Method and system for detecting and analyzing performance of integrated circuit mainboard

Abstract

The invention relates to the technical field of integrated circuit testing, and discloses a performance detection analysis method and system for an integrated circuit mainboard. The method comprises the following steps: acquiring and cleaning historical voltage, current and temperature data, and generating a clean historical data set; evaluating parameter fluctuation of the data set, and calculating to obtain stability estimated values of voltage, current and temperature; the estimated values are input into a pre-trained detection period recommendation network, and the network outputs customized detection frequency by comparing with a historical mode library; real-time monitoring is deployed based on the frequency, instantaneous voltage, current and temperature readings and signal quality indexes are collected, and a real-time data set is assembled; and performing multi-dimensional analysis on the real-time data set by utilizing a pattern recognition engine, dividing performance state categories of the mainboard, and outputting an evaluation label. According to the invention, adaptive optimization of the detection frequency and intelligent diagnosis of the performance state are realized, and the detection efficiency and the early warning accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated circuit testing, in particular to a performance detection and analysis method and system for an integrated circuit mainboard. BACKGROUND

[0002] Currently, performance detection and health state monitoring of integrated circuit mainboards generally rely on fixed-period detection strategies. The monitoring system collects basic parameters such as voltage, current and temperature at preset time intervals, and compares the collected real-time data with preset fixed safety thresholds. Once a parameter exceeds the upper or lower threshold, the system triggers an alarm. This time-fixed, threshold-fixed monitoring method is a widely used basic means in the industry.

[0003] The fixed-period detection mode has obvious defects. For mainboards with stable performance and gentle load, too frequent detection will waste computing and storage resources and increase unnecessary monitoring costs. For mainboards in complex and variable working conditions or mainboards whose performance has begun to deteriorate slightly, a fixed long detection period may fail to capture key transient abnormalities or early signs of performance degradation in a timely manner, causing early warning delays. Fixed threshold alarms can only reflect the instantaneous overrun of parameters and cannot assess the fluctuation trend of parameters within the normal range and the long-term risks it indicates, lacking the ability to predict the trend of the health state of the mainboard.

[0004] Existing alarm mechanisms are usually limited to threshold judgments of a single parameter and fail to comprehensively analyze multiple parameters such as voltage, current, temperature, etc. Conventional techniques rarely include signal quality, a key indicator directly reflecting circuit timing and logic correctness, in the scope of real-time monitoring and analysis. This isolated, single-point judgment method cannot identify complex fault modes caused by subtle changes in multiple parameters. The output result is only a binary judgment of "normal" or "overrun", which cannot classify the performance state of the mainboard more finely. SUMMARY

[0005] The present application aims to provide a performance detection and analysis method and system for an integrated circuit mainboard to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a performance detection and analysis method for an integrated circuit mainboard, which comprises:

[0007] Deriving voltage sampling data, current sampling data and temperature sampling data in a historical time range from a monitoring system of the integrated circuit mainboard, performing integrity verification and outlier rejection on the voltage sampling data, current sampling data and temperature sampling data, and generating a clean historical data set;

[0008] performing parameter fluctuation evaluation on the clean historical data set, generating voltage stability evaluation, current stability evaluation and temperature stability evaluation by calculating oscillation amplitude of voltage data, drift range of current data and fluctuation degree of temperature data;

[0009] inputting the voltage stability evaluation, current stability evaluation and temperature stability evaluation into a pre-trained detection cycle recommendation network, the network outputting a customized detection frequency by comparing a historical pattern library;

[0010] deploying a real-time monitoring process according to the customized detection frequency, capturing instantaneous voltage readings, instantaneous current readings, instantaneous temperature readings and signal quality indicators of the integrated circuit motherboard, and assembling a real-time data set;

[0011] performing multi-dimensional analysis on the real-time data set using a pattern recognition engine, dividing performance state categories, and outputting performance evaluation labels.

[0012] Preferably, the integrity check and outlier removal of the voltage sampling data, current sampling data and temperature sampling data include:

[0013] checking the time stamp continuity of the voltage sampling data, current sampling data and temperature sampling data, and identifying data points corresponding to missing time stamps;

[0014] for the data points with missing time stamps, performing linear interpolation filling using adjacent time stamp data points to form a continuous data sequence;

[0015] calculating the deviation of each data point in the continuous data sequence from adjacent data points, and if the deviation exceeds a preset tolerance threshold, marking it as an outlier and removing it;

[0016] performing normalization processing on the data sequence after removing outliers to unify the scales of all data, obtaining a clean historical data set.

[0017] Preferably, the parameter fluctuation evaluation on the clean historical data set includes:

[0018] extracting the voltage data sequence in the clean historical data set, calculating the average value of the absolute value of the difference between adjacent sampling points in the voltage data sequence as the oscillation amplitude;

[0019] extracting the current data sequence in the clean historical data set, calculating the difference between the maximum value and the minimum value of the current data sequence divided by the average value as the drift range;

[0020] extracting the temperature data sequence in the clean historical data set, calculating the ratio of the standard deviation to the average value of the temperature data sequence as the fluctuation degree;

[0021] The oscillation amplitude, the drift range and the fluctuation degree are mapped to the range of zero to one respectively to obtain the voltage stability estimate, the current stability estimate and the temperature stability estimate.

[0022] Preferably, the pre-trained detection cycle recommendation network is constructed by the following way:

[0023] A plurality of sets of historical voltage stability estimates, historical current stability estimates and historical temperature stability estimates, and corresponding historical detection cycle data are collected to form a training sample set;

[0024] A three-layer neural network structure is constructed, the input layer includes three nodes corresponding to the voltage stability estimate, the current stability estimate and the temperature stability estimate respectively, and the output layer is one node corresponding to the detection cycle;

[0025] The neural network is supervised learning using the training sample set, and the network weight is adjusted through the back propagation algorithm until the prediction error is lower than the set threshold;

[0026] The trained network parameters and structure are saved as the detection cycle recommendation network.

[0027] Preferably, the input of the voltage stability estimate, the current stability estimate and the temperature stability estimate into the pre-trained detection cycle recommendation network includes:

[0028] The voltage stability estimate, the current stability estimate and the temperature stability estimate are combined into an input vector;

[0029] The input vector is fed into the input layer of the detection cycle recommendation network, and the weighted sum is calculated through the hidden layer node and the activation function is applied;

[0030] The output layer node calculates the final value, and maps the final value to the actual detection cycle range to obtain the customized detection frequency.

[0031] Preferably, the deployment of the real-time monitoring process includes:

[0032] A timer is set according to the customized detection frequency to trigger a data collection event;

[0033] When the timer triggers, the voltage sensor reading, the current sensor reading, the temperature sensor reading and the jitter index output by the signal analyzer of the integrated circuit mainboard are synchronously read;

[0034] Each time the reading is marked with a time stamp and stored as a real-time data record;

[0035] The real-time data records at multiple time points are accumulated to assemble a real-time data set.

[0036] Preferably, the multi-dimensional analysis of the real-time data set by the pattern recognition engine includes:

[0037] extracting a voltage reading sequence, a current reading sequence, a temperature reading sequence and a signal quality indicator sequence from the real-time data set;

[0038] calculating statistical features of each sequence, including mean, variance, skewness and kurtosis, forming a feature vector;

[0039] inputting the feature vector into a pre-trained classification model, which is based on decision tree rules to divide states;

[0040] outputting the classification result as a performance evaluation label.

[0041] Preferably, the pre-trained classification model is trained by the following process:

[0042] collecting historical real-time data sets and their corresponding known performance state labels as training data sets;

[0043] randomly extracting subsets from the training data sets, constructing multiple decision trees, each using a different subset of features;

[0044] integrating the voting results of multiple decision trees to determine the final classification rule;

[0045] saving the decision tree parameters and voting mechanism as a classification model.

[0046] Preferably, the calculation of statistical features of each sequence includes:

[0047] For the voltage reading sequence, the arithmetic mean of all voltage values is calculated as the mean, the average square of the deviation of the voltage value from the mean is calculated as the variance, and the asymmetry of the voltage value distribution is calculated as the skewness, and the sharpness of the voltage value distribution is calculated as the kurtosis;

[0048] The same calculation is performed on the current reading sequence and the temperature reading sequence to obtain their respective mean, variance, skewness and kurtosis;

[0049] For the signal quality indicator sequence, the average value and the coefficient of variation of the jitter value are calculated.

[0050] Preferably, the present application further includes a performance detection and analysis system for an integrated circuit mainboard, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the performance detection and analysis method for the integrated circuit mainboard as described above.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] Based on the parameter fluctuation evaluation and detection cycle recommendation network, the detection frequency is fundamentally changed from fixed mode to dynamic self-adaptation. First, the historical operation data is quantitatively analyzed, the voltage oscillation amplitude, current drift range and temperature fluctuation degree are calculated, and accurate stability evaluation is generated. These evaluations are input to the trained network, which can identify the running stability characteristics of the current mainboard by comparing the built-in historical mode library. This mechanism makes the detection frequency no longer fixed, but personalized according to the stability level actually shown by the mainboard. For the mainboard that runs smoothly for a long time, the system will automatically recommend a lower detection frequency, effectively saving computing, storage and communication resources. For the mainboard with specific fluctuation mode or showing potential instability trend, the system will recommend a higher detection frequency, so as to capture data more densely and realize early warning of performance degradation or potential failure. This data-driven intelligent scheduling mechanism improves the resource utilization efficiency of the whole monitoring system while ensuring the monitoring coverage.

[0053] By introducing signal quality indicators and applying pattern recognition engine for multi-dimensional analysis, the performance evaluation is upgraded from single parameter threshold judgment to comprehensive state classification. The real-time monitoring process not only captures the conventional voltage, current and temperature instantaneous readings, but also specially integrates the synchronous collection of signal quality indicators, building a more comprehensive real-time data set reflecting the actual working state of the mainboard. The pattern recognition engine conducts comprehensive analysis on the data set, which can mine the deep correlation and complex patterns between different parameters. This analysis capability enables it to identify abnormal states predicted by the coordinated changes of multiple parameters, although a single parameter is within the normal range. The system output is a classification label of the overall health status of the mainboard, rather than a simple binary normal / abnormal judgment. This fine-grained state classification provides more rich and accurate decision basis for predictive maintenance, so that maintenance actions can be arranged based on the urgency of the state. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The working principle diagram of the performance detection and analysis method of the integrated circuit mainboard described in the application;

[0055] Figure 2 The flowchart for data integrity check and outlier rejection;

[0056] Figure 3 The flowchart for parameter fluctuation evaluation;

[0057] Figure 4 The comparison diagram of the integrated circuit mainboard multi-parameter stability evaluation;

[0058] Figure 5 The signal quality jitter index statistical analysis diagram of the integrated circuit mainboard. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figure 1 This invention provides a performance testing and analysis method for integrated circuit motherboards. The method includes: exporting voltage sampling data, current sampling data, and temperature sampling data collected within a historical time range from the monitoring system of the integrated circuit motherboard; performing integrity verification and outlier removal on these sampling data to generate a clean historical data set; performing parameter fluctuation assessment on the clean historical data set, specifically by calculating the oscillation amplitude of voltage data, the drift range of current data, and the fluctuation degree of temperature data to generate quantified voltage stability estimates, current stability estimates, and temperature stability estimates; feeding these stability estimates as input into a pre-trained detection cycle recommendation network, which outputs a customized detection frequency for the current motherboard state by comparing with an internally stored historical pattern library; deploying a real-time monitoring process based on this customized detection frequency, which captures instantaneous voltage readings, instantaneous current readings, instantaneous temperature readings, and signal quality indicators of the integrated circuit motherboard at a set frequency, and assembling these readings into a real-time data set; and using a pattern recognition engine to perform multi-dimensional analysis on the real-time data set, classifying the motherboard's performance state according to the analysis results, and outputting corresponding performance evaluation labels, thereby completing a comprehensive testing and analysis of the motherboard performance.

[0061] Example 1: See Figure 2In practical implementation, the data preprocessing stage of the integrated circuit motherboard performance testing and analysis method involves rigorous integrity verification and outlier removal of voltage, current, and temperature sampling data exported from the monitoring system. The first step in integrity verification is to check whether the timestamps of all sampled data remain continuous. When the monitoring system collects data within a historical time range, some timestamps may be missing due to temporary equipment failures or communication interference. The system scans the timestamp intervals of the entire data sequence, identifies the data positions corresponding to the missing timestamps, and adds specific markers to these positions. In some embodiments, for the identified missing timestamp data points, a linear interpolation algorithm is used to fill them. The linear interpolation algorithm uses the effective data point values ​​of the adjacent timestamps before and after the missing point to calculate the linear estimate between the two and insert it into the missing position, thereby forming a continuous and uninterrupted voltage, current, and temperature data sequence in the time dimension. It can be understood that linear interpolation can better maintain the local trend of the data and avoid introducing drastic fluctuations. After interpolation, the data sequence achieves complete coverage on the time axis.

[0062] In practice, the outlier removal process is performed on continuous voltage, current, and temperature data sequences. The deviation of each data point from its neighboring data points within a certain window is calculated. This deviation is typically calculated using absolute difference or standard deviation methods. If the calculated deviation exceeds a preset tolerance threshold for that type of parameter, the data point is considered an outlier and removed from the sequence. Optionally, the tolerance threshold can be dynamically adjusted based on historical data statistical characteristics, such as by setting it based on moving averages or standard deviation multiples. After removing outliers, gaps may remain in the data sequence, but these gaps are not subject to secondary interpolation to maintain data accuracy. In some embodiments, the data sequence after outlier removal needs to be normalized. Normalization scales the voltage, current, and temperature data to the same numerical range, such as zero to one. The scaling formula uses a minimum-maximum normalization method, mapping the minimum value of each sequence to zero, the maximum value to one, and the intermediate values ​​linearly, thus unifying all data scales and eliminating dimensional effects. Understandably, normalization helps subsequent fluctuation assessment algorithms treat data of different physical quantities fairly, avoids certain parameters from dominating the assessment results due to their large values, and ultimately provides a clean historical data set to provide high-quality input for parameter fluctuation assessment.

[0063] In practical implementation, the timestamp continuity check for integrity verification involves parsing the timestamp information attached to each sampled data point. Timestamps are typically stored in Unix timestamp or standard date and time format. The system sorts all data points chronologically and calculates the interval between adjacent timestamps. If the interval exceeds the sampling period threshold, a missing data point is identified. Optionally, the sampling period threshold is determined based on the monitoring system's preset sampling frequency. For example, if the sampling frequency is once per second, the threshold can be set to 1.5 seconds. Data points exceeding the threshold are considered missing. After identifying the data point corresponding to the missing timestamp, the system records the missing location index for subsequent processing. When filling missing timestamp data points using linear interpolation, the algorithm needs to obtain the data values ​​of the preceding and following valid timestamps. Then, it performs a linear weighted calculation based on the relative position of the missing timestamp within the preceding and following timestamps. The filled value equals the preceding value plus the product of the difference between the preceding and following values ​​and the time ratio, thus ensuring a smooth transition in the filled data. The deviation calculation in the outlier removal phase can employ a sliding window method. The window size is set according to the data characteristics; for example, the window might include five points before and after the current point. The average or maximum deviation between the current point and its neighboring points within the window is calculated. If the deviation exceeds a tolerance threshold, it is marked as an outlier. The tolerance threshold might be set to 5% of the rated voltage for voltage data, 10% of the rated current for current data, and 15% of the allowable temperature rise for temperature data. Removing outliers may shorten the sequence length but improve data quality. The normalization phase requires calculating the minimum and maximum values ​​for the voltage, current, and temperature data sequences separately. Then, a scaling formula is applied to each data point. After scaling, all data points fall within the zero-to-one range, making the voltage, current, and temperature sampled data numerically comparable. This clean historical data set is thus generated and stored for subsequent analysis.

[0064] In practice, the integrity verification and outlier removal process can be accomplished by writing a dedicated data cleaning module. This module reads the original voltage sampling data, current sampling data, and temperature sampling data files, parses the data format, and performs the above steps. The module outputs a clean historical data collection file or database records.

[0065] Example 2: See Figure 3In practical implementation, the parameter fluctuation assessment phase analyzes the cleanroom historical data set to quantify the stability of voltage, current, and temperature signals. The assessment process calculates specific fluctuation indices for voltage, current, and temperature data sequences respectively. For the voltage data sequence, the oscillation amplitude index is calculated. The oscillation amplitude is obtained by calculating the absolute value of the difference between all adjacent sampling points in the voltage data sequence, and then averaging these absolute values. This average value reflects the average intensity of voltage change within a unit sampling interval. In some embodiments, the calculation of the oscillation amplitude of the voltage data sequence requires traversing every position in the sequence from the first sampling point to the second-to-last sampling point, calculating the absolute value of the difference between the value of the corresponding sampling point at each position and the value of the next sampling point, summing all absolute values, and dividing by the total number of absolute values ​​to obtain the oscillation amplitude value. It can be understood that the smaller the oscillation amplitude value, the more stable the voltage data sequence; the larger the value, the more severe the voltage fluctuation.

[0066] In practical implementation, for current data sequences, the drift range index is evaluated. Calculating the drift range requires identifying the maximum and minimum values ​​in the current data sequence and calculating their difference. The current data sequence is a sequence of current sampled values ​​extracted from a clean historical data set and arranged in chronological order. The maximum value is found by traversing the entire current data sequence. A variable is initialized to store the current maximum value, initially set to the first element of the current data sequence. Then, starting from the second element, each element is compared with the current maximum value. If the current element value is greater than the current maximum value, the current maximum value is updated to that element value. After traversal, the current maximum value is the maximum value of the current data sequence. The minimum value is found using a similar method. The current minimum value is initialized to the first element of the current data sequence. Then, subsequent elements are compared sequentially. If the current element value is less than the current minimum value, the current minimum value is updated to that element value. After traversal, the minimum value of the current data sequence is obtained. The difference between the maximum and minimum values ​​is calculated by subtracting the minimum value from the maximum value, yielding the range. This range reflects the absolute range of change of the current data sequence throughout the entire observation period. The current maximum and minimum values ​​are initialized as the first elements of the current data sequence. Then, starting from the second element, the sequence iterates. In each iteration, the current element is compared with the current maximum value; if greater, the current maximum value is updated. Then, the current element is compared with the current minimum value; if less, the current minimum value is updated. This synchronous traversal method requires only one linear scan of the current data sequence, with a time complexity of O(n), where n is the length of the current data sequence. The range is calculated after the traversal is complete, ensuring numerical stability and avoiding floating-point calculation errors. Optionally, for particularly long current data sequences, a divide-and-conquer strategy can be used to find the maximum and minimum values ​​in parallel. The current data sequence is divided into multiple subsequences, and the local maximum and local minimum values ​​of each subsequence are found in different computational units. All local maximum values ​​are compared to obtain the global maximum, and all local minimum values ​​are compared to obtain the global minimum. The range is still calculated by subtracting the global minimum from the global maximum. This method is suitable for distributed computing environments or scenarios requiring the processing of massive historical data.

[0067] In practice, after calculating the range, it needs to be divided by the arithmetic mean of the current data sequence to obtain the drift range. The arithmetic mean is calculated by summing all elements of the current data sequence and dividing by the number of elements. The summation operation is achieved by traversing the current data sequence and accumulating all current values, then dividing by the sequence length n to obtain the average value. When calculating the range divided by the average, it is necessary to check whether the average value is zero. If the average value is close to zero, the drift range may be meaningless, requiring special handling, such as directly using the range or marking the data as abnormal. As a dimensionless relative indicator, the drift range can eliminate the influence of different current reference values ​​on fluctuation judgment, allowing for comparison of motherboard current stability under different operating currents. It is understandable that this drift range calculation method based on the range and average value is simple, intuitive, and computationally efficient, making it suitable for integration into automated performance testing and analysis processes.

[0068] For temperature data series, the degree of fluctuation is evaluated using an index. This fluctuation is calculated by dividing the standard deviation of the temperature data series by its mean; this ratio, known as the coefficient of variation, reflects the relative dispersion of temperature values ​​around the mean. Calculating the standard deviation of a temperature data series involves first obtaining the mean, then calculating the square of the difference between each temperature sample and the mean, averaging these squares, taking the square root, and finally dividing the standard deviation by the mean to obtain the fluctuation index. In essence, the fluctuation index eliminates the interference of a high or low temperature baseline on the judgment of fluctuation, allowing for comparison of fluctuations at different average temperatures.

[0069] In practical implementation, the calculated oscillation amplitude, drift range, and fluctuation degree are all numerical values ​​with original physical dimensions. They need to be mapped to a zero-to-one range to generate standardized stability estimates. The mapping process uses a minimum-maximum normalization method. The voltage stability estimate is generated by mapping the voltage oscillation amplitude value to a zero-to-one range. During mapping, a reasonable range for the voltage oscillation amplitude needs to be determined based on historical data or prior knowledge. The minimum value within the range is mapped to zero, the maximum value to one, and intermediate values ​​are linearly interpolated. The current stability estimate is generated by mapping the current drift range value to a zero-to-one range, similarly using minimum-maximum normalization based on the possible value range of the current drift range. The temperature stability estimate is generated by mapping the temperature fluctuation degree value to a zero-to-one range, determining the mapping interval based on the historical statistical characteristics of the temperature fluctuation degree. Optionally, the mapping interval can be determined based on the minimum and maximum values ​​of each indicator obtained from a large amount of historical data, or theoretical limit values ​​can be set as boundaries according to circuit design specifications. The calculation of voltage stability estimates, current stability estimates, and temperature stability estimates can be implemented through a separate fluctuation assessment module. This module receives a clean historical dataset as input and processes the voltage, current, and temperature data sequences respectively. The fluctuation assessment module contains three parallel processing units, each specifically calculating the oscillation amplitude, drift range, and fluctuation degree. Then, a normalization unit maps the three index values ​​to the zero-to-one range. It can be understood that the closer the stability estimate is to zero, the more unstable the parameter; the closer it is to one, the more stable the parameter. This normalization process allows the stability of different physical quantities to be compared and integrated on the same scale. The output of the fluctuation assessment module is three values ​​within the zero-to-one range, which serve as the voltage stability estimate, current stability estimate, and temperature stability estimate, respectively. These estimates will be used as input features for the detection cycle recommendation network.

[0070] In practice, oscillation amplitude calculation involves differential operations on the voltage data sequence. The differential operation yields a sequence of absolute values ​​of the differences between adjacent points. Then, the arithmetic mean of this absolute value sequence is calculated; this average is the original value of the oscillation amplitude. Drift range calculation requires extreme value searching on the current data sequence to find the global maximum and minimum values. The range is calculated, and then divided by the average to obtain the original value of the drift range. Fluctuation degree calculation requires calculating the standard deviation of the temperature data sequence. The standard deviation is the positive square root of the variance, which is the average of the squares of the differences between each temperature sample and the average. Finally, the standard deviation is divided by the average to obtain the original value of the fluctuation degree. These three original values ​​are converted into voltage stability estimates, current stability estimates, and temperature stability estimates within the range of zero to one using normalization functions. The normalization functions require preset theoretical minimum and maximum values ​​for each indicator; these boundary values ​​can be pre-configured according to circuit characteristics. The output of the fluctuation assessment stage is three standardized estimates that comprehensively reflect the voltage stability, current stability, and temperature stability performance of the integrated circuit motherboard during the historical observation period.

[0071] See Figure 4 This figure is a key visualization result of the parameter fluctuation assessment stage in the integrated circuit motherboard performance testing and analysis method. The horizontal axis represents the motherboard number, and the vertical axis represents the stability estimate; in the legend, the blue, cyan, and yellow bars correspond to voltage stability, current stability, and temperature stability, respectively. Technically, these stability estimates are the core inputs to the testing cycle recommendation network. Motherboards with lower stability require higher testing frequencies to avoid performance degradation or delayed fault warnings; motherboards with higher stability can have lower testing frequencies, saving resources. This figure provides an intuitive decision-making basis for adaptive testing cycles by quantitatively comparing the stability differences of different motherboards in the voltage, current, and temperature dimensions. It is a crucial link in achieving a breakthrough from fixed-cycle testing to data-driven dynamic testing technology, effectively solving the shortcomings of fixed-cycle testing in the background technology, such as resource waste or untimely warnings.

[0072] Example 3: In specific implementation, the construction of the detection cycle recommendation network is a supervised learning process, requiring the collection of a large amount of historical data as training samples. The training samples include multiple sets of historical voltage stability estimates, historical current stability estimates, and historical temperature stability estimates, as well as historical detection cycle data corresponding to each set of estimates that have been proven reasonable in practice. These historical detection cycle data represent the optimal detection interval under specific stability states, collectively constituting the training sample set. Each sample in the training sample set is a quadruple, containing three input feature values ​​(historical voltage stability estimate, historical current stability estimate, and historical temperature stability estimate) and a target output value (historical detection cycle data). The number of samples needs to be sufficiently large to cover various possible combinations of stability states. In some embodiments, the training samples can be collected from monitoring records of multiple long-running integrated circuit motherboards. The appropriate detection cycle for each historical time period is determined through expert experience or offline analysis, forming labeled data.

[0073] In implementation, the network structure employs a three-layer feedforward neural network. The input layer consists of three nodes, each receiving one of three input features: historical voltage stability estimates, historical current stability estimates, and historical temperature stability estimates. The output layer comprises one node, whose output corresponds to the recommended detection period, typically expressed in seconds or milliseconds. The number of nodes in the hidden layers needs to be determined based on the problem complexity and the amount of training data. Too many hidden layer nodes may lead to overfitting, while too few may lead to underfitting; the optimal number needs to be determined experimentally. It is understood that the three-layer feedforward neural network structure can learn the nonlinear mapping relationship between input features and output periods, and the nonlinear activation function of the hidden layers is key to achieving this capability.

[0074] In practice, the training process employs backpropagation for supervised learning. During training, samples from the training set are input into the network individually or in batches. The error between the network output value and the actual historical detection period data is calculated. Then, the weights of each connection node in the network are adjusted layer by layer based on the error. The error calculation typically uses the mean squared error function, and the weight adjustment is based on the gradient descent principle. The direction and magnitude of the adjustment are determined by calculating the partial derivative of the error with respect to each weight, and the learning rate controls the adjustment step size. The training process is iterated multiple times until the prediction error falls below a set threshold or the training epochs reach their limit, indicating that the network has learned the mapping relationship from stability estimation to detection period. The trained network parameters and structure are saved. The network parameters include the weight matrix from the input layer to the hidden layer, the weight vector from the hidden layer to the output layer, the bias vector of the hidden layer nodes, and the bias value of the output layer nodes. The network structure includes the number of nodes in each layer and the type of activation function, which together form the detection period recommendation network.

[0075] In practical implementation, when inputting the real-time calculated voltage stability estimate, current stability estimate, and temperature stability estimate into the pre-trained detection period recommendation network, the three estimates need to be combined into a three-dimensional input vector. This input vector is fed into the input layer of the detection period recommendation network. The three nodes in the input layer receive the voltage stability estimate, current stability estimate, and temperature stability estimate, respectively, and directly pass the values ​​to the hidden layer without transformation. Each node in the hidden layer calculates the weighted sum of the input vector and its corresponding weight vector, adds its bias value, and then performs a nonlinear transformation through an activation function. The activation function can be either a sigmoid function or a rectified linear unit function. The output calculation of the hidden layer nodes can be expressed as:

[0076]

[0077] in: It is the first The output of each hidden layer node It is an activation function. It is the input layer. The node is connected to the hidden layer. The connection weight of each node It is the first of the input vectors Each component (i.e., voltage stability estimate, current stability estimate, or temperature stability estimate). It is the hidden layer. The bias values ​​of each node are calculated. The output layer node receives the outputs of all nodes in the hidden layer, calculates the weighted sum, adds the bias values ​​of the output layer nodes, and then generates the final output value through the output activation function. The output activation function can be selected as a linear function or a sigmoid function depending on the range of the detection period.

[0078] In some embodiments, the calculation of the output layer node can be represented as:

[0079]

[0080] in: It is the output value of the output layer node. It is the output activation function. It is the hidden layer. The connection weights from each node to the output layer node It is the number of hidden layer nodes. It is the bias value of the output layer node. It is the first The output of each hidden layer node. Output value It needs to be mapped to the actual detection cycle range. The mapping function is designed according to the actual application requirements. For example, it can be... The detection period is linearly scaled between the minimum and maximum detection periods. This means the detection period recommendation network can adaptively recommend a detection frequency based on the motherboard's current stability; a longer detection period is recommended when stability is high to reduce resource consumption, while a shorter detection period is recommended when stability is low to enhance monitoring.

[0081] In some embodiments, the training of the detection period recommendation network can employ various techniques to improve performance, such as using momentum to accelerate gradient descent convergence, using regularization to prevent overfitting, and using early stopping to avoid overtraining. Network evaluation after training should be performed on an independent test set that does not overlap with the training set. Evaluation metrics include mean absolute error or root mean square error to ensure the network has good generalization ability. Optionally, the detection period recommendation network can also be implemented using other machine learning models, but a three-layer feedforward neural network has the advantages of being easy to train and capable of approximating any continuous function.

[0082] Example 4: In specific implementation, the real-time monitoring process is deployed based on the customized detection frequency recommended by the network output during the detection cycle. The customized detection frequency is a value in units of time, representing the ideal interval for data acquisition. For example, the customized detection frequency may be 100 milliseconds or 1 second. Starting the real-time monitoring process requires initializing a timer component. The timer component can be a software-implemented counting timer or a hardware-supported high-precision timer. The configuration parameters of the timer component are directly derived from the customized detection frequency, and the trigger period of the timer component is set equal to the customized detection frequency value. After the timer component starts running, it periodically generates data acquisition events according to the customized detection frequency. A data acquisition event is a system interrupt or software signal used to notify the data acquisition module to perform a reading operation. The data acquisition event handler needs to ensure timely response upon triggering to avoid missing acquisition opportunities. In some embodiments, the implementation of the timer component can be based on the time service provided by the real-time operating system. The precision adjustment of the timer component is achieved by setting the interval time in the itimerspec structure, ensuring strict synchronization with the customized detection frequency.

[0083] In practical implementation, when the timer component triggers a data acquisition event, the data acquisition module is activated. The data acquisition module needs to synchronously read data from various sensors connected to the integrated circuit motherboard, including voltage sensor readings, current sensor readings, temperature sensor readings, and jitter indicators output by the signal analyzer. Synchronous reading requires all data points to be captured at the same time point as much as possible to reduce time deviations introduced by sequential readings. Synchronous reading can be achieved by accessing the sensor interfaces in parallel or by using a synchronization signal to trigger multiple sensors to sample simultaneously. Voltage sensor readings are acquired through an ADC chip, which typically connects to test points on the motherboard. The ADC chip converts analog voltage values ​​into digital values. Current sensor readings are obtained through a Hall effect sensor or a sampling resistor circuit; the current value is amplified and digitized. Temperature sensor readings are read through a thermistor or a digital temperature sensor such as the DS18B20. The jitter indicators output by the signal analyzer are measured by dedicated analyzer hardware or a timer circuit integrated on the motherboard, reflecting the stability of the signal timing. All readings are read immediately after the data acquisition event is triggered and temporarily stored in a buffer.

[0084] Optionally, the synchronous reading process can be optimized using multi-threading techniques, such as creating an independent reading thread for each sensor type. All threads start synchronously when a data acquisition event is triggered, and a barrier synchronization mechanism ensures consistent reading completion times. If a sensor fails or times out during the reading process, the data acquisition module can record an error flag but continue reading data from other available sensors to maintain partial data integrity. The raw values ​​of the voltage sensor readings, current sensor readings, temperature sensor readings, and jitter index need to undergo preliminary verification, such as range checks, to exclude obviously invalid values, before being prepared for time stamping. Each successfully acquired reading must be accurately time-stamped using a high-precision clock source, such as a Unix timestamp synchronized with the system clock via NTP, or a nanosecond-level timestamp provided by a hardware counter. The additional time-stamping operation is performed immediately after data reading to ensure accurate correspondence between the time stamp and the reading. After time-stamping, the readings are assembled into a real-time data record, which is a structured data object containing a timestamp field, a voltage reading field, a current reading field, a temperature reading field, and a jitter index field. Real-time data records can be stored in memory or persistently. Memory storage uses a circular buffer to manage and save the latest record in a circular overwrite manner. Persistent storage writes records to a database or file system, such as using an SQLite database table, with each record inserted as a row. The structure of real-time data records is shown in Table 1, which illustrates the field definitions and data types of the records.

[0085] Table 1: Real-time Data Recording Structure Table

[0086] Field Name Data Type Description Time Stamp Timestamp The precise time of data collection, formatted as YYYY-MM-DD HH:MM:SS.ssssss, for example 2023-10-01 12:00:00.123456 Voltage Reading Floating Point Number The voltage value read by the voltage sensor, in volts, with a range determined by the mainboard design, for example 0-5V Current Reading Floating Point Number The current value read by the current sensor, in amperes, with a range determined by the mainboard load, for example 0-10A Temperature Reading Floating Point Number The temperature value read by the temperature sensor, in degrees Celsius, with a range determined by the heat dissipation conditions, for example -40-125°C Jitter Index Floating Point Number The jitter value output by the signal analyzer, in nanoseconds, representing the time deviation of the signal edge

[0087] Optionally, real-time data recordings can be stored using compression algorithms to reduce storage space, such as differential encoding or quantization techniques for floating-point segments. However, the compression process must ensure reversible decompression to avoid information loss. Indexing mechanisms, such as time-stamped B-tree indexes, can also be added during storage to accelerate subsequent data retrieval and collection assembly operations. Real-time data records accumulate sequentially over time, forming a serialized data stream.

[0088] In practice, the accumulation of real-time data records across multiple time points is achieved through a data management module. This module monitors the number of records in the storage area and triggers the collection assembly process when the number of records reaches a preset threshold or after a fixed time window. The collection assembly process retrieves a series of real-time data records from the storage medium at consecutive time points, arranges them in ascending order by timestamp, and assembles them into a real-time data set. The real-time data set can be represented as a time-series array or a data frame structure, with each time point corresponding to one record. The overall size of the real-time data set depends on the customized detection frequency and the accumulated time length. For example, if the customized detection frequency is 1 second and the accumulated time is 1 minute, the real-time data set contains 60 records. The real-time data set is maintained in memory as a dynamic array or linked list, supporting fast iterative access, and can also be exported to file formats such as CSV or JSON for external analysis. During the assembly process, the data management module verifies the temporal continuity of the records. If a timestamp interval does not match the customized detection frequency, an empty record is inserted or a missing point is marked, but the originality of the data is prioritized.

[0089] Example 5: In specific implementation, the pattern recognition engine performs multi-dimensional analysis on the real-time data set to classify the performance status categories of the integrated circuit motherboard. The analysis process first extracts voltage reading sequences, current reading sequences, temperature reading sequences, and signal quality index sequences from the real-time data set. The real-time data set consists of multiple real-time data records arranged in chronological order. Each real-time data record contains a timestamp, voltage reading, current reading, temperature reading, and jitter index fields. The sequence extraction operation sorts and groups the field values ​​based on the timestamp. The voltage reading sequence is a list of voltage readings from all real-time data records arranged in chronological order; the current reading sequence is a list of current readings arranged in chronological order; the temperature reading sequence is a list of temperature readings arranged in chronological order; and the signal quality index sequence is a list of jitter index values ​​arranged in chronological order. After sequence extraction, completeness needs to be verified to check for missing time points or outliers. However, since the continuity of the real-time data set is usually processed during the assembly stage, it directly enters the feature calculation stage. In some embodiments, sequence extraction can be achieved using database query operations, such as executing a SELECT statement on an SQLite database storing a real-time dataset to retrieve specific fields sorted by timestamp, or using list operations in a programming language to slice sequence values ​​from an in-memory data structure.

[0090] In practical implementation, calculating the statistical characteristics of each sequence is the core step of feature engineering. For voltage reading sequences, the statistical characteristics include mean, variance, skewness, and kurtosis. The mean is calculated as the arithmetic mean of all voltage readings in the sequence, i.e., the sum of all voltage values ​​divided by the number of voltage values. The variance is calculated as the average of the squares of the differences between each voltage reading and the mean, reflecting the dispersion of the voltage values. Skewness is calculated as the third standard moment of the voltage reading sequence, measuring the asymmetry of the voltage value distribution. Kurtosis is calculated as the fourth standard moment of the voltage reading sequence, measuring the sharpness of the voltage value distribution compared to a normal distribution. For current reading sequences, the mean, variance, skewness, and kurtosis are calculated similarly, using the same methods as for voltage reading sequences, but applied to current data. For temperature reading sequences, the mean, variance, skewness, and kurtosis are calculated. For a signal quality index sequence, statistical characteristics include the mean and coefficient of variation of jitter values. The mean jitter value is calculated as the arithmetic mean of all jitter index values, and the coefficient of variation is calculated as the standard deviation of the jitter index sequence divided by the mean jitter value, reflecting the relative volatility of jitter. Optionally, the mean can be calculated using an iterative algorithm to avoid large-number accumulation errors, the variance can be calculated using a two-pass algorithm or an online algorithm, and the skewness and kurtosis are calculated based on the sample moment estimation formula, but unbiased correction should be noted.

[0091] In some embodiments, the specific implementation of statistical feature calculation can be achieved by writing a feature extraction function, which takes a numerical sequence as input and outputs feature values. Skewness calculation requires first calculating the mean and mean, then using a formula based on third-moment normalization; kurtosis calculation requires normalization based on fourth-moment normalization; and the coefficient of variation calculation for the signal quality index sequence requires first calculating the standard deviation and mean. All calculations ensure numerical stability, avoiding division-by-zero errors or overflow. Optionally, for short sequences, statistical features may not be robust; therefore, the real-time dataset needs to contain a sufficient number of real-time data records, such as at least 30 points, to ensure feature representativeness.

[0092] In practice, forming a feature vector is the process of combining all calculated statistical features into a high-dimensional vector. The feature vector serves as input to the classification model, and its dimension depends on the number of sequences and the number of features calculated for each sequence. For example, voltage reading sequences contribute four features (mean, variance, skewness, and kurtosis), current reading sequences contribute four features, temperature reading sequences contribute four features, and signal quality index sequences contribute two features (mean and coefficient of variation), resulting in a total feature dimension of 14. Feature vectors are assembled in a fixed order, such as voltage features first, then current features, then temperature features, and finally signal features, ensuring consistent feature vector structure each time. Feature vectors require normalization to ensure that all feature values ​​are at a similar scale. For example, min-max normalization can be used to scale each feature to the range of zero to one, preventing certain features from dominating the classification result due to their large values. Normalization parameters are obtained statistically from the training data and reused in applications. Optionally, feature vectors can be dimensionality-reduced, such as through principal component analysis to reduce redundancy, but this increases complexity; usually, the original feature vectors are used directly.

[0093] In practice, the pre-trained classification model uses decision tree rules for state partitioning. The training process employs the random forest algorithm, an ensemble learning method that improves classification accuracy by constructing multiple decision trees and combining voting results. Training the classification model requires collecting historical real-time data sets and their corresponding known performance status labels. The historical real-time data set is a collection of previously collected real-time data records, and the known performance status labels are assigned by experts based on motherboard behavior, such as "normal," "warning," and "abnormal," together forming the training dataset. The training process randomly selects multiple subsets from the training dataset. Each subset is obtained through sampling with replacement, called a bootstrap set, and a decision tree is constructed for each bootstrap set. When constructing the decision tree, each node split only considers a randomly selected feature subset. The size of the feature subset is typically the square root or logarithm of the total number of features, increasing the diversity between trees. Decision tree growth involves recursively splitting nodes, selecting the optimal splitting features and thresholds to maximize information gain or reduce Gini impurity. Tree growth stops when node purity is too high or the number of samples is too low. After training, the structural parameters and splitting rules of all decision trees, as well as the voting mechanism, are saved to form the classification model. Optionally, hyperparameters of random forests, such as the number of trees and maximum depth, can be adjusted through cross-validation to avoid overfitting.

[0094] In practical implementation, when the feature vector is input into the pre-trained classification model, each decision tree within the model makes predictions independently. Each decision tree starts from the root node and traverses the tree branches to the leaf nodes based on the feature values ​​of the feature vector. The leaf nodes store a class probability or a direct class label. The random forest integrates the prediction results of all decision trees and determines the final classification result through a majority voting mechanism; that is, the class with the most votes is output as the performance evaluation label. For example, if there are 100 decision trees, 60 predict "normal," 30 predict "warning," and 10 predict "abnormal," then the output performance evaluation label is "normal." The voting process can be weighted, assigning weights based on the accuracy of the trees, but usually simple majority voting is sufficient. The application of the classification model needs to be implemented efficiently, especially for scenarios with high real-time requirements. Decision tree predictions can be optimized through pre-compiled rules or lookup tables. It can be understood that the random forest classification model can handle non-linear relationships, is robust to noise, and is suitable for classifying the performance status of integrated circuit motherboards. Optionally, the classification model can be updated periodically, retrained with new data to adapt to motherboard aging or environmental changes.

[0095] In practical implementation, the overall implementation of the pattern recognition engine can be encapsulated as a software module. The module takes a real-time dataset as input and outputs performance evaluation labels. The internal process includes sequence extraction, feature calculation, feature vector assembly, and classification prediction. In the sequence extraction stage, voltage reading sequences, current reading sequences, temperature reading sequences, and signal quality index sequences are parsed from the real-time dataset. In the feature calculation stage, statistical functions are called to calculate the mean, variance, skewness, kurtosis, and signal characteristics of each sequence. In the feature vector assembly stage, feature values ​​are combined sequentially and normalized. In the classification prediction stage, a pre-trained classification model file is loaded for inference. The module is implemented using programming languages ​​such as Python or C++, utilizes scientific computing libraries for numerical calculations, and loads a random forest model using machine learning libraries. The module is deployed on a performance analysis server and integrated with the real-time monitoring process to achieve automated performance evaluation. Optionally, the module can add a logging function to record intermediate results during the classification process for debugging and auditing.

[0096] See Figure 5 This image illustrates four key statistical data points related to jitter in integrated circuit motherboard performance testing. These data are crucial inputs for the pattern recognition engine to analyze real-time datasets, perfectly aligning with the technical workflow. During real-time monitoring, jitter metrics output by the signal analyzer are simultaneously acquired. Subsequently, the average and coefficient of variation of the jitter values ​​are calculated to form the statistical characteristics of the signal quality index sequence. The minimum and maximum values ​​define the fluctuation range of the jitter index, the average value reflects the overall level of jitter, and the coefficient of variation quantifies the relative dispersion of jitter. The data format in this image directly corresponds to the technical requirements for calculating the statistical characteristics of the signal quality index sequence, providing crucial data support for performance status classification and serving as a vital step in achieving multi-dimensional intelligent diagnostics.

[0097] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of performance analysis of an integrated circuit motherboard, characterized by, The method comprises: Deriving voltage sampling data, current sampling data and temperature sampling data in a historical time range from a monitoring system of an integrated circuit mainboard, performing integrity check and outlier removal on the voltage sampling data, current sampling data and temperature sampling data, and generating a clean historical data set; Performing parameter fluctuation evaluation on the clean historical data set, generating voltage stability estimate, current stability estimate and temperature stability estimate by calculating oscillation amplitude of voltage data, drift range of current data and fluctuation degree of temperature data; Inputting the voltage stability estimate, current stability estimate and temperature stability estimate into a pre-trained detection cycle recommendation network, the network outputting a customized detection frequency by comparing a historical mode library; Deploying a real-time monitoring process according to the customized detection frequency, capturing instantaneous voltage readings, instantaneous current readings, instantaneous temperature readings and signal quality indicators of the integrated circuit mainboard, and assembling a real-time data set; Using a pattern recognition engine to perform multi-dimensional analysis on the real-time data set, dividing performance state categories, and outputting performance evaluation labels.

2. The method of claim 1, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The integrity check and outlier removal on the voltage sampling data, current sampling data and temperature sampling data comprises: Checking the time stamp continuity of the voltage sampling data, current sampling data and temperature sampling data, and identifying data points corresponding to missing time stamps; For data points with missing time stamps, linear interpolation is used to fill in adjacent time stamp data points to form a continuous data sequence; Calculate the deviation of each data point in the continuous data sequence from the adjacent data point, if the deviation exceeds the preset tolerance threshold, mark it as an outlier and remove it; Normalizing the data sequence after removing outliers to unify the scales of all data, and obtaining a clean historical data set.

3. The method of claim 2, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The parameter fluctuation evaluation on the clean historical data set comprises: Extracting the voltage data sequence in the clean historical data set, calculating the average value of the absolute value of the difference between adjacent sampling points in the voltage data sequence as the oscillation amplitude; Extracting the current data sequence in the clean historical data set, calculating the difference between the maximum value and the minimum value of the current data sequence divided by the average value as the drift range; Extracting the temperature data sequence in the clean historical data set, calculating the ratio of the standard deviation to the average value of the temperature data sequence as the fluctuation degree; Map the oscillation amplitude, drift range and fluctuation degree to the range of zero to one respectively to obtain the voltage stability estimate, current stability estimate and temperature stability estimate.

4. The method of claim 3, wherein the performance of the integrated circuit main board is analyzed by using a performance analysis program. The pre-trained detection cycle recommendation network is constructed by: Collecting multiple sets of historical voltage stability estimates, historical current stability estimates and historical temperature stability estimates, and corresponding historical detection cycle data to form a training sample set; Constructing a three-layer neural network structure, the input layer contains three nodes corresponding to voltage stability estimate, current stability estimate and temperature stability estimate respectively, and the output layer is one node corresponding to detection cycle; Using the training sample set to supervise the learning of the neural network, adjusting the network weights through the back propagation algorithm until the prediction error is lower than the set threshold; Save the trained network parameters and structure as the detection cycle recommendation network.

5. The method of claim 4, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The inputting of the voltage stability estimate, the current stability estimate, and the temperature stability estimate into the pre-trained detection cycle recommendation network includes: Combining the voltage stability estimate, the current stability estimate, and the temperature stability estimate into an input vector; Feeding the input vector into the input layer of the detection cycle recommendation network, calculating the weighted sum through the hidden layer nodes, and applying the activation function; The output layer node calculates the final value, maps the final value to the actual detection cycle range, and obtains the customized detection frequency.

6. The method of claim 5, wherein the performance analysis of the integrated circuit main board is performed by a performance analysis program. The deployment of the real-time monitoring process includes: Setting a timer according to the customized detection frequency to trigger a data collection event; When the timer triggers, synchronously read the voltage sensor readings, current sensor readings, temperature sensor readings of the integrated circuit motherboard, and the jitter index output by the signal analyzer; Time-stamp each collection of readings and store them as real-time data records; Accumulate the real-time data records at multiple time points to assemble a real-time data set.

7. The method of claim 6, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The multi-dimensional analysis of the real-time data set by the pattern recognition engine includes: Extracting the voltage reading sequence, current reading sequence, temperature reading sequence, and signal quality index sequence from the real-time data set; Calculating the statistical characteristics of each sequence, including mean, variance, skewness, and kurtosis, to form a feature vector; Inputting the feature vector into a pre-trained classification model, which performs state division based on decision tree rules; Outputting the classification result as a performance evaluation label.

8. The method of claim 7, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The pre-trained classification model is trained through the following process: Collecting historical real-time data sets and their corresponding known performance state labels as a training data set; Randomly extracting a subset from the training data set to build multiple decision trees, each using a different subset of features; Integrating the voting results of multiple decision trees to determine the final classification rule; Saving the decision tree parameters and voting mechanism as a classification model.

9. The method of claim 8, wherein the performance analysis of the integrated circuit mainboard is performed by a performance analysis program. The calculation of the statistical characteristics of each sequence includes: For the voltage reading sequence, calculate the arithmetic mean of all voltage values as the mean, calculate the average square of the voltage value deviation from the mean as the variance, and calculate the asymmetry of the voltage value distribution as the skewness, and calculate the sharpness of the voltage value distribution as the kurtosis; Perform the same calculation on the current reading sequence and the temperature reading sequence to obtain their respective mean, variance, skewness, and kurtosis; For the signal quality index sequence, calculate the average value and coefficient of variation of the jitter value.

10. A performance detection analysis system for an integrated circuit mainboard, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the integrated circuit motherboard performance detection analysis method according to any one of claims 1 to 9.

Images