of temperature of all pixels in the four time windows are calculated respectively. Each window is 3-dimensional, for a total of 12 dimensions.

[0080] (2) Spatial fine-grained statistical features (192 dimensions): For each of the four time windows, the average temperature, maximum temperature and standard deviation of the temperature of 16 image blocks are calculated respectively. Each window has 48 dimensions (16 image blocks × 3 statistics), for a total of 192 dimensions.

[0081] (3) Overall temperature sequence characteristics (4-dimensional): Calculate the average temperature, standard deviation of temperature, maximum temperature and minimum temperature of the overall temperature sequence formed by the frame average temperature of all 48 frames of images (i.e. the average temperature of all pixels in each frame of image).

[0082] (4) Temperature rise amplitude and linear fitting slope of the overall temperature sequence (2D): The difference between the average temperature of the overall temperature sequence in window 4 (steady-state response window) and the average temperature of the overall temperature sequence in window 1 (early response window) is used as the temperature rise amplitude; the overall temperature sequence is linearly fitted to obtain the linear fitting slope of temperature change over time.

[0083] (5) Average frame temperature of each time window (4-dimensional): Calculate the average temperature of the 12 frames in each of the four time windows.

[0084] (6) Differences in average frame temperature between adjacent time windows (3D): Calculate the differences in average frame temperature between window 2 and window 1, window 3 and window 2, and window 4 and window 3 respectively.

[0085] In a preferred embodiment, the training method of the gradient boosting decision tree classification model (hereinafter referred to as the XGBoost classification model) is specifically described as follows:

[0086] (a) Acquisition of training samples:

[0087] Thermal imaging data samples of multiple memory chip samples are acquired during the execution of the power consumption test algorithm, along with the corresponding power consumption level labels for each memory chip sample. The power consumption level labels can be calibrated using a traditional resistor + analog-to-digital converter (ADC) scheme, or by labeling a reference chip with a known power consumption level. In this embodiment, for example, the power consumption levels are divided into three categories: 0 represents low heat, 1 represents medium heat, and 2 represents high heat.

[0088] (II) Construction of the training sample set:

[0089] For each thermal imaging data sample, perform temporal and spatial decomposition according to steps S121-S122 above, extract 217-dimensional multi-dimensional features, and construct a power consumption classification training sample set:

[0090] ,in Let i be the 217-dimensional feature vector of the i-th sample. This corresponds to the power consumption level label.

[0091] (III) Model Training and Hyperparameter Optimization and Tuning:

[0092] The XGBoost classifier is trained using a power consumption classification training sample set. XGBoost is an ensemble learning algorithm based on gradient boosting decision trees. This embodiment of the invention selects the XGBClassifier classifier, employing a softmax function combined with cross-entropy loss as the objective function.

[0093] The mathematical expression for a model learning K decision trees is:

[0094] ,in, Let K be the model's final power consumption level prediction for the i-th sample (a probability vector of length 3), where softmax() is the activation function that transforms any real-valued vector into a probability distribution, and K is the total number of decision trees. Let k be the k-th decision tree (Classification and Regression Tree, CART). Let be the input feature vector (217 dimensions) for the i-th sample. The hypothesis space is formed by all possible decision tree functions.

[0095] The expression for the loss function is:

[0096] Where n is the total number of samples in the training set, and c is the category index (c=0 represents low fever, c=1 represents medium fever, and c=2 represents high fever). This is an indicator function (value 1 when the actual label is equal to c, otherwise 0). Let be the probability predicted by the model that the i-th sample belongs to class c. It is the natural logarithm. This is the complexity penalty term for the k-th decision tree. Through gradient boosting iterative optimization, each new tree fits the negative gradient of the previous prediction.

[0097] Furthermore, we employ GridSearchCV in XGBoost combined with Stratified K-Fold cross-validation to optimize the hyperparameters. The specific search space is shown in Table 1.

[0098] Table 1. Meaning of Search Space Parameters and Comparison of Candidate Values

[0099]

[0100] Based on the parameter settings in Table 1, the optimal hyperparameter combination is selected using GridSearchCV based on the average accuracy of 5-fold cross-validation. In one specific embodiment, the optimal hyperparameter combination is:

[0101] `colsample_bytree=0.8` (each tree uses 80% of the features), `learning_rate=0.05` (small learning rate), `max_depth=4` (tree depth is 4 layers), `n_estimators=200` (number of decision trees is 200), `subsample=1.0` (each tree uses all samples). A specific program example is as follows:

[0102] {

[0103] 'colsample_bytree': 0.8, # Use 80% of the features per tree

[0104] 'learning_rate': 0.05, # Small learning rate

[0105] 'max_depth': 4, # Shallow tree depth (4 levels)

[0106] 'n_estimators': 200, # Large number of trees

[0107] 'subsample': 1.0 # Use all samples for each tree

[0108] }

[0109] The combination of shallow trees, small learning rates, and multiple trees forms a robust learning strategy that can effectively prevent overfitting while improving generalization ability.

[0110] The model is trained using thermal imaging data from a large number of memory chips under the above-mentioned optimal hyperparameter configuration. During the training process, each decision tree continuously splits nodes and adjusts leaf weights by minimizing the loss function. At the end of the training, all tree structures and leaf weights are saved as model parameters.

[0111] In a preferred embodiment, the specific implementation for evaluating the XGBoost classification model is as follows:

[0112] After training the best / optimal XGBoost classification model, the performance of the XGBoost classification model is evaluated using an independent test set. Specifically:

[0113] (a) Cross-validation evaluation:

[0114] Five-fold stratified cross-validation was used to evaluate the training set. The training data was randomly divided into five parts, keeping the distribution of each compromise class consistent with the overall distribution. Four parts were used as the training set and one part as the validation set in turn, and this process was repeated five times, with the average accuracy being calculated.

[0115] In one specific embodiment, the average accuracy of 5-fold cross-validation reached 0.972 (±0.018), with a standard deviation of only 1.8%, indicating that the XGBoost classification model performs stably on different training subsets.

[0116] (ii) Test set evaluation:

[0117] The performance of the XGBoost classification model was evaluated using an independent test set (20% of the total samples) that was not used in training.

[0118] In one specific embodiment, 500 samples were taken, including 400 in the training set and 100 in the test set. The accuracy of the test set reached 0.990. The classification report is shown in Table 2.

[0119] Table 2 Classification Report

[0120]

[0121] The precision, recall, and F1-score for the low-heat category are 1.00, 0.97, and 0.99, respectively; for the medium-heat category, the precision, recall, and F1-score are 0.97, 1.00, and 0.98, respectively; and for the high-heat category, the precision, recall, and F1-score are all 1.00. The macro-average F1-score for the three power consumption levels is 0.99, and the weighted average F1-score is also 0.99. This XGBoost classification model demonstrates highly consistent and excellent performance across the three categories.

[0122] (III) Feature Importance Analysis:

[0123] The feature importance distribution of the XGBoost classification model is analyzed. In a specific embodiment, the top 10 important features and their contributions are shown in Table 3:

[0124] Table 3. Importance of Features (Top Ten)

[0125]

[0126] The most important feature, rise_amplitude, contributes approximately 22.32%, while the top four features contribute a combined total of approximately 57.5%, indicating that the model primarily relies on global dynamic features. Individual image patch features have low importance (all below 5%), but multiple features stacked together still have some impact. The feature importance distribution is relatively dispersed, suggesting that the model's decisions depend on multidimensional features rather than a few, demonstrating good robustness.

[0127] In a preferred embodiment, the inference process for power level prediction using the trained XGBoost classification model is implemented as follows:

[0128] During the inference phase, the input features are optimized according to the parameters of the trained model and flow to a leaf node of each decision tree. The weight values ​​stored at that node are retrieved and added together to obtain the cumulative score.

[0129] Taking a specific memory chip sample as an example, assume the model outputs the following three scores: ;

[0130] Corresponding calculation index: ;

[0131] sum: ;

[0132] Probability:

[0133] .

[0134] because maximum, The highest percentage (approximately 52.4%) of this sample was predicted as category 0 (low heat). The XGBoost classification model achieves its classification objective through the aforementioned softmax probability calculation.

[0135] In a preferred embodiment, the trained optimal model is bound to the feature sequence through the model persistence module and saved as a model file for subsequent actual detection and inference.

[0136] Specifically, the best XGBoost classification model object can be serialized and saved as a .pkl file using Python's Pickle or Joblib libraries (e.g., using joblib.dump() or pickle.dump()). During inference, the model file can be loaded using joblib.load(), allowing the model to be directly called to classify and predict new samples.

[0137] Those skilled in the art should understand that the model file saving format is not limited to .pkl format. In actual production, the model can be converted to other formats to facilitate cross-platform or cross-language calls, depending on the specific deployment environment and development language requirements.

[0138] For example, when cross-language deployment (e.g., C++, Java) is required, the model can be saved in .onnx (OpenNeural Network Exchange) format; when integration in a Java environment is required, the model can be saved in .pmml (Predictive Model Markup Language) or .model format; when it needs to be used in a system that supports the PMML standard, the .pmml format can also be chosen. The specific format choice depends on the deployment requirements of the actual application scenario.

[0139] This invention also provides a memory chip power consumption classification system 200 and a memory chip power consumption classification device 300 based on thermal imaging recognition.

[0140] See Figure 3 The diagram illustrates the principle structure of a memory chip power consumption classification system 200 according to one embodiment.

[0141] like Figure 3 As shown, the system includes, but is not limited to, a processor 210, a memory chip testing engine 220, a thermal imaging device 230 (e.g., a thermal imager), and a memory chip under test 240.

[0142] The processor 210, acting as the control center of the memory chip power consumption classification system, sends start / stop signals to the thermal imaging device 230 to control the start and stop of thermal imaging data acquisition. Simultaneously, it sends power consumption test parameters (such as power consumption test algorithms and power consumption test frequencies) to the memory chip test engine 220 and receives power consumption test status feedback from the engine. The processor 210 is also responsible for receiving thermal radiation image data frames transmitted by the thermal imaging device 230, running the feature subdivision module and XGBoost classification model to perform power consumption level inference, and outputting the power consumption classification results.

[0143] The memory chip test engine 220 is communicatively connected to both the processor 210 and the memory chip 240. It sends operation commands to the memory chip 240 based on power consumption test parameters sent by the processor 210, driving the memory chip 240 to execute the power consumption test algorithm. The memory chip test engine 220 is also responsible for monitoring the data status of the memory chip 240 and reporting the power consumption test status back to the processor 210. Power consumption test parameters include, but are not limited to, the command to start the memory chip test engine, the power consumption test algorithm, and the power consumption test frequency.

[0144] The memory chip 240, as the object under test, receives operation commands from the memory chip test engine 220 to operate its internal storage units and feeds back the data status to the memory chip test engine 220. During operation, the memory chip 240 generates heat, which is emitted outward in the form of thermal radiation and detected by the thermal imaging device 230, which then acquires thermal radiation image data frames.

[0145] The thermal imaging device 230 is responsible for detecting and recording thermal changes in the memory chip 240 through a lens, generating thermal radiation image data frames that are transmitted to the processor 210. The thermal imaging device 230 starts or stops acquiring thermal radiation image data frames by receiving start / stop signals from the processor 210.

[0146] In a preferred embodiment, the processor 210 simultaneously sends power consumption test parameters to the memory chip test engine 220 and a start signal to the thermal imaging device 230 to initiate thermal imaging data acquisition, controlling the thermal imaging data acquisition of the thermal imaging device 230 and the execution of the power consumption test algorithm driven by the memory chip test engine 220 to start synchronously in time. This timing synchronization control ensures the time alignment of subsequent feature extraction, guaranteeing that the thermal radiation image acquired by each memory chip corresponds precisely in time to the corresponding response stage (each time window).

[0147] See Figure 4 The diagram illustrates the principle structure of a memory chip power consumption classification device 300 according to one embodiment.

[0148] Combination Figure 3 and Figure 4 As shown, the memory chip power consumption classification device 300 mainly includes the following software functional modules:

[0149] The thermal imaging data receiving module 310 is configured to acquire thermal imaging data from the memory chip 240 during the execution of a power consumption test algorithm. This thermal imaging data includes multiple frames of thermal radiation images acquired in a time sequence. The thermal imaging data receiving module 310 is responsible for data communication with the hardware of the thermal imaging device 230, receiving raw thermal radiation image data frames transmitted by the thermal imaging device 230, and performing data verification and preprocessing.

[0150] The feature subdivision module 320 is configured to perform temporal and spatial decomposition on the thermal imaging data to extract multi-dimensional features, including temporal and spatial features. The specific implementation of this feature subdivision module 320 is described in the embodiment of the memory chip power consumption classification method described above, and will not be repeated here.

[0151] The model inference module 330, which has a built-in trained XGBoost classification model, is configured to receive multi-dimensional features extracted by the feature subdivision module 320 and output the power consumption level of the memory chip 240 from the XGBoost classification model. Specifically, the model inference module 330 includes an XGBoost training and tuning module 331, a model evaluation module 332, and a model persistence module 333, used to execute the above-described embodiment of the memory chip power consumption classification method. In particular, the specific implementation steps regarding the XGBoost classification model training and evaluation are described in the above embodiment description and will not be repeated here.

[0152] The power consumption classification output module 340 is configured to classify the power consumption of the memory chip 240 according to the power consumption level output by the XGBoost classification model and output the power consumption classification result. The power consumption classification output module 340 converts the power consumption level (0 / 1 / 2) predicted by the XGBoost classification model into readable classification labels (low heat / medium heat / high heat) and outputs the classification result for use in subsequent production processes.

[0153] The four modules described above constitute a complete software processing pipeline of "data acquisition, feature extraction, inference, and classification." For example, the four modules can be deployed on the same processor or distributed across different computing nodes.

[0154] In a specific embodiment, the testing procedure for memory chip power consumption levels, based on the memory chip power consumption classification system 200 and the memory chip power consumption classification device 300, is described in detail below:

[0155] Step 1: Configure the processor 210 to use the memory chip test engine 220 to use a special algorithm for testing the power consumption of the memory chip 240, and set the power consumption test parameters, such as the power consumption test algorithm type and power consumption test frequency.

[0156] Step 2: The processor 210 simultaneously starts the memory chip test engine 220 and the thermal imaging device 230 to ensure that the thermal radiation image data frames collected by each memory test chip 240 are consistent in timing.

[0157] Step 3: The memory chip 240 begins executing the power consumption test algorithm driven by the memory chip test engine 220. Simultaneously, the thermal imaging device 230 begins detecting and recording the thermal radiation data of the memory chip 240, forming a thermal radiation image data frame. The execution of the power consumption test algorithm covers four stages: early response, rising response, mid-term response, and steady-state response.

[0158] Step 4: The memory chip test engine 220 finishes executing the power consumption test algorithm, the thermal imaging device 230 stops recording the thermal radiation data of the memory chip 240, and transmits the collected thermal imaging data to the processor 210 in the form of thermal radiation image data frames.

[0159] Step 5: After receiving the thermal imaging data, the processor 210 (thermal imaging data receiving module 310) inputs the thermal imaging data to the feature subdivision module 320. The feature subdivision module 320 decomposes the thermal imaging data into 217-dimensional multi-dimensional features according to the memory chip power consumption classification method of the aforementioned embodiment, and inputs them into the trained XGBoost classification model.

[0160] Step 6: The XGBoost classification model infers based on the input features, calculates the cumulative score based on the weight values ​​of each decision tree leaf node, converts it into a probability distribution through the softmax function, and predicts the power consumption level (low heat / medium heat / high heat) corresponding to the currently tested memory chip 240.

[0161] Step 7: Classify the power consumption of the currently tested memory chip 240 according to the power consumption level output by the XGBoost classification model.

[0162] See Figure 5 The execution result of a memory chip heat level classification script based on the XGBoost classification model is shown in one embodiment.

[0163] A Python script was used to classify the heat dissipation levels of memory chips (simulating power consumption testing). This generated 500 samples, each containing 214 features. The script has slightly fewer than 217 features because the simulation code did not generate all features for simplification. The heat dissipation score (0-100) was calculated using a designed formula. The correlation coefficient between the formula and the actual heat dissipation score reached 0.897, indicating that the features effectively reflect the heat dissipation trend. The category distribution was as follows: low heat (172 samples), medium heat (159 samples), and high heat (169 samples), with a relatively balanced distribution across categories.

[0164] GridSearchCV was used to perform 5-fold cross-validation on 72 parameter combinations, with a total of 360 fits. The optimal parameter combination was: colsample_bytree=0.8, learning_rate=0.05, max_depth=4, n_estimators=200, subsample=1.0.

[0165] The average accuracy of the 5-fold cross-validation was 0.972 (±0.018), and the accuracy on the test set was 0.990. The difference between the cross-validation and test set accuracies was minimal (<2%), indicating no overfitting.

[0166] The classification report shows: for low-heat category, precision is 1.00, recall is 0.97, and F1-score is 0.99; for medium-heat category, precision is 0.97, recall is 1.00, and F1-score is 0.98; for high-heat category, precision is 1.00, recall is 1.00, and F1-score is 1.00.

[0167] The top 10 features by importance and their contributions are shown in Table 3 above. The highest feature importance accounts for 22.32%, indicating that the feature importance is dispersed, which means that the model's decision-making is robust and will not lead to misjudgment due to errors of a single feature.

[0168] The memory chip power consumption classification method, system, and apparatus based on thermal imaging identification provided in the embodiments of the present invention have at least one or a portion of the following advantages:

[0169] (1) By acquiring thermal imaging data of the memory chip during the execution of the power consumption test algorithm and performing temporal and spatial decomposition, and then inputting it into the gradient boosting decision tree classification model, the power consumption level of the memory chip can be automatically classified in a non-contact manner, avoiding the parasitic parameters and interference introduced by the series sampling resistor on the power supply line, and improving the stability and reliability of the detection system.

[0170] (2) By dividing thermal imaging data into multiple time windows (e.g., early response window, rising response window, mid-term response window and stable response window) according to time series, and dividing the thermal radiation image in each time window into spatial quadrants, multi-dimensional features including global temperature statistics of each window, temperature statistics of each image block, overall temperature sequence statistics, temperature rise amplitude and linear fitting slope can be extracted to comprehensively characterize the heating characteristics of memory chips in different response stages. The importance of features is dispersed (the highest proportion is about 22%), does not depend on a few features, and has good decision robustness.

[0171] (3) By controlling the acquisition of thermal imaging data and the execution of power consumption test algorithm to start synchronously in time, the frames of thermal imaging data and the response stages of memory chip are aligned in time, ensuring the time correspondence of subsequent feature extraction and the accuracy of power consumption classification results.

[0172] (4) The gradient boosting decision tree classification model uses the softmax function as the output layer activation function and cross-entropy loss as the loss function. The hyperparameters are optimized by grid search (GridSearchCV) combined with 5-fold hierarchical cross-validation. The average accuracy of 5-fold cross-validation reaches 0.972 (±0.018), the accuracy of the test set reaches 0.990, and the F1-score of each type reaches above 0.98. The model has high prediction accuracy and strong generalization ability.

[0173] (5) The memory chip power consumption classification system of the present invention achieves full automation from thermal imaging data acquisition and feature extraction to model inference through the collaborative work of thermal imaging equipment, processor and memory chip test engine. The single prediction time is in the millisecond level (<10ms), which meets the real-time requirements of chip test production line.

[0174] (6) The memory chip power consumption classification device of the present invention realizes a complete software processing pipeline of “acquisition-feature extraction-reasoning-classification” through the modular design of thermal imaging data receiving module, feature subdivision module, model inference module and classification output module, which is convenient for integration and deployment in hardware devices and software maintenance.

[0175] While some embodiments of the present general inventive concept have been shown and described, those skilled in the art will understand that changes may be made to these embodiments without departing from the principles and spirit of the present general inventive concept, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for classifying the power consumption of memory chips based on thermal imaging recognition, characterized in that, The memory chip power consumption classification method includes: Acquire thermal imaging data of the memory chip during the execution of a power consumption test algorithm, the thermal imaging data including multiple frames of thermal radiation images acquired in a time sequence; The thermal imaging data is decomposed in the temporal domain and in the spatial domain to extract multi-dimensional features, which include temporal features and spatial features. The extracted multi-dimensional features are input into a trained gradient boosting decision tree classification model, which then outputs the power consumption level of the memory chip. The memory chip is classified by power consumption level, and the power consumption classification result is output; wherein... The step of performing temporal and spatial decomposition on the thermal imaging data to extract multi-dimensional features includes: The multi-frame thermal radiation images acquired in a time series are divided into multiple time windows, each time window corresponding to a response phase of the power consumption test algorithm executed by the memory chip; The thermal radiation image within each time window is spatially divided, and each frame of thermal radiation image is divided into multiple image blocks. The multi-dimensional features are extracted based on the time window and the image patch; The multidimensional features include at least one of the following features or any combination thereof: Window-level global statistical features, which include the average temperature, maximum temperature, and standard deviation of all pixels within each time window; Spatial fine-grained statistical features, which include the average temperature, maximum temperature, and standard deviation of the temperature for each image patch in each time window; The temperature mean and temperature standard deviation of the overall temperature series formed by all the aforementioned time windows; The overall temperature sequence features include the average temperature, standard deviation of temperature, maximum temperature, and minimum temperature of all frames of the thermal radiation image. The overall temperature sequence shows the temperature rise rate and the slope of the linear fit of the temperature change over time. The average frame temperature for each of the aforementioned time windows; The difference between the average frame temperatures of adjacent time windows.

2. The memory chip power consumption classification method according to claim 1, characterized in that, The acquisition of thermal imaging data by the memory chip during the execution of the power consumption test algorithm is synchronized in time with the execution of the power consumption test algorithm by the memory chip, so that each frame of the thermal imaging data is aligned in time with each response stage of the memory chip.

3. The memory chip power consumption classification method according to claim 2, characterized in that, The multiple time windows can be 3 time windows, 4 time windows, or 5 time windows; among them, When the multiple time windows are three time windows, the three time windows include an early response window, a mid-term response window, and a late response window; When the multiple time windows are four time windows, the four time windows include an early response window, a rising response window, a mid-term response window, and a stable response window; When the multiple time windows consist of 5 time windows, the 5 time windows include an early response window, a rising response window, a mid-term response window, a stable response window, and a late response window.

4. The memory chip power consumption classification method according to any one of claims 1-3, characterized in that, The gradient boosting decision tree classification model uses the softmax function as the activation function of the output layer and the cross-entropy loss as the loss function. It outputs the probability distribution of the memory chip belonging to each preset power consumption level, and uses the power consumption level with the highest probability as the power consumption level prediction result.

5. The memory chip power consumption classification method according to claim 4, characterized in that, The gradient boosting decision tree classification model was pre-trained as follows: Acquire thermal imaging data samples of multiple memory chip samples during the execution of the power consumption test algorithm, as well as the power consumption level label corresponding to each memory chip sample; The thermal imaging data samples are decomposed in the temporal domain and spatial domain to extract multi-dimensional features and construct a power consumption classification training sample set. The gradient boosting decision tree classifier is trained using the power consumption classification training sample set, and the hyperparameters of the gradient boosting decision tree classifier are optimized and adjusted using grid search combined with cross-validation to obtain the gradient boosting decision tree classification model.

6. The memory chip power consumption classification method according to claim 5, characterized in that, The hyperparameters include at least one of the following, or any combination thereof: number of decision trees, maximum tree depth, learning rate, sample sampling ratio, and feature sampling ratio.

7. A memory chip power consumption classification system based on thermal imaging recognition, characterized in that, The memory chip power consumption classification system includes: A thermal imaging device configured to acquire thermal imaging data of a memory chip during the execution of a power consumption test algorithm, the thermal imaging data including multiple frames of thermal radiation images acquired in a time sequence. A processor, communicatively connected to the thermal imaging device, is configured to perform the following steps: Acquire the thermal imaging data collected by the thermal imaging device; The thermal imaging data is decomposed in the temporal domain and in the spatial domain to extract multi-dimensional features, which include temporal features and spatial features. The extracted multi-dimensional features are input into a trained gradient boosting decision tree classification model, which then outputs the power consumption level of the memory chip. The memory chip is classified by power consumption level, and the power consumption classification result is output; wherein... The processor performs temporal and spatial decomposition on the thermal imaging data, extracting multi-dimensional features including: The multi-frame thermal radiation images acquired in a time series are divided into multiple time windows, each time window corresponding to a response phase of the power consumption test algorithm executed by the memory chip; The thermal radiation image within each time window is spatially divided, and each frame of thermal radiation image is divided into multiple image blocks. The multi-dimensional features are extracted based on the time window and the image patch; The multidimensional features include at least one of the following features or any combination thereof: Window-level global statistical features, which include the average temperature, maximum temperature, and standard deviation of all pixels within each time window; Spatial fine-grained statistical features, which include the average temperature, maximum temperature, and standard deviation of the temperature for each image patch in each time window; The temperature mean and temperature standard deviation of the overall temperature series formed by all the aforementioned time windows; The overall temperature sequence features include the average temperature, standard deviation of temperature, maximum temperature, and minimum temperature of all frames of the thermal radiation image. The overall temperature sequence shows the temperature rise rate and the slope of the linear fit of the temperature change over time. The average frame temperature for each of the aforementioned time windows; The difference between the average frame temperatures of adjacent time windows.

8. The memory chip power consumption classification system according to claim 7, characterized in that, The memory chip power consumption classification system also includes: A memory chip testing engine is communicatively connected to both the processor and the memory chip, and is configured to send operation commands to the memory chip based on power consumption test parameters sent by the processor, thereby driving the memory chip to execute the power consumption test algorithm. The processor is also configured to send a start signal to the thermal imaging device and send the power consumption test parameters to the memory chip test engine, so as to control the thermal imaging data acquisition of the thermal imaging device and the execution of the power consumption test algorithm driven by the memory chip test engine by the memory chip to start synchronously in time.

9. The memory chip power consumption classification system according to claim 8, characterized in that, The multiple time windows can be 3 time windows, 4 time windows, or 5 time windows; among them, When the multiple time windows are three time windows, the three time windows include an early response window, a mid-term response window, and a late response window; When the multiple time windows are four time windows, the four time windows include an early response window, a rising response window, a mid-term response window, and a stable response window; When the multiple time windows consist of 5 time windows, the 5 time windows include an early response window, a rising response window, a mid-term response window, a stable response window, and a late response window.

10. The memory chip power consumption classification system according to any one of claims 7-9, characterized in that, The processor is also configured to pre-train the gradient boosting decision tree classification model, including: Acquire thermal imaging data samples of multiple memory chip samples during the execution of the power consumption test algorithm, as well as the power consumption level labels corresponding to the multiple memory chip samples; The thermal imaging data samples are decomposed in the temporal domain and spatial domain to extract multi-dimensional features and construct a power consumption training sample set. The gradient boosting decision tree classifier is trained using the power consumption training sample set, and the hyperparameters of the gradient boosting decision tree classifier are optimized and adjusted using grid search combined with cross-validation to obtain the gradient boosting decision tree classification model.

11. A memory chip power consumption classification device based on thermal imaging recognition, wherein the memory chip power consumption classification device is integrated into the memory chip power consumption classification system according to any one of claims 7-10, characterized in that, The memory chip power consumption classification device includes: A thermal imaging data receiving module is configured to acquire thermal imaging data of a memory chip during the execution of a power consumption test algorithm, the thermal imaging data including multiple frames of thermal radiation images acquired in a time sequence. The feature segmentation module is configured to perform temporal and spatial decomposition on the thermal imaging data and extract multi-dimensional features, including temporal and spatial features. The model inference module has a built-in trained gradient boosting decision tree classification model. The model inference module receives the multi-dimensional features extracted by the feature subdivision module and outputs the power consumption level of the memory chip by the gradient boosting decision tree classification model. A power consumption classification output module is configured to classify the power consumption of the memory chip according to the power consumption level and output the power consumption classification result.

Images