Server PCB automatic detection system based on operation data collection and analysis

Abstract

The present invention belongs to the field of server detection technology and discloses a server PCB automatic detection system based on operation data collection and analysis. The system consists of an operation data distributed acquisition module, a preprocessing and synchronous timing correction module, a real-time signal integrity detection module, a thermal field dynamic distribution analysis module, a timing drift intelligent analysis module, a multi-dimensional joint feature extraction module, an abnormal pattern adaptive recognition module, a fault trend deep prediction module, and a self-learning feedback and online optimization module. Acquisition units equipped with FPGA hardware are deployed at key nodes on the server PCB, utilizing its high-speed parallel acquisition characteristics to achieve distributed acquisition of multi-source electrical data such as current, voltage, and temperature. After the collected data is preprocessed and subjected to asynchrony and noise interference by the synchronous timing correction module, the various functional modules work together to build a full-process online dynamic detection system from data acquisition to abnormality analysis, which can quickly detect abnormalities at the incipient stage of failures.

Description

Technical Field

[0001] The present invention belongs to the technical field of server detection, and in particular relates to a server PCB automatic detection system based on operation data collection and analysis. Background Art

[0002] Amid the rapid development of server technology, printed circuit boards (PCBs), as key components of servers, have a significant impact on overall server performance and system reliability. PCBs house complex power management, high-speed signal transmission, and thermal management networks. Any minor electrical anomalies or structural flaws can lead to server performance fluctuations, data transmission errors, and even system failures. However, existing server PCB inspection methods still primarily rely on spot checks during production or static electrical testing during maintenance. These tests have long inspection cycles and poor real-time performance, making it difficult to promptly respond to dynamic faults during server operation. This is especially true in environments where servers operate under continuous high loads and for extended periods. Existing inspection methods are unable to capture transient anomalies or early-stage faults that may arise in real-time, resulting in significant blind spots. Furthermore, limited by labor costs, inspection tools, and testing time, traditional inspection methods have limited coverage and are unable to fully scan the operating status of every PCB node. This makes it difficult to promptly expose potential faults, increasing the risk of server system crashes or downtime.

[0003] With the increasing complexity of server structures and the continuous expansion of data center scale, PCB operation data has shown strong multi-source and complex correlation, involving multi-dimensional parameters such as current, voltage, temperature, signal integrity, and timing offset. However, existing technologies usually only focus on a single electrical indicator, such as current monitoring or local temperature detection, and lack real-time synchronous collection and coupled analysis of multi-source data, making it impossible to dynamically grasp the health status of the PCB throughout its life cycle. In addition, existing data processing mostly uses static threshold judgment or a single algorithm, lacking intelligent anomaly recognition and trend prediction capabilities. It cannot effectively identify minor defects, progressive aging, or complex fault paths, and the detection results are prone to misjudgment or missed detection. Summary of the Invention

[0004] The purpose of the present invention is to provide a server PCB automatic detection system based on operation data collection and analysis to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a server PCB automatic detection system based on operation data collection and analysis, the system comprising an operation data distributed collection module, a preprocessing and synchronization timing correction module, a signal integrity real-time detection module, a thermal field dynamic distribution analysis module, a timing drift intelligent analysis module, a multi-dimensional joint feature extraction module, an abnormal pattern adaptive recognition module, a fault trend deep prediction module, and a self-learning feedback and online optimization module;

[0006] Operational data distributed acquisition module: As the starting point of the system, acquisition units are deployed at key nodes on the PCB. FPGA hardware is used to collect multi-source electrical data such as current and voltage in parallel, and the raw data is synchronously transmitted via a high-speed bus to provide a data foundation for subsequent processing.

[0007] Preprocessing and synchronization timing correction module: This module receives raw data and, given the asynchrony of multi-source data, applies time alignment and multi-clock domain compensation algorithms based on timestamps to preprocess and synchronize the data. It also performs filtering and denoising to improve data time consistency and accuracy for analysis by subsequent modules.

[0008] Signal integrity real-time detection module: Receives corrected data, focuses on high-frequency key signals, analyzes signal waveform parameters through wavelet denoising and reflection attenuation detection algorithms, identifies abnormal features such as distortion, and outputs signal integrity labels, providing key clues for fault analysis.

[0009] Thermal field dynamic distribution analysis module: Based on the synchronously corrected temperature data and signal integrity labels, a dynamic thermal field fitting algorithm is used to construct a thermal field distribution map, associate high-temperature points with signal anomaly points, implement coupled modeling, and output multi-source coupling anomaly signs to indicate potential fault risks;

[0010] Timing Drift Intelligent Analysis Module: Based on the corrected data stream, combined with thermal anomalies and signal anomalies, it uses sliding window PLL phase locking technology to monitor the phase information of high-speed data channels, detect timing offset and jitter, and output risk labels to assist in fault diagnosis.

[0011] Multi-dimensional joint feature extraction module: This module integrates the output data of the first five modules and uses feature embedding and principal component analysis algorithms to map the data to the same space and reduce the dimensionality. It then extracts comprehensive feature vectors and outputs them to the anomaly recognition module, laying the foundation for anomaly judgment.

[0012] Abnormal pattern adaptive recognition module: Receives feature vectors and uses the LSTM recursive neural network and density clustering algorithm to learn the time series characteristics of the data and classify them according to distribution density. It identifies abnormal types such as sudden and gradual abnormalities and feeds the results back to the fault prediction module.

[0013] Fault trend deep prediction module: Based on abnormal pattern results and historical data, it builds a dynamic Bayesian network and time series prediction model to analyze fault probability and development trends, output risk level-driven alarms, and facilitate timely troubleshooting;

[0014] Self-learning feedback and online optimization module: Dynamically adjusts anomaly thresholds, model weights, and feature extraction priorities based on alarm, operation and maintenance, and fault verification results, supports online learning optimization, and continuously improves system adaptability and detection accuracy.

[0015] Preferably, the distributed operation data collection module includes:

[0016] (1) Scientifically arrange the acquisition units at multiple nodes: Run the distributed data acquisition module at key nodes such as the power supply line, high-frequency signal transmission line, and high-heat chip periphery of the server PCB, and scientifically arrange the acquisition units. Different from the existing centralized single-point acquisition mode, this distributed layout can cover more PCB areas, realize the comprehensive acquisition of multi-source electrical data, and provide a rich data foundation for system detection;

[0017] (2) Using FPGA to achieve high-speed parallel acquisition: Each acquisition unit of this module is equipped with high-performance FPGA hardware, which uses its high-speed parallel acquisition characteristics to simultaneously acquire current, voltage, temperature, signal waveform and data transmission timing. Compared with traditional acquisition methods, FPGA can ensure the integrity of signal timing capture, and cooperate with high-precision Hall current sensors, isolation op amp circuits and digital temperature sensors to achieve accurate data acquisition. The multi-node data synchronization sampling formula ensures the consistency of data sampling at each node. The original data obtained during the acquisition process is transmitted to the next module in a synchronous manner through a high-speed data transmission bus. This transmission method ensures the timeliness and accuracy of the data, lays a reliable foundation for subsequent data preprocessing, feature extraction and other links, and enables the data processing process of the entire detection system to proceed smoothly;

[0018] The expression formula for multi-node data synchronous sampling is:

[0019]

[0020] Where: For the A collection node at time The data values ​​collected; Run signals for real PCB; is the sampling error (random interference term).

[0021] Preferably, the preprocessing and synchronization timing correction module includes:

[0022] (1) Accepting data and clarifying processing objectives: The preprocessing and synchronous timing correction module accepts the multi-source electrical data output by the distributed data acquisition module. Due to the asynchronous nature of the data during the acquisition process, this module needs to preprocess and perform synchronous timing correction on the raw data. This is a key link that conventional systems lack, laying a solid foundation for subsequent accurate analysis.

[0023] (2) Realize time synchronization and multi-clock domain correction: The module is based on timestamps and uses a time alignment algorithm. Through a specific time alignment correction formula, it accurately compares and adjusts the timestamps of different data sources to make the data uniform on the time axis. At the same time, it uses multi-clock domain compensation technology to finely adjust the clock signal for acquisition delay errors, eliminate time deviations, and ensure data time consistency. In the data preprocessing stage, the module filters the raw data to effectively remove noise interference. After processing by this module, the time accuracy and data quality of the data are greatly improved, providing high-precision basic data for subsequent modules such as signal integrity detection and feature extraction, ensuring the reliability of system analysis;

[0024] Time alignment correction expression formula:

[0025]

[0026] Where: For the Node and Time synchronization deviation of nodes; Timestamp for each node is the reference time delay correction value.

[0027] Preferably, the signal integrity real-time detection module includes:

[0028] (1) Joint algorithm deeply processes the corrected data: The real-time signal integrity detection module receives the accurate data after pre-processing and synchronization timing correction, and adopts the wavelet denoising + reflection detection joint algorithm, which is different from the existing simple discrimination method based on Fourier analysis. The wavelet denoising algorithm can effectively decompose the signal frequency band and accurately remove noise based on a specific formula; the reflection detection algorithm uses relevant formulas to monitor the signal transmission reflection in real time. The two work together to deeply restore the true characteristics of the signal and lay a solid data foundation for abnormal analysis;

[0029] Wavelet denoising expression formula:

[0030]

[0031] Where: is the denoised signal; is the wavelet decomposition coefficient For the Order wavelet basis function; is the decomposition order;

[0032] (2) Accurately identify anomalies and output key tags: The module accurately calculates and compares parameters such as the amplitude, phase, and rise time of high-frequency signals, and uses a joint algorithm to capture abnormal characteristics such as signal distortion, crosstalk, and reflection in real time. When an anomaly is detected, the signal integrity tag is immediately output, providing key clues for subsequent fault analysis, achieving efficient and accurate detection of PCB high-frequency signal integrity and ensuring stable server operation.

[0033] Preferably, the thermal field dynamic distribution analysis module includes:

[0034] (1) Constructing thermal field distribution maps based on multi-source data: The thermal field dynamic distribution analysis module uses the synchronously corrected temperature data and signal integrity labels as the basis, applies the thermal field anomaly fitting formula, and uses the dynamic thermal field fitting algorithm to study the thermal field distribution of the server PCB. By correlating temperature data with signal integrity data, an accurate PCB thermal field distribution map is constructed, providing intuitive data support for subsequent analysis;

[0035] Thermal field anomaly fitting expression formula:

[0036]

[0037] Where: is the degree of thermal abnormality; is the thermal weight coefficient; For the Temperature of each measuring point; is the overall average temperature of the PCB; is the number of temperature measurement points;

[0038] (2) Coupling modeling to identify potential failure risks: This module breaks the limitations of single thermal analysis and integrates thermal field and signal anomalies for coupled modeling. By analyzing the thermal field distribution map, high-temperature points are accurately located and associated with signal anomaly points to achieve signal-thermal anomaly coupled modeling. Once a correlation is found between the two, a multi-source coupled anomaly flag is output, effectively indicating potential failure risks. Compared with traditional detection methods, it is more comprehensive and accurate.

[0039] Preferably, the timing drift intelligent analysis module includes:

[0040] (1) Multi-source data fusion and detection technology: The timing drift intelligent analysis module is based on the corrected data stream, integrating the thermal anomaly points of the thermal field dynamic distribution module and the signal anomaly points of the signal integrity module, focusing on the timing drift detection of high-speed data channels. It uses enhanced sliding window PLL phase locking technology and, based on the sliding window timing drift detection formula, monitors the data stream phase information in real time and accurately analyzes potential timing offsets and phase jitter.

[0041] Sliding window timing drift detection expression formula:

[0042]

[0043] Where: Timing drift is the number of sampling points in the sliding window; is the ideal signal sampling time; is the actual signal sampling time;

[0044] (2) Accurately identify anomalies and output key labels: Different from traditional timing analysis methods, this module makes timing analysis more advanced by setting an adaptive sliding window size and phase lock threshold. Once an abnormal timing change is detected, the timing drift risk label is immediately output, providing a core basis for subsequent fault diagnosis, effectively improving the early warning capability of high-speed data transmission anomalies on server PCBs, and ensuring data transmission stability and reliability.

[0045] Preferably, the multi-dimensional joint feature extraction module includes:

[0046] (1) Multi-source data fusion and innovative feature embedding: The multi-dimensional joint feature extraction module takes over the multi-type data such as electrical parameters and signal integrity labels output by the first five modules. Different from the traditional single-variable extraction method, this module proposes a feature embedding path for coupled data. Through feature embedding technology, data of different dimensions and properties are mapped to the same feature space, breaking the data type restrictions, enhancing data compatibility and analyzability, and building a unified foundation for deep feature extraction;

[0047] (2) Principal Component Analysis: After completing data fusion and feature embedding, the principal component analysis (PCA) algorithm is used to deeply process the data. Based on the principal component feature extraction formula, the fused data is subjected to dimensionality reduction operations to screen out the most representative key features, eliminate data redundancy, and extract a feature vector that can fully reflect the operating status of the server PCB. This feature vector is then output to the anomaly identification module, providing core data support for accurately identifying PCB operation anomalies.

[0048] The principal component feature extraction expression formula is:

[0049]

[0050] Where: is the principal component eigenvector; For the The contribution rate of each feature; For the Normalized input features; is the total number of input features.

[0051] Preferably, the abnormal pattern adaptive recognition module includes:

[0052] (1) Dual algorithms collaboratively process feature vectors: The abnormal pattern adaptive recognition module receives the joint feature vector output by the multi-dimensional joint feature extraction module and introduces a dual algorithm architecture of LSTM recursive neural network and density clustering. LSTM, through its unique time recursive formula, can deeply mine the time series characteristics of data and conduct a comprehensive and integrated analysis of historical and current data; the density clustering algorithm effectively distinguishes different types of data based on the data distribution density. Compared with a single algorithm, the combination of the two significantly improves the ability to identify complex anomalies;

[0053] LSTM time recursive expression formula:

[0054]

[0055] is the hidden state at the current moment; is the hidden state of the previous moment; For current input data; is the network weight; is the bias term; Activation function;

[0056] (2) Accurately classify anomalies and output results: Based on the collaborative operation of LSTM and density clustering algorithms, this module can accurately identify complex anomaly types such as sudden anomalies, gradual aging, and periodic anomalies of server PCBs. Through in-depth analysis of feature vectors, detailed anomaly classification results are generated and promptly fed back to the fault prediction module, providing key information for subsequent prediction of potential fault development paths, ensuring the accuracy and reliability of server PCB anomaly diagnosis.

[0057] Preferably, the fault trend depth prediction module includes:

[0058] (1) Dual model construction to achieve intelligent prediction: The fault trend deep prediction module is based on abnormal pattern results and historical data streams, and introduces dynamic Bayesian networks and time series prediction models. The dynamic Bayesian network models and analyzes the probability of fault occurrence and causal relationship based on the dynamic Bayesian prediction formula, and can simulate the multi-path development trend of faults under different conditions, which is different from the single judgment method of simple threshold prediction; the time series prediction model is based on the trend of historical data changes and accurately deduces the future operating status. The dual models provide data support for fault prediction from different dimensions;

[0059] (2) Collaborative prediction aids troubleshooting: Through the collaborative work of dynamic Bayesian networks and time series prediction models, this module can deeply analyze the potential development path of server PCB failures and output accurate failure risk levels in real time. The prediction results directly drive the automatic alarm module, prompting the system to respond promptly, providing a reliable basis for operation and maintenance personnel to take quick measures to troubleshoot, effectively reducing the losses caused by server PCB failures and improving the overall stability and reliability of the system.

[0060] Dynamic Bayesian prediction expression formula:

[0061]

[0062] Where: is the probability of failure occurring at the next moment; is the implicit fault path in the current state is the total number of hidden states.

[0063] Preferably, the self-learning feedback and online optimization module includes:

[0064] (1) Constructing a data feedback closed loop to achieve real-time adjustment: The self-learning feedback and online optimization module constructs a complete data feedback closed loop from automatic alarm, actual operation and maintenance to fault verification, which is different from the existing offline training mode. The module dynamically adjusts the entire detection system in real time based on the results of the automatic alarm module, the actual operation and maintenance situation, and the subsequent fault verification results. By dynamically adjusting the abnormal threshold, the system can flexibly adapt to different operating environments and fault types, providing a basic guarantee for the system's accurate detection;

[0065] (2) Innovative online incremental learning to improve system performance: This module clarifies the online incremental learning path and uses the incremental learning weight update formula based on feedback data to optimize the model weights, effectively improving the accuracy and reliability of the model. At the same time, the feature extraction priority is adjusted according to actual needs, focusing on key features to improve detection efficiency. Through continuous online incremental learning and dynamic optimization, the adaptability of the system is continuously enhanced to ensure that it always maintains efficient and accurate detection capabilities in long-term operation;

[0066] Incremental learning weight update expression formula:

[0067]

[0068] is the updated model weight; is the existing model weight; is the learning rate; is the loss function; is the weight gradient.

[0069] The beneficial effects of the present invention are as follows:

[0070] 1. The present invention runs a distributed data acquisition module and deploys acquisition units equipped with FPGA hardware at key nodes of the server PCB. By utilizing its high-speed parallel acquisition characteristics, it realizes real-time distributed acquisition of multi-source electrical data such as current, voltage, and temperature. After the collected data is pre-processed and synchronized with the timing correction module to eliminate asynchrony and noise interference, the various functional modules work together to build a full-process online dynamic detection system from data acquisition to anomaly analysis. This system can quickly capture anomalies at the incipient stage of faults, greatly shorten the fault discovery time, effectively improve the reliability of server operation, and reduce the risk of downtime and maintenance costs caused by sudden faults.

[0071] 2. The present invention uses real-time comprehensive monitoring based on multi-source operating data, a multi-dimensional joint feature extraction module, and feature embedding and principal component analysis algorithms to fuse and reduce multi-dimensional data such as current, voltage, and signal integrity, and extract the most representative feature vectors. The abnormal pattern adaptive recognition module uses LSTM recursive neural network and density clustering algorithm to perform in-depth analysis of feature vectors and accurately distinguish various abnormal types such as sudden abnormalities and progressive aging; the fault trend deep prediction module uses dynamic Bayesian network and time series prediction model to further analyze the fault development path; each module is linked together to form a complete intelligent analysis chain from data acquisition, feature extraction to fault prediction, which can accurately locate the fault point, effectively avoid systemic failures caused by minor defects or progressive aging, and significantly improve the accuracy and comprehensiveness of fault diagnosis.

[0072] 3. The present invention builds a complete data feedback closed loop through self-learning feedback and online optimization modules. Based on the automatic alarm results, actual operation and maintenance conditions and fault verification information, it dynamically adjusts the abnormality threshold, optimizes the model weight, and flexibly adjusts the feature extraction priority in real time. At the same time, through online incremental learning and dynamic optimization, the system can continuously optimize the detection strategy and algorithm parameters according to the operating environment, hardware configuration and fault type of different servers. This adaptive capability enables the system to not only cope with complex and changeable operating conditions, but also seamlessly adapt to the PCB detection needs of servers of different models and configurations, greatly improving the versatility and scalability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flow chart of the server PCB automatic detection system based on operation data collection and analysis of the present invention. DETAILED DESCRIPTION

[0074] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0075] like Figure 1 As shown, an embodiment of the present invention provides a server PCB automatic detection system based on operation data collection and analysis, which consists of an operation data distributed collection module, a preprocessing and synchronization timing correction module, a signal integrity real-time detection module, a thermal field dynamic distribution analysis module, a timing drift intelligent analysis module, a multi-dimensional joint feature extraction module, an abnormal pattern adaptive recognition module, a fault trend deep prediction module, and a self-learning feedback and online optimization module;

[0076] Operational data distributed acquisition module: As the starting point of the system, acquisition units are deployed at key nodes on the PCB. FPGA hardware is used to collect multi-source electrical data such as current and voltage in parallel, and the raw data is synchronously transmitted via a high-speed bus to provide a data foundation for subsequent processing.

[0077] Preprocessing and synchronization timing correction module: This module receives raw data and, given the asynchrony of multi-source data, applies time alignment and multi-clock domain compensation algorithms based on timestamps to preprocess and synchronize the data. It also performs filtering and denoising to improve data time consistency and accuracy for analysis by subsequent modules.

[0078] Signal integrity real-time detection module: Receives corrected data, focuses on high-frequency key signals, analyzes signal waveform parameters through wavelet denoising and reflection attenuation detection algorithms, identifies abnormal features such as distortion, and outputs signal integrity labels, providing key clues for fault analysis.

[0079] Thermal field dynamic distribution analysis module: Based on the synchronously corrected temperature data and signal integrity labels, a dynamic thermal field fitting algorithm is used to construct a thermal field distribution map, associate high-temperature points with signal anomaly points, implement coupled modeling, and output multi-source coupling anomaly signs to indicate potential fault risks;

[0080] Timing Drift Intelligent Analysis Module: Based on the corrected data stream, combined with thermal anomalies and signal anomalies, it uses sliding window PLL phase locking technology to monitor the phase information of high-speed data channels, detect timing offset and jitter, and output risk labels to assist in fault diagnosis.

[0081] Multi-dimensional joint feature extraction module: This module integrates the output data of the first five modules and uses feature embedding and principal component analysis algorithms to map the data to the same space and reduce the dimensionality. It then extracts comprehensive feature vectors and outputs them to the anomaly recognition module, laying the foundation for anomaly judgment.

[0082] Abnormal pattern adaptive recognition module: Receives feature vectors and uses the LSTM recursive neural network and density clustering algorithm to learn the time series characteristics of the data and classify them according to distribution density. It identifies abnormal types such as sudden and gradual abnormalities and feeds the results back to the fault prediction module.

[0083] Fault trend deep prediction module: Based on abnormal pattern results and historical data, it builds a dynamic Bayesian network and time series prediction model to analyze fault probability and development trends, output risk level-driven alarms, and facilitate timely troubleshooting;

[0084] Self-learning feedback and online optimization module: Dynamically adjusts anomaly thresholds, model weights, and feature extraction priorities based on alarm, operation and maintenance, and fault verification results, supports online learning optimization, and continuously improves system adaptability and detection accuracy.

[0085] In terms of detection timeliness, the distributed acquisition module for operating data realizes real-time acquisition of multi-source electrical data by deploying acquisition units equipped with FPGA hardware at key nodes. After processing by the pre-processing module, each functional module collaborates to build an online dynamic detection system, which can quickly capture anomalies at the incipient stage of faults, reducing downtime risks and maintenance costs. In terms of fault diagnosis accuracy, relying on real-time monitoring of multi-source data, the multi-dimensional joint feature extraction module integrates dimensionality reduction data, and the abnormal pattern adaptive recognition module and the fault trend deep prediction module use a variety of algorithms to form a complete intelligent analysis chain, accurately locate faults, and avoid minor defects causing system failures. In terms of system adaptability, the self-learning feedback and online optimization modules build a data feedback loop, adjust detection parameters in real time, and support online incremental learning, so that the system can adapt to different operating environments and server models, greatly improving versatility and scalability.

[0086] The distributed operational data collection module, through innovative layout and technological application, efficiently collects multi-source electrical data from server PCBs. This module strategically places collection units at key nodes on the server PCB, such as the power supply lines, high-frequency signal transmission lines, and around high-heat chips. This overcomes the limitations of traditional centralized, single-point collection models, achieving multi-region coverage and providing a rich source of data for system testing.

[0087] Each acquisition unit is equipped with high-performance FPGA hardware. With its high-speed parallel acquisition characteristics, it can simultaneously collect current, voltage, temperature, signal waveform and data transmission timing. Combined with high-precision Hall current sensors, isolated op amp circuits and digital temperature sensors, it ensures the accuracy of the collected data.

[0088] The multi-node data synchronous sampling formula is used to ensure the consistency of data sampling at each node. The collected raw data is synchronously transmitted to the next module via a high-speed data transmission bus to ensure data timeliness and accuracy, laying a solid foundation for subsequent data processing and ensuring the smooth operation of the detection system process.

[0089] The preprocessing and synchronization timing correction module receives multi-source electrical data from the distributed operating data acquisition module. Addressing the asynchrony of the data during the acquisition process, the module performs preprocessing and synchronization timing correction on the raw data based on timestamps. The module utilizes a time alignment algorithm and a time alignment correction formula to accurately compare and adjust timestamps from different data sources, aligning the data on a consistent timeline.

[0090] Multi-clock domain compensation technology is also used to fine-tune clock signals to eliminate time deviations caused by acquisition delay errors. During the data preprocessing stage, raw data is filtered and denoised. This module improves the temporal consistency and accuracy of the data, providing high-precision foundational data for subsequent modules such as signal integrity testing and feature extraction, ensuring the reliability of system analysis.

[0091] The real-time signal integrity monitoring module receives precise data after preprocessing and synchronization timing correction. It utilizes a combined wavelet denoising and reflection detection algorithm, overcoming the limitations of existing simple Fourier analysis-based discrimination. The wavelet denoising algorithm, based on a specific formula, effectively removes noise interference by decomposing the signal's frequency bands. The reflection detection algorithm, using a related formula, monitors the signal's reflections during transmission in real time. These two algorithms work together to deeply restore the signal's true characteristics, providing a reliable data foundation for anomaly analysis.

[0092] This module accurately calculates and compares key parameters of high-frequency signals, such as amplitude, phase, and rise time. Using a joint algorithm, it can capture abnormalities such as signal distortion, crosstalk, and reflections in real time. Once an anomaly is detected, it immediately outputs a signal integrity tag, providing critical information for subsequent fault analysis, enabling efficient and accurate detection of PCB high-frequency signal integrity.

[0093] The thermal field dynamic distribution analysis module, based on synchronized and corrected temperature data and signal integrity labels, utilizes thermal field anomaly fitting formulas and dynamic thermal field fitting algorithms to correlate and analyze the two types of data to construct a PCB thermal field distribution map, providing intuitive data support for subsequent analysis. This module transcends the limitations of single thermal analysis by integrating thermal field and signal anomalies for coupled modeling. It accurately locates high-temperature points in the thermal field distribution map and correlates them with signal anomaly points, achieving signal-thermal anomaly coupled modeling. When a high-temperature point is found to be associated with a signal anomaly point, a multi-source coupled anomaly flag is output, indicating potential fault risks. Compared to traditional methods, this detection is more comprehensive and accurate, effectively identifying complex anomaly hazards in server PCB operation.

[0094] The Timing Drift Intelligent Analysis Module, based on the corrected data stream, integrates the thermal anomaly points output by the Thermal Field Dynamic Distribution Module and the signal anomaly points output by the Signal Integrity Module to conduct focused detection of timing drift in high-speed data channels. This module utilizes enhanced sliding window PLL phase locking technology and, based on the sliding window timing drift detection formula, monitors and analyzes the phase information in the data stream in real time, accurately capturing potential timing offsets and phase jitter.

[0095] Compared to traditional timing analysis methods, this module significantly improves the cutting-edge nature and accuracy of timing analysis by scientifically setting the sliding window size and phase lock threshold. Once an abnormal timing change is detected, it quickly outputs a timing drift risk tag, providing key evidence for subsequent fault diagnosis. This effectively enhances the ability to early-warn anomalies in high-speed data transmission on server PCBs, effectively ensuring the stability and reliability of data transmission.

[0096] The multidimensional joint feature extraction module integrates multi-source data, such as electrical parameters and signal integrity labels, output by the first five modules. This module breaks through the limitations of traditional single-variable extraction and proposes a feature embedding path for coupled data. Through feature embedding technology, data of different dimensions and properties are mapped into the same feature space, breaking down data type barriers, enhancing compatibility and analyzability, and establishing a unified foundation for deep feature extraction.

[0097] After data fusion and feature embedding, the principal component analysis (PCA) algorithm is used to reduce the data dimension based on the principal component feature extraction formula. Key features are screened to eliminate redundancy, and feature vectors that fully reflect the server PCB operating status are extracted and output to the anomaly identification module. This module, through multi-source data fusion and feature dimensionality reduction, provides core data support for accurately identifying PCB operating anomalies and enhances the system's comprehensive analysis capabilities.

[0098] The Adaptive Abnormal Pattern Recognition Module receives the joint feature vector output by the Multi-Dimensional Joint Feature Extraction Module and uses a dual architecture, a LSTM recursive neural network and a density clustering algorithm, to collaboratively process the data. The LSTM uses a time-recursive formula to deeply mine the time series characteristics of the data, comprehensively analyzing historical and current data to identify abnormal patterns. The density clustering algorithm distinguishes different types of anomalies based on the density of the data distribution. The combination of the two overcomes the limitations of a single algorithm and significantly improves the ability to identify complex anomalies.

[0099] Based on the collaborative operation of dual algorithms, the module can accurately identify complex types of server PCB abnormalities such as sudden anomalies, gradual aging and periodic anomalies. Through in-depth analysis of feature vectors, it forms detailed classification results and promptly feeds back to the fault prediction module, providing key information for predicting potential fault development paths, ensuring the accuracy and reliability of PCB abnormality diagnosis, and realizing intelligent classification and early warning of server operation status.

[0100] The fault trend prediction module, based on anomaly pattern results and historical data streams, constructs a dual-model architecture consisting of a dynamic Bayesian network and a time series prediction model. The dynamic Bayesian network uses dynamic Bayesian prediction formulas to model fault probability and causal relationships. This model can simulate the multi-path evolution of faults under diverse conditions, overcoming the limitations of simple threshold predictions. The time series prediction model, based on the changing patterns of historical data, accurately predicts future operating states. The two complement each other from the perspectives of probability analysis and trend deduction, providing multi-perspective data support for fault prediction.

[0101] Working together, the two models can deeply explore the potential development paths of server PCB failures and output accurate fault risk levels in real time. The prediction results directly trigger the automatic alarm module, driving a rapid system response and providing operational personnel with a scientific basis for decision-making. This allows them to take timely measures to eliminate potential faults, thereby reducing losses from PCB failures and effectively improving the stability and reliability of server system operations.

[0102] The self-learning feedback and online optimization module establishes a complete data feedback loop from automatic alarms, actual operations and maintenance, to fault verification, completely different from traditional offline training methods. Based on automatic alarm results, actual operations and maintenance, and fault verification information, this module dynamically adjusts the entire detection system in real time. By flexibly adjusting anomaly thresholds, the system can adapt to diverse operating environments and complex fault types, laying a solid foundation for accurate detection.

[0103] This module innovatively utilizes an online incremental learning approach. Based on feedback data, it applies an incremental learning weight update formula to fine-tune model weights, significantly improving model accuracy and reliability. Furthermore, it flexibly adjusts feature extraction priorities based on actual needs, focusing on key feature analysis to effectively improve detection efficiency. Through continuous online incremental learning and dynamic optimization, the system's adaptability is continuously enhanced, ensuring efficient and accurate detection capabilities throughout long-term operation, enabling the detection system to evolve and continuously optimize.

[0104] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0105] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The server PCB automatic detection system based on operation data collection and analysis is characterized by: include: Distributed data acquisition module for operation: Collection units are placed at key nodes of the PCB, electrical data is collected in parallel using FPGA hardware, and the original data is transmitted synchronously via a high-speed bus; Preprocessing and synchronization timing correction module: After receiving the raw data, it uses time alignment and multi-clock domain compensation algorithms based on the timestamp to preprocess and synchronize the data, and filter and remove noise; Signal integrity real-time detection module: After receiving pre-processed and corrected data, it uses wavelet denoising and reflection attenuation detection algorithms to analyze signal waveform parameters, identify distortion and abnormal characteristics, and output signal integrity labels; Thermal field dynamic distribution analysis module: Based on the synchronously corrected data and signal integrity labels, a dynamic thermal field fitting algorithm is used to construct a thermal field distribution map, associate high-temperature points with signal anomaly points, and output multi-source coupling anomaly signs; Timing Drift Intelligent Analysis Module: Based on the corrected data stream, combined with thermal anomalies and signal anomalies, it uses sliding window PLL phase locking technology to monitor the phase information of high-speed data channels, detect timing offset and jitter, and output risk labels. Multi-dimensional joint feature extraction module: uses feature embedding and principal component analysis algorithms to map data into the same space and reduce dimensionality to extract comprehensive feature vectors; Abnormal pattern adaptive recognition module: Receives feature vectors, uses LSTM recursive neural networks and density clustering algorithms to learn data time series features, classifies data based on distribution density, and identifies abnormal types; Fault trend deep prediction module: Based on the results of the abnormal pattern adaptive recognition module and historical data, it builds a dynamic Bayesian network and time series prediction model to analyze the fault probability and development trend, and outputs risk level-driven alarms; Self-learning feedback and online optimization module: Dynamically adjusts abnormality thresholds, model weights, and feature extraction priorities based on alarm and fault verification results.

2. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The distributed operation data collection module includes: (1) Scientifically arrange the acquisition units at multiple nodes: Scientifically arrange the acquisition units at key nodes around the server PCB’s power supply lines, high-frequency signal transmission lines, and high-heat chips; (2) Use FPGA to achieve high-speed parallel acquisition: Each acquisition unit is equipped with high-performance FPGA hardware to collect current, voltage, temperature, signal waveform and data transmission timing, and cooperate with high-precision Hall current sensor, isolation op amp circuit and digital temperature sensor, and the multi-node data synchronous sampling formula ensures the consistency of data sampling at each node.

3. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The preprocessing and synchronization timing correction module includes: (1) Accept data and clarify processing objectives: After receiving the multi-source electrical data output by the distributed operation data acquisition module, pre-process the raw data and perform synchronous timing correction; (2) Realize time synchronization and multi-clock domain correction: Based on the timestamp, use the time alignment algorithm and the time alignment correction formula to make the data uniform on the time axis. Use multi-clock domain compensation technology to finely adjust the clock signal according to the acquisition delay error, eliminate time deviation, filter the original data, and remove noise interference.

4. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The signal integrity real-time detection module includes: (1) Joint algorithm for deep processing of corrected data: The data after receiving pre-processing and synchronization timing correction is processed using a wavelet denoising + reflection detection joint algorithm. The wavelet denoising algorithm can effectively decompose the signal frequency band and accurately remove noise; the reflection detection algorithm monitors the signal transmission reflection in real time; (2) Accurately identify anomalies and output key labels: Calculate and compare the amplitude, phase, and rise time parameters of high-frequency signals, and use joint algorithms to capture signal distortion, crosstalk, and reflection anomaly characteristics in real time. When an anomaly is detected, the signal integrity label is immediately output.

5. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The thermal field dynamic distribution analysis module includes: (1) Constructing a thermal field distribution map based on multi-source data: Based on the synchronously corrected temperature data and signal integrity labels, the thermal field anomaly fitting formula is used to construct a PCB thermal field distribution map by correlating the temperature data with the signal integrity data; (2) Coupling modeling to identify potential failure risks: By analyzing the thermal field distribution map, high-temperature points are accurately located and associated with signal abnormality points to achieve signal-thermal abnormality coupling modeling.

6. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The timing drift intelligent analysis module includes: (1) Multi-source data fusion and detection technology: Based on the corrected data stream, thermal anomalies and signal anomalies are integrated, and sliding window PLL phase locking technology is used. According to the sliding window timing drift detection formula, the data stream phase information is monitored in real time; (2) Accurately identify anomalies and output key labels: By setting the adaptive sliding window size and phase lock threshold, when abnormal timing changes are detected, the timing drift risk label is immediately output.

7. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The multi-dimensional joint feature extraction module includes: (1) Multi-source data fusion and innovative feature embedding: After receiving the multi-type data of electrical parameters and signal integrity labels output by the first five modules, feature embedding technology is used to map data of different dimensions and properties into the same feature space; (2) Principal component analysis: Use the principal component analysis algorithm to deeply process the data. Based on the principal component feature extraction formula, perform dimensionality reduction on the fused data to screen out key features, eliminate data redundancy, and extract feature vectors.

8. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The abnormal pattern adaptive recognition module includes: (1) Dual algorithms collaboratively process feature vectors: After receiving the feature vectors, a dual algorithm architecture of LSTM recursive neural network and density clustering is introduced. LSTM uses a time recursive formula to deeply mine the time series features of the data; (2) Accurately classify anomalies and output results: Based on the collaborative operation of LSTM and density clustering algorithms, it accurately identifies sudden anomalies, progressive aging, and periodic anomaly types of server PCBs. Through in-depth analysis of feature vectors, it forms anomaly classification results and feeds the results back to the fault prediction module.

9. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The fault trend depth prediction module includes: (1) Dual model construction to achieve intelligent prediction: Based on abnormal pattern results and historical data streams, dynamic Bayesian network and time series prediction model are introduced to model and analyze the probability of failure and causal relationship; (2) Collaborative prediction helps troubleshooting: It can deeply analyze the potential fault development path of the server PCB and output accurate fault risk level in real time. The prediction results directly drive the automatic alarm module.

10. The server PCB automatic detection system based on operation data collection and analysis according to claim 1 is characterized in that: The self-learning feedback and online optimization module includes: (1) Build a data feedback closed loop to achieve real-time adjustment: Dynamically adjust the abnormality threshold based on the results of the automatic alarm module, actual operation and maintenance conditions, and subsequent fault verification results; (2) Innovative online incremental learning to improve system performance: clarify the online incremental learning path, use the incremental learning weight update formula based on feedback data to optimize the model weight, adjust the feature extraction priority according to actual needs, and focus on key features.

Images