Industrial computer predictive maintenance method based on artificial intelligence

Abstract

The invention provides an artificial intelligence-based industrial computer predictive maintenance method, which comprises the following steps of: acquiring real-time monitoring values of clock signal jitter, signal integrity parameters and electromagnetic compatibility characteristics from mainboard circuit operation data to form multi-dimensional time sequence data, and performing feature extraction to obtain initial distribution of abnormal signals related to aging; carrying out primary distribution clustering analysis on the abnormal signals to obtain signal abnormal groups in an aging mode, and carrying out weighted analysis on interaction influence of clock signal jitter and microcrack propagation in the groups by adopting a preset convolution kernel to obtain abnormal signal group features; and identifying interference factors influencing system stability evaluation from the electromagnetic compatibility characteristic variation coupling characteristic data, and classifying the interference factors in combination with the priority initial sequence.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of information technology, and in particular to an artificial intelligence-based predictive maintenance method for industrial computers. Background Art

[0002] The application of artificial intelligence in predictive maintenance for industrial computers is becoming a key technological enabler for improving equipment reliability and reducing operating costs. With the increasing complexity of industrial systems, the aging of motherboard circuits is becoming increasingly prominent, directly impacting system stability and long-term operational performance. Especially under high loads and long operating times, circuit aging not only leads to performance degradation but can also cause systemic failures. Therefore, research on monitoring and maintenance strategies for these circuits is of considerable practical significance. Traditional maintenance methods rely on periodic inspections or passive repairs based on experience, which are often inefficient and lack foresight when dealing with dynamic aging processes. While existing monitoring technologies can identify circuit anomalies to a certain extent, they are insufficient to capture the microscopic changes in motherboard aging, particularly the complex electrical characteristic changes caused by the interaction of clock signal jitter and material fatigue, making accurate prediction and timely intervention difficult. Current research still faces several key challenges, the most prominent of which include the dynamic deterioration of signal integrity parameters, the uncontrollable increase in the accumulation rate of clock offset, and the nonlinear variation of electromagnetic compatibility characteristics. These factors interact with each other during the growth of microcracks in circuit boards, making traditional static threshold detection methods difficult to adapt to the real-time changes of the aging process. Deteriorating signal integrity reduces data transmission reliability, increasing clock offset accumulation rates further exacerbate synchronization errors, and variations in electromagnetic compatibility characteristics can amplify external interference. These unresolved technical challenges collectively constitute bottlenecks in the accuracy and priority determination of predictive maintenance. Therefore, how to dynamically adjust the criteria for determining electrical characteristic anomalies through convolutional neural networks and optimize repair priority ranking to accommodate the changing correlation between signal jitter and material fatigue during motherboard circuit aging has become a key issue in improving overall system stability evaluation metrics. Solving this problem requires accurate identification and adaptive adjustment of electrical parameter anomalies in complex and changing aging environments, thereby providing more intelligent technical support for the reliable operation of industrial computers. Summary of the Invention

[0003] The present invention provides an artificial intelligence-based predictive maintenance method for industrial computers, which mainly includes:

[0004] Real-time monitoring values ​​of clock signal jitter, signal integrity parameters, and electromagnetic compatibility characteristics are obtained from the motherboard circuit operation data to form multi-dimensional time series data. Feature extraction is performed to obtain the preliminary distribution of aging-related abnormal signals.

[0005] A preliminary cluster analysis of the abnormal signal distribution was performed to obtain abnormal signal groups under the aging mode. A preset convolution kernel was used to perform a weighted analysis of the interaction between clock signal jitter and microcrack propagation within the group to obtain the abnormal signal group characteristics.

[0006] Extract the coupling characteristics of the electromagnetic compatibility characteristics variation of the corresponding signal abnormality group from the signal integrity parameters of each group, and compress the dimension of the coupling characteristic data by combining the abnormal signal group characteristics to obtain the refined expression of the abnormal signal of each group;

[0007] Identify the clock jitter offset rate of the abnormal signal expressed in each refined group. If the clock jitter offset rate exceeds the preset offset threshold, identify the signal deterioration trend caused by microcrack expansion in the corresponding group, and perform nonlinear fitting to identify the distribution of acceleration points.

[0008] Obtain the statistical characteristics of abnormal signals corresponding to the distribution of acceleration points of each aging mode in different time windows. By comparing the changes in statistical characteristics of different time windows, the evolution law of abnormal signals in different time windows is obtained. The long short-term memory network is used to perform time series prediction on the evolution law and determine the initial priority sequence of maintenance sorting.

[0009] Identify interference factors that affect system stability assessment from electromagnetic compatibility characteristic variation coupling feature data, and classify interference factors based on the initial priority sequence;

[0010] According to the interference factor classification, the abnormal judgment standard threshold of aging pattern recognition is adjusted, the matching degree between the priority sequence change before and after adjustment and the improvement of system stability is compared, the high-priority aging abnormality area is obtained, the electrical characteristics adjustment instruction is generated, and the real-time maintenance plan of the mainboard circuit is determined.

[0011] Furthermore, real-time monitoring values ​​of clock signal jitter, signal integrity parameters, and electromagnetic compatibility characteristics are obtained from the motherboard circuit operation data to form multidimensional time series data. Feature extraction is performed to obtain a preliminary distribution of aging-related abnormal signals. This includes: collecting clock signal jitter data from the phase error sequence based on the clock signal sampling rate during the operation of the motherboard circuit; collecting signal integrity parameters of key nodes in the circuit based on the signal integrity transmission rate; collecting electromagnetic compatibility characteristics based on the electromagnetic field intensity threshold; and calculating the signal distortion to obtain a first time series data set. The first time series data set is normalized using the maximum and minimum normalization method to standardize the voltage fluctuation amplitude, signal attenuation, and motherboard current change rate to obtain a second time series data set. Feature vectors are constructed based on the voltage fluctuation amplitude and signal attenuation in the second time series data set. Time series clustering of the feature vectors is performed using the radial basis kernel function of the support vector machine clustering algorithm to obtain a first abnormal feature cluster. For the first abnormal feature cluster, electromagnetic interference intensity and jitter frequency distribution features are extracted. Anomaly detection rules are constructed using an anomaly detection algorithm based on a Gaussian distribution. The abnormality threshold interval is divided according to the signal transmission delay time to obtain a second abnormal feature cluster. An aging feature vector is constructed for the second abnormal feature cluster and classified using the multi-layer perceptron structure in the convolutional neural network. The signal integrity transmission rate is used to calculate the aging correlation matrix, and the circuit aging distribution map is obtained based on the voltage fluctuation amplitude and signal attenuation degree.

[0012] Furthermore, a cluster analysis of the initial distribution of abnormal signals is performed to obtain signal anomaly groups under the aging model. A preset convolution kernel is used to perform a weighted analysis of the interaction between clock signal jitter and microcrack growth within the group to obtain abnormal signal group characteristics. This includes: constructing an original feature matrix based on clock jitter amplitude and microcrack length, normalizing the feature matrix using a maximum-minimum normalization method to obtain a first standardized feature matrix. Statistics of signal anomaly frequency and duration are calculated for the first standardized feature matrix. Time series statistical features are extracted using a sliding time window, and these time series statistical features are grouped using a hierarchical clustering method to obtain a first feature group set. An association mapping is established between the time series correlation and signal fluctuation trend within the first feature group set. A radial basis function in a support vector machine is used to perform a mapping transformation on the crack stress distribution data, and a second feature group set is obtained based on the signal cluster distance. A Gaussian convolution kernel weight matrix is ​​applied to the second feature group set. A multi-scale convolution transformation is performed on the clock signal jitter data using the jitter propagation path, and a weighted superposition is performed on the crack growth rate data to obtain a first weighted feature vector. The stress concentration index is extracted according to the first weighted eigenvector, and the stress concentration index is grouped using a clustering method based on local density. The interactive feature term is constructed through the coupling relationship between the clock jitter amplitude and the crack growth rate, and the abnormal signal characteristic distribution is obtained.

[0013] Furthermore, coupling features related to the electromagnetic compatibility characteristic variations of the corresponding signal anomaly group are extracted from the signal integrity parameters of each group. The dimensions of the coupling feature data are compressed using the abnormal signal group features to obtain a refined representation of the abnormal signals in each group. This includes: constructing an original feature matrix based on signal integrity and electromagnetic interference intensity, normalizing the feature matrix using a maximum-minimum normalization method to obtain a first standardized feature matrix. A feature map is established based on the degree of waveform distortion and parameter variation amplitude in the first standardized feature matrix. Compressed features are extracted using the encoding layer of a multi-layer autoencoder, and the first compressed feature matrix is ​​obtained based on the reconstruction error. Coupling metrics are calculated based on the signal transmission stability and anti-interference capability in the first compressed feature matrix. The coupling metrics are grouped using a density-based adaptive clustering method, and a second compressed feature matrix is ​​obtained based on parameter clustering distance. A covariance representation of electromagnetic radiation intensity is constructed for the second compressed feature matrix. Eigenvalue decomposition is used to extract the main feature directions. Eigenvectors are selected based on the cumulative contribution rate to obtain a first principal component matrix. Information gain is calculated for each feature dimension based on the first principal component matrix. Key dimensions are extracted using a mutual information-based feature selection method. Coupling feature vectors are constructed based on the abnormal frequency distribution to obtain a refined feature representation.

[0014] Furthermore, the clock jitter offset rate of the abnormal signal expressed in each refined group is identified. If the clock jitter offset rate exceeds a preset offset threshold, the signal deterioration trend caused by microcrack propagation in the corresponding group is identified, and nonlinear fitting is performed to identify the distribution of acceleration points. This includes: extracting clock jitter data based on the refined group expression, eliminating measurement noise through median filtering, and synchronizing the data using timestamp alignment to obtain a first jitter sequence. For the first jitter sequence, the offset rates of adjacent sampling points are calculated, the offset rates are normalized within a preset threshold range, and outliers are marked according to the triple standard deviation principle to obtain a first rate sequence. Continuous outlier intervals are extracted from the first rate sequence, and a mapping relationship between crack length and jitter rate is established using support vector regression based on a radial basis kernel function. The kernel function parameters are determined through cross-validation to obtain a first trend function. For the first trend function, the local rate of change is calculated, the trend slope characteristics are extracted using a sliding differential window, and a stress intensity factor is constructed based on the crack propagation rate to obtain a second trend function. The second trend function was loaded into a Gaussian process regression model, and a nonlinear mapping was constructed using the mean and covariance functions. A probability density estimate was performed on the stress intensity factor to obtain the third trend function. Local acceleration regions were extracted based on the third trend function, and the acceleration points were spatially grouped using a density clustering method. A damage evolution curve was constructed based on the degree of signal degradation, resulting in a distribution map of the acceleration points.

[0015] Furthermore, the statistical characteristics of the abnormal signals corresponding to the distribution of acceleration points of each aging mode are obtained in different time windows. By comparing the changes in statistical characteristics in different time windows, the evolution pattern of the abnormal signals in different time windows is obtained. The evolution pattern is then predicted using a long short-term memory network to determine the initial priority sequence for maintenance sorting. This includes: data quality assessment based on the distribution of aging mode acceleration points, outliers are removed using a median filter, and the data is normalized using the standard deviation method to obtain a first data sequence. A fixed-length sliding window is constructed for the first data sequence, the sampling interval is set by the window sliding step, and the mean, variance, and kurtosis values ​​within the window are calculated based on the characteristic statistical period to obtain the first characteristic sequence. Based on the first characteristic sequence, the characteristic difference values ​​between adjacent time windows are calculated, a baseline threshold is established based on the fault evolution speed, and mutation points are marked based on the signal abnormality to obtain a second characteristic sequence. For the second characteristic sequence, time series features are extracted in segments, key variables are selected using a gradient-based feature screening method, and training samples are constructed based on the time series change trend to obtain the first training matrix. The first training matrix is ​​loaded into the long-short-term memory network structure. The hidden layer state vector records the temporal dependencies. A prediction model is established based on the signal variation pattern to obtain the first prediction sequence. Aging acceleration indicators are extracted from this first prediction sequence. The hazard level is calculated using a quantitative fault severity assessment method. A priority scoring function is constructed based on maintenance resource constraints to obtain the initial priority sequence.

[0016] Furthermore, based on the interference factor classification, the threshold for abnormality determination criteria in aging pattern recognition is adjusted. The change in the priority sequence before and after the adjustment is compared with the degree of match between the system stability improvement and the high-priority aging abnormality area. Electrical characteristic adjustment instructions are generated, and a real-time maintenance plan for the mainboard circuit is determined. This process includes: normalizing the raw interference data according to the interference classification level, eliminating measurement noise using a median filter, and segmenting the data based on the abnormal area to obtain a first interference feature matrix. Based on the first interference feature matrix, an aging severity index is calculated. The threshold adjustment function is parameter-optimized using a recursive iterative method. The adjustment step size is determined based on the statistical characteristics of historical data to obtain a first threshold adjustment matrix. A priority change curve is constructed based on the first threshold adjustment matrix. The difference between the sequence before and after the threshold adjustment is calculated using a stability index. A matching quantization function is established based on the difference data to obtain a second threshold adjustment matrix. Based on the second threshold adjustment matrix, the electrical parameter variation pattern is extracted. A recursive neural network is used to perform feature mapping on the parameter space. Based on the mapping results, a parameter correction vector is constructed to obtain a first parameter optimization sequence. Based on the first parameter optimization sequence, maintenance instruction generation rules are established. The electrical characteristic correction is converted using an instruction template library. Based on the conversion results, an instruction execution sequence is constructed to obtain a second parameter optimization sequence. The stability evaluation index is calculated for the second parameter optimization sequence, the maintenance instructions are prioritized using a priority sorting algorithm, and a scheduling sequence is generated based on the dependencies between instructions to obtain a real-time maintenance plan.

[0017] Furthermore, interference factor classification is used to identify high-priority aging anomaly areas. The location coordinates, abnormal signal type, and severity of these high-priority aging anomaly areas are identified. Circuit simulation verification is then performed based on the adjustment ranges and step sizes of voltage, current, and impedance. Electrical characteristic adjustment instructions are generated, and maintenance time windows and frequencies are determined for each anomaly area based on the severity of the anomaly. Dynamic maintenance scheduling is then developed for the target time period. This process involves standardizing and preprocessing the raw data according to the interference factor classification rules, eliminating noise interference through median filtering, and extracting features from the abnormal signals to generate a first feature sequence. The coordinates of the abnormal areas are extracted from the first feature sequence, and the anomaly type and severity are classified and labeled using a region location algorithm. High-risk areas are screened based on priority thresholds to generate a first abnormal area table. A circuit parameter adjustment matrix is ​​constructed based on the first abnormal area table. Constraints are established using the voltage adjustment range and current variation range, and adjustment step sizes are set based on impedance matching criteria to generate a first parameter adjustment table. The first parameter adjustment table is then loaded into a support vector regression algorithm to construct a simulation mapping function. The parameter adjustment solution is validated using the regression results, and the optimization direction is determined based on performance evaluation indicators to generate a second parameter adjustment table. Based on the second parameter adjustment table, electrical characteristic adjustment instructions are generated. The adjustment steps are converted into code using an instruction template library. Conflict detection is performed on the instruction sequence to obtain a first maintenance instruction set. Maintenance time estimates are calculated for the first maintenance instruction set. Time windows are divided using a maintenance cycle optimization algorithm. The execution frequency is determined based on the severity of the anomaly to obtain a first maintenance schedule. A resource allocation matrix is ​​constructed based on the first maintenance schedule. A scheduling optimization algorithm is used to map maintenance tasks to timelines. Dynamic adjustments are made to the priority sequence to obtain a dynamic maintenance scheduling solution.

[0018] The technical solutions provided by the embodiments of the present application may have the following beneficial effects:

[0019] The embodiment of the present application discloses a predictive maintenance method for industrial computers based on artificial intelligence. The method forms multidimensional time series data by real-time monitoring of clock signal jitter, signal integrity parameters and electromagnetic compatibility characteristics in the mainboard circuit operation data. Feature extraction and cluster analysis are performed on these data to identify signal anomalies in the aging mode. The preset convolution kernel is used to analyze the interactive influence of clock signal jitter and microcrack extension, and a refined expression of the abnormal signal is obtained in combination with the variation of electromagnetic compatibility characteristics. The evolution law of the abnormal signal is predicted in time series by a long short-term memory network to determine the maintenance priority. Finally, according to the interference factor in the variation of electromagnetic compatibility characteristics, the abnormality judgment standard is adjusted, the electrical characteristic adjustment instruction is generated, and the real-time maintenance plan is determined. The embodiment of the present application can effectively identify and predict the aging anomaly of the mainboard circuit, improve the stability of the system, and realize proactive preventive maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flowchart of an artificial intelligence-based predictive maintenance method for industrial computers according to an embodiment of the present application.

[0021] Figure 2 This is a schematic diagram of an artificial intelligence-based predictive maintenance method for industrial computers according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly and in detail describe the technical solutions in the embodiments of the present application. The described embodiments are only part of the embodiments of the present application.

[0023] like Figure 1-2 In this embodiment, an artificial intelligence-based predictive maintenance method for industrial computers may specifically include:

[0024] S101: Obtain operating data of a mainboard circuit, wherein the operating data of the mainboard circuit includes at least real-time monitoring values ​​of clock signal jitter, signal integrity parameters, and electromagnetic compatibility characteristics collected at different time points.

[0025] By collecting real-time monitoring values ​​of clock signal jitter, signal integrity parameters and electromagnetic compatibility characteristics from the operating data of the mainboard circuit, multi-dimensional time series data is generated and feature extraction is performed to obtain the preliminary distribution of abnormal signals related to aging.

[0026] In this embodiment, the motherboard circuit's clock signal is first monitored in real time through high-frequency sampling to obtain jitter data. Signal integrity parameters and electromagnetic compatibility characteristics of key nodes are simultaneously collected to form multidimensional time series data. Feature extraction techniques are then used to identify abnormal signal distributions related to aging. The specific collection and processing procedures can be flexibly adjusted according to actual scenarios.

[0027] S1011. Collect a phase error sequence according to the clock signal sampling rate during operation of the mainboard circuit to obtain jitter data, and collect signal integrity parameters and electromagnetic compatibility characteristic values ​​in combination with the signal integrity transmission rate and the electromagnetic field strength threshold to generate a first time series data set.

[0028] In an embodiment of the present application, the clock signal jitter is collected using a high-frequency sampling method, by recording a phase error sequence near the rising edge of the clock signal, and the sampling rate is set to 32 times the clock frequency. For example, when the clock frequency is 100MHz, the sampling interval is 0.3125ns, and the phase error sequence records jitter values ​​within the range of ±50ps. Signal integrity parameters are collected by arranging probe points at key nodes of the circuit to record characteristics such as signal rise time, fall time, and overshoot value. Electromagnetic compatibility characteristics are measured by scanning the surface of the circuit board through a near-field probe array, measuring the local electromagnetic field intensity distribution, and generating a first time series data set to provide basic data for analysis.

[0029] S1012. Standardize the voltage fluctuation amplitude, signal attenuation degree, and mainboard current change rate in the first time series data set, use the support vector machine clustering algorithm to perform time series clustering on the standardized data, extract the first abnormal feature cluster and construct an aging feature vector, classify and calculate the aging correlation matrix through the convolutional neural network, and generate a circuit aging distribution map.

[0030] In an embodiment of the present application, data preprocessing is performed on the first time series data set, and the voltage fluctuation amplitude is standardized to the [-3, 3] interval using the maximum and minimum normalization method, the signal attenuation degree is mapped to the [0, 1] interval, and the mainboard current change rate is normalized based on the rated current. Subsequently, a feature vector is constructed based on the voltage fluctuation amplitude and the signal attenuation degree, and the radial basis kernel function in the support vector machine is used for time series clustering. The kernel function parameter γ is set to 0.1 to obtain the first abnormal feature cluster. The electromagnetic wave interference intensity and jitter frequency distribution characteristics are extracted for the first abnormal feature cluster, and an aging feature vector is constructed. Classification is performed through a three-layer convolutional neural network. The input layer receives a 32×32 feature map, and high-level features are extracted through the convolution layer and the fully connected layer. The aging correlation matrix is ​​calculated based on the signal complete transmission rate, and the circuit aging distribution map is generated by combining the voltage fluctuation amplitude and the signal attenuation degree.

[0031] In an embodiment of the present application, S101 can effectively identify the preliminary distribution of abnormal signals during the aging process of the mainboard circuit through multi-dimensional data acquisition and feature extraction, laying the foundation for subsequent cluster analysis and abnormality judgment. During the acquisition process, the sampling frequency and detection point position can be dynamically adjusted according to the circuit operation status to meet the needs of different aging scenarios. The monitoring of signal integrity parameters focuses on the quality of the signal waveform. For example, the signal rise time is controlled within 70ps, and the overshoot value does not exceed 15% of the amplitude. If it exceeds, it is marked as abnormal. In the electromagnetic compatibility characteristic monitoring, when the electromagnetic field strength exceeds 40dBμV / m, it is marked as a potential interference source, and it is aligned with the jitter data time to establish a correlation between interference and signal degradation. This multi-level analysis method ensures the comprehensiveness and accuracy of the distribution of abnormal signals.

[0032] In some optional embodiments, if the operating environment of the motherboard circuit changes, the standardization method or clustering algorithm parameters can be adjusted according to actual needs to further improve the accuracy of feature extraction. This embodiment of the application does not impose too many restrictions on the specific algorithm implementation details, and technicians can customize them according to the application scenario. The resulting circuit aging distribution map provides intuitive data support for subsequent aging pattern identification, effectively improving the targetedness and reliability of predictive maintenance.

[0033] S102. Perform cluster analysis on the preliminary distribution of abnormal signals to identify abnormal signal groups under the aging mode, use a preset convolution kernel to analyze the interactive effect of clock signal jitter and microcrack extension, and extract abnormal signal group features.

[0034] In an embodiment of the present application, based on the preliminary distribution of abnormal signals, the signal anomaly groups in the aging mode are divided by clustering technology, and the convolution kernel weighted method is used to extract the features of the coupling effect of clock signal jitter and microcrack extension to provide data support for maintenance decisions.

[0035] S1021. Construct an original feature matrix based on the clock jitter amplitude and the microcrack length and normalize it using the maximum and minimum value normalization method to generate a first standardized feature matrix. Then, extract the timing statistical features through a sliding time window and group the timing statistical features using a hierarchical clustering method to obtain a first feature grouping set.

[0036] In an embodiment of the present application, the clock signal jitter data and microcrack length data of the mainboard circuit are first collected to construct the original feature matrix. The collection of clock jitter data is achieved through high-frequency sampling, and the sampling rate is set to 32 times the clock frequency. For example, at a 100MHz clock, the sampling interval is 0.3125ns, and the jitter amplitude is recorded in the range of 10ps to 100ps. The microcrack length data is obtained by optical detection or ultrasonic scanning, and the typical value fluctuates between 0 and 1mm. The original feature matrix is ​​normalized by the maximum and minimum values, and the jitter amplitude is normalized to the [-1,1] interval, and the microcrack length is mapped to [0,1]

[0037] interval to eliminate dimensional differences. Subsequently, a 60-second sliding time window was used to extract time series statistical features, with the window overlap rate set to 50%. The frequency of signal anomalies and the duration of anomalies within the window were calculated. For example, when the frequency of anomalies exceeded 10 times per minute and the duration exceeded 100 microseconds, the window was marked as an abnormal interval. The hierarchical clustering method used the Ward minimum variance method to achieve grouping by minimizing the intra-cluster variance. The number of clusters was dynamically determined by the silhouette coefficient, usually between 3 and 5 clusters, to generate the first feature grouping set, laying the foundation for analysis.

[0038] S1022. Establish an association mapping for the timing correlation and signal fluctuation trend in the first feature grouping set and use the radial basis function in the support vector machine to perform a nonlinear mapping transformation on the crack stress distribution data, generate a second feature grouping set based on the signal clustering distance, and then load the Gaussian convolution kernel weight matrix to perform a multi-scale convolution transformation on the clock signal jitter data and perform weighted superposition on the crack propagation rate data to extract the first weighted feature vector.

[0039] In an embodiment of the present application, an in-depth analysis is conducted on the first feature grouping set, and an association mapping is established by calculating the correlation coefficient between the timing correlation and the signal fluctuation trend. When the correlation coefficient exceeds 0.7, it is considered that there is a significant coupling relationship. The crack stress distribution data is obtained by finite element analysis, and the stress concentration coefficient at the crack tip is recorded. Usually, when it is above 2.5, it indicates that the stress distribution is abnormal. The radial basis function in the support vector machine is used for nonlinear mapping, and the kernel function parameter γ is set to 0.1. The stress distribution characteristics are mapped to a high-dimensional space. The signal clustering distance is measured according to the Euclidean distance. The distance threshold is set to 2 times the standard deviation of the cluster center to generate a second feature grouping set. Next, a Gaussian convolution kernel weight matrix of size 5×5 is loaded, and the standard deviation is set to 1.5. The impact of the jitter propagation path on the clock signal jitter data is analyzed by multi-scale convolution transformation. The edge processing uses mirror filling to retain boundary information. The crack growth rate data is weighted and superimposed. The higher the rate, the greater the weight. For example, when the rate reaches 0.1 mm / hour, the weight increases to 2.0. Finally, the first weighted eigenvector is extracted to reflect the interactive characteristics of vibration and crack growth.

[0040] In some optional embodiments, a stress concentration index is extracted based on the first weighted eigenvector and grouped using a local density-based clustering method. An interactive feature item is constructed through the coupling relationship between the clock jitter amplitude and the crack propagation rate to generate an abnormal signal feature distribution. The stress concentration index is calculated by the stress state at the crack tip. When the index value exceeds 3.0, it indicates the presence of severe stress concentration. The local density clustering method automatically identifies the cluster center based on the density distribution of data points, and combines cross-correlation analysis to determine the strong coupling area between jitter and crack propagation. For example, a 0.8mm long microcrack was detected in a high-speed DDR interface area, the clock jitter amplitude increased from 45ps to 68ps, the stress concentration coefficient reached 2.8, and the coupling coefficient was 0.85, indicating that the signal quality degradation in this area is closely related to stress damage.

[0041] In the embodiment of the present application, through multi-level clustering and convolution analysis, the interaction between clock signal jitter and microcrack extension can be accurately identified. The extraction process of abnormal signal feature distribution fully considers the impedance discontinuity caused by signal reflection. For example, in a 20-layer PCB board, when the crack length exceeds 0.5mm, the local signal integrity is significantly reduced and the jitter amplitude increases by more than 20%. This method not only improves the accuracy of feature extraction, but also provides a reliable data basis for subsequent aging pattern recognition. The convolution kernel size or clustering parameters can be adjusted according to the actual circuit complexity to meet the needs of different application scenarios. The specific implementation details can be flexibly set by technical personnel.

[0042] S103 , extracting coupling features with electromagnetic compatibility characteristic variations from the signal integrity parameters of each abnormal signal group, and compressing the dimensions of the coupling feature data in combination with the abnormal signal group features to generate a refined expression of the abnormal signal of each group.

[0043] In the embodiment of the present application, feature extraction is performed on the signal integrity parameters and electromagnetic compatibility characteristics of the signal anomaly group, and the data dimension is compressed by dimensionality reduction technology to generate a refined expression.

[0044] S1031. Construct an original feature matrix based on signal integrity and electromagnetic interference intensity and perform normalization processing using the maximum and minimum value normalization method to generate a first standardized feature matrix. Subsequently, the encoding layer of the multi-layer autoencoder extracts compression features based on the waveform distortion degree and parameter variation amplitude to obtain a first compressed feature matrix.

[0045] In an embodiment of the present application, signal integrity data, including parameters such as signal rise time, overshoot, and jitter, is collected from the operating data of the motherboard circuit. Electromagnetic interference intensity data, such as radiated field strength and conducted interference values, is also recorded to construct a multidimensional original feature matrix. Signal integrity is measured by eye diagram opening. For example, in high-speed circuits, an opening below 70% indicates signal quality degradation. Electromagnetic interference intensity is obtained through near-field scanning and typically ranges from 40dBμV / m to 60dBμV / m. The original feature matrix is ​​normalized to the [0, 1] interval using a maximum-minimum normalization method to ensure consistency of data across different dimensions. Feature compression is then performed using a multi-layer autoencoder. The autoencoder adopts a symmetrical structure, for example, with a 64-dimensional input layer, and encoding layers that are gradually reduced to 32, 16, and 8 dimensions, before the output layer is restored to 64 dimensions. Compressed features are extracted by minimizing the reconstruction error. When the reconstruction error is less than 0.05, the first compressed feature matrix retains the key information of the original data, providing a low-dimensional data foundation for analysis.

[0046] In some optional embodiments, a coupling metric of signal transmission stability and anti-interference capability is calculated based on the first compressed feature matrix and grouped using a density-based adaptive clustering method to generate a second compressed feature matrix. Subsequently, a covariance representation of the electromagnetic radiation intensity is constructed and the main eigendirections are extracted through eigenvalue decomposition to obtain a first principal component matrix. The coupling metric is calculated based on signal delay jitter and bit error rate. For example, when the delay jitter exceeds 15% of the unit interval or the bit error rate is higher than 10^-12, it indicates the presence of significant interference. The density clustering method determines the cluster center through local density estimation, and the cluster radius is set to 1.5 times the standard deviation of the parameter space to generate a second compressed feature matrix. The covariance matrix of the electromagnetic radiation intensity is calculated for the second compressed feature matrix to reflect the correlation between the measurement points. The main eigendirections are extracted through eigenvalue decomposition, and eigenvectors with a cumulative contribution rate of 85% are screened. Typically, 3 to 5 principal components are retained to form the first principal component matrix, further reducing data redundancy.

[0047] S1032. Calculate the information gain of each feature dimension according to the first principal component matrix and extract key dimensions using a feature selection method based on mutual information. Then, construct a coupled feature vector for the abnormal frequency distribution to generate a refined feature expression.

[0048] In an embodiment of the present application, an information gain analysis is performed on each characteristic dimension of the first principal component matrix, and the most discriminative features are screened based on the Shannon entropy calculation conditional entropy and information gain ratio. For example, the abnormal frequency distribution presents a bimodal feature, the low-frequency interference around 100MHz is related to the switching power supply ripple, and the high-frequency interference around 2.4GHz originates from the digital circuit harmonics, and the characteristic dimensions with mutual information values ​​exceeding 0.6 are retained. A refined feature expression is generated by constructing a coupling feature vector. For example, in a certain high-speed memory interface area, the electromagnetic coupling gain reaches 0.8, far exceeding the normal value of 0.3, resulting in a 25% decrease in signal integrity. Finally, the 2048-dimensional original data is reduced to a 16-dimensional refined expression, retaining more than 90% of the discriminant information. This efficient expression method significantly improves the computational efficiency of anomaly detection.

[0049] In the embodiment of the present application, the coupling relationship between signal integrity and electromagnetic compatibility characteristics can be effectively captured through multi-layer feature extraction and dimensionality reduction technology. The application of multi-layer autoencoders ensures that information loss is minimized during feature compression. For example, in 64-dimensional input data, the 8-dimensional coding layer can still reflect the main trends of signal distortion and interference. Density clustering and eigenvalue decomposition further optimize the data structure, so that the refined feature expression can significantly reduce the computational complexity while maintaining high discriminability. In practical applications, this method can quickly identify potential signal quality problems, such as transmission errors caused by electromagnetic interference in a certain area, and provide an accurate basis for maintenance. The number of autoencoder layers or clustering parameters can be adjusted according to the circuit operation requirements to adapt to the specific requirements of different scenarios, and the specific implementation method can be flexibly set by technical personnel.

[0050] S104. Identify the clock jitter offset rate of the abnormal signal in each group refined expression, analyze the signal deterioration trend caused by microcrack expansion when the offset rate exceeds a preset threshold, and identify the acceleration point distribution through a nonlinear fitting method to optimize maintenance decisions.

[0051] In an embodiment of the present application, by dynamically monitoring the clock jitter offset rate and combining it with the nonlinear characteristics of crack propagation, a signal degradation trend model is constructed and the acceleration point distribution is extracted, providing an accurate basis for identifying abnormal aging areas of industrial computers.

[0052] In some optional embodiments, clock jitter data is extracted from the refined representation and measurement noise is removed using a median filter. Timestamp alignment is then performed to generate a first jitter sequence. The offset rates of adjacent sampling points are then calculated and outliers are marked using normalization and the triple standard deviation principle to obtain a first rate sequence. The sampling frequency for the clock jitter data is set to 100 MHz, and the phase offset value is recorded at each sampling point. Under normal circumstances, the jitter value fluctuates within a range of ±50 ps, ​​and the measurement noise is approximately 5 ps. A median filter is used for smoothing using a 5-point sliding window to remove the effects of sudden noise. Timestamp alignment ensures that the synchronization accuracy of data at different measurement points is better than 1 ns. The offset rate is calculated based on the difference between adjacent sampling points. Under normal circumstances, the rate variation does not exceed 2 ps / μs. A preset threshold range is set based on historical data statistics, and the offset rate is normalized to the interval [-1, 1]. If the rate value exceeds three standard deviations and five consecutive points are abnormal, they are marked as outliers, forming the first rate sequence, providing basic data for trend analysis.

[0053] S1041. Based on the first rate sequence, the radial basis kernel function in support vector regression is used to establish a mapping relationship between crack length and jitter rate, and the kernel function parameters are optimized through cross-validation to generate a first trend function. Subsequently, the local change rate is calculated and the trend slope characteristics are extracted through a sliding differential window, and the second trend function is constructed in combination with the stress intensity factor.

[0054] In an embodiment of the present application, the nonlinear relationship between crack length and jitter rate is analyzed using the first rate sequence, and a mapping model is constructed using the support vector regression method. The radial basis kernel function controls the model complexity through the kernel parameter γ, and the γ value is determined by ten-fold cross validation. For example, when the optimized γ is 0.05, the model prediction accuracy is the highest. The crack length data is obtained through optical or ultrasonic testing, and the typical range is 0 to 1 mm. The jitter rate reflects the signal degradation rate. For example, when the crack length in a DDR interface area is 0.8 mm, the rate increases from 1.5 ps / μs to 4.8 ps / μs. After the first trend function is generated, the local change rate is calculated through a 20-point sliding window, and the trend slope feature is extracted. When the slope exceeds 0.5, it indicates that the crack extension has entered the acceleration stage. The stress intensity factor is calculated based on the crack length and material properties. It increases exponentially with the crack growth. When the critical value is close to 20 MPa√m, the extension rate is significantly accelerated. The second trend function further characterizes the coupling effect of signal degradation and material damage by incorporating the stress intensity factor.

[0055] S1042. Input the second trend function into the Gaussian process regression model and construct a nonlinear mapping through the mean function and covariance function to generate the third trend function. Then, extract the local acceleration area and use the density clustering method to spatially group the acceleration points. Construct a damage evolution curve according to the degree of signal degradation to generate an acceleration point distribution map.

[0056] In an embodiment of the present application, the second trend function is optimized using a Gaussian process regression model. The mean function describes the overall trend of the jitter rate. The covariance function uses a square exponential form to characterize the local correlation between sampling points. Model training is based on maximum likelihood estimation, and the prediction confidence interval is set to 95%. After the third trend function is generated, the key stages of crack propagation are identified by analyzing the local acceleration area. For example, three acceleration points with stress intensity factors of 15, 18, and 22 MPa√m are predicted from the measured data. The density clustering method groups acceleration points based on local density estimation. The cluster radius is set to 1.5 times the standard deviation, the minimum number of points is 5, and the spatial distribution of high-risk areas is identified. The degree of signal degradation is measured by the eye opening. When the decrease exceeds 30%, the corresponding stress intensity factor is close to the critical value. A damage evolution curve was constructed based on the third trend function. The curve shows three stages: microcrack initiation, stable expansion, and rapid expansion. For example, the inflection point jitter rates in a certain high-speed interface area were 2.5ps / μs, 3.8ps / μs, and 5.2ps / μs, respectively. This clearly reflects the nonlinear process of damage evolution and generates an acceleration point distribution map, providing a basis for maintenance priority sorting.

[0057] In the embodiments of the present application, through multi-level trend analysis and probability modeling, it is possible to accurately capture abnormal changes in the clock jitter offset rate and its association with microcrack propagation. The application of Gaussian process regression not only improves the prediction accuracy, but also quantifies the uncertainty through confidence intervals. For example, in areas where the crack propagation rate increases significantly, the model prediction accuracy reaches more than 90%. This method is particularly effective in complex aging scenarios. For example, the degradation of the signal path in a multi-layer PCB board due to impedance discontinuity can quickly locate the problem area through the distribution of acceleration points. The filter window size or clustering parameters can be adjusted according to actual needs to improve the adaptability of the model. The specific implementation details are optimized by technical personnel according to the scenario.

[0058] S105. Obtain statistical features of abnormal signals corresponding to the distribution of acceleration points of each aging mode in different time windows and analyze the evolution law of the abnormal signals by comparing feature changes. Then, use a long short-term memory network to perform time series prediction on the evolution law to generate an initial sequence of maintenance priorities.

[0059] In an embodiment of the present application, statistical feature extraction and time series modeling are used to analyze the dynamic evolution of abnormal signals in the aging mode, providing an intelligent basis for the maintenance sorting of industrial computers and ensuring efficient resource allocation.

[0060] S1051. Use the median filtering method to remove measurement noise from the original data sequence and perform standard deviation normalization to generate a first data sequence. Then, construct a fixed-length sliding window to calculate feature statistics to obtain a first feature sequence, and mark mutation points based on the difference values ​​between adjacent windows to generate a second feature sequence.

[0061] In this embodiment, raw data is collected from abnormal signals corresponding to the acceleration point distribution. The sampling frequency is set to 100 MHz, and parameters such as clock signal jitter and impedance variation are recorded. A five-point median filter is used to smooth the data and remove measurement noise. For example, a noise amplitude of approximately 5 ps can effectively filter out sudden interference. Data normalization is based on the standard deviation method. Outliers exceeding three standard deviations from the mean are identified and removed, generating a first data sequence. A sliding window with a fixed length of 60 seconds and a 50% overlap is constructed. Feature statistics are extracted by calculating the mean, variance, and kurtosis within the window to form a first feature sequence. The mean reflects the overall signal level, the variance indicates the amplitude of fluctuations, and the kurtosis reveals the distribution shape. For example, an increase in kurtosis from 3.2 to 4.8 in a high-speed interface region indicates that the signal distribution deviates from normality. The Euclidean distance between adjacent windows is calculated as the feature difference value. The baseline threshold is set to twice the standard deviation of the historical data. If the difference exceeds the threshold and persists for three windows, it is marked as a sudden change point, generating a second feature sequence, providing key data for subsequent time series analysis.

[0062] In some optional embodiments, the second feature sequence is segmented to extract timing features and key variables are selected using a gradient-based feature screening method to construct a first training matrix. This matrix is ​​then loaded into a long-short-term memory network (LSTM) to capture timing dependencies and predict signal variation patterns through hidden layer states, generating a first prediction sequence. The second feature sequence is segmented by mutation points, and timing features, such as signal degradation rate and cumulative change trend, are extracted for each segment. A gradient-based method is used to screen key variables, and the gradient values ​​of each feature are calculated. The top 20% of features, such as clock jitter and signal reflection coefficient, are selected for training, with each sample containing data from 10 consecutive windows. The first training matrix is ​​then input into a long-short-term memory network (LSTM) with a three-layer architecture, 64 hidden layer neurons, and a sequence length of 10. Training on historical fault data yields a prediction accuracy of 92% on the validation set. The network records timing dependencies through hidden layer states and predicts signal variation patterns. For example, it can identify power integrity degradation in a server motherboard 72 hours in advance, generating a first prediction sequence that lays the foundation for priority assessment.

[0063] S1052: extracting an aging acceleration index based on the first prediction sequence and calculating a hazard level using a fault severity quantification method, and then constructing a priority scoring function in combination with maintenance resource constraints to generate an initial priority sequence.

[0064] In an embodiment of the present application, an aging acceleration index is extracted from the first prediction sequence, and the signal degradation rate and cumulative effect are integrated. For example, after a certain PCB board has been running for 4000 hours, the jitter increases from 45ps to 75ps, and the acceleration index reaches 0.95. The severity of the fault is evaluated by a quantitative method and divided into 5 levels. When the index exceeds 0.8 and rises for three consecutive prediction cycles, it is included in the high-risk level. Combined with maintenance resource constraints, such as the difficulty of repair and the scope of fault impact, a priority scoring function is constructed with a scoring range of 0 to 1. For example, the jitter acceleration of the core processor power supply area of ​​a data center switch motherboard is 0.95, and it is predicted to reach the critical value within 48 hours, with a score of 0.92; the degradation rate of the high-speed backplane interface is 0.3 / hour, with a score of 0.85. The initial priority sequence is generated based on the score, and it is recommended to give priority to the power supply area to avoid coupling effects that exacerbate system risks.

[0065] In the embodiments of the present application, the evolution law of abnormal signals can be accurately captured through multi-dimensional feature extraction and timing prediction. The application of long short-term memory networks effectively records the long-term dependence of signal changes. For example, in the aging of DDR interfaces, the standard deviation increases from 2.8ps to 6.5ps, and the network accurately predicts the fault point. Priority scoring is combined with actual resource constraints to ensure the feasibility of maintenance strategies. For example, in multi-point aging scenarios, areas with a score exceeding 0.9 are prioritized, greatly improving system stability. The window length or network parameters can be adjusted according to the application scenario to optimize the prediction accuracy, and the specific implementation is flexibly set by technical personnel.

[0066] S106. Identify interference factors that affect system stability assessment from electromagnetic compatibility characteristic variation coupling feature data, and classify the interference factors based on the initial priority sequence.

[0067] Spectral decomposition is commonly used to analyze electromagnetic compatibility (EMC) data. Its purpose is to decompose complex signals into distinct frequency components to reveal potential interference sources. For example, in PCB EMI analysis, the raw data may contain broadband noise. Fast Fourier transform (FFT) is used to convert the time-domain signal into the frequency domain, generating a first frequency-domain sequence.

[0068] For example, electromagnetic radiation data from a high-speed circuit board was collected at a 100MHz sampling rate. Spectral decomposition revealed a significant peak in the 2GHz band, indicating possible harmonic interference. This decomposition helps locate the interference source and improves the targeting of subsequent analysis. Statistical analysis of the first frequency domain sequence using a fixed-length sliding window is performed to calculate the mean and variance, generating a second statistical sequence.

[0069] Specifically, the sliding window can be set to 1 second, with a step size of 0.5 seconds and an overlap rate of 50% to capture the dynamic changes of frequency components.

[0070] In a possible implementation, for a frequency domain sequence in the 2 GHz band, the mean within the window reflects the average interference intensity, and the variance represents the degree of fluctuation.

[0071] For example, the mean value in a certain window is -60dBm and the variance is 5dBm5, indicating that the interference intensity is relatively stable.

[0072] Preferably, variance analysis can highlight abnormal fluctuations and provide a basis for interference point marking. If the variance of the second statistical sequence exceeds a preset threshold, it is marked as a potential interference point and a third marking sequence is generated.

[0073] It will be appreciated that the threshold may be set based on two standard deviations of the historical data.

[0074] For example, when a circuit board operates normally, its variance is 3dBm5, and the threshold is set at 6dBm5. When the variance in a window reaches 8dBm5, it is marked as an interference point. This marking method effectively screens out abnormal areas, facilitating subsequent partition analysis. Time series partitioning is performed based on the third marker sequence, and gradient analysis is used to extract the changing trends of each interval to generate the fourth feature matrix.

[0075] In one embodiment, the time series can be partitioned by the time of occurrence of the interference point, with each segment containing 10 consecutive windows. Gradient analysis focuses on the rate of change of interference intensity. For example, if the interference intensity in a partition increases from -55dBm to -50dBm, with a gradient of 0.5dBm / window, this indicates increased interference. The fourth feature matrix integrates the gradient information of each partition and provides input for dimensionality reduction. Principal component analysis is used to reduce the dimensionality of the fourth feature matrix, screening key variables and generating the fifth training matrix.

[0076] For example, principal component analysis can retain 80% of the variance contribution rate and reduce high-dimensional features to 5 dimensions.

[0077] For example, among variables such as interference gradient and frequency offset for a certain circuit board, the interference gradient has the highest contribution rate and is therefore selected as the key variable. This dimensionality reduction reduces computational complexity while preserving key information. A long short-term memory network is used to perform time series modeling on the fifth training matrix to generate the sixth prediction sequence.

[0078] Specifically, the network can be configured with a two-layer structure, with 32 neurons per layer and a sequence length of 10. The training data includes historical interference patterns, such as periodic interference observed on a circuit board after 500 hours of operation. By learning the temporal dependencies of the interference, the network can predict interference trends for the next 10 windows. This modeling approach captures long-term patterns and improves prediction accuracy. The classification label for the interference factor is determined based on the degree of match between the sixth predicted sequence and the initial priority sequence.

[0079] In one possible implementation, the initial priority sequence is predefined based on the impact of interference on system performance. For example, high priority corresponds to interference with a signal attenuation exceeding 3dB. The sixth prediction sequence indicates that interference in a certain frequency band will reach -45dBm within 24 hours, matching the high priority sequence and classifying it as a high-risk interference factor. This classification helps optimize resource allocation and prioritize the handling of severe interference.

[0080] In other embodiments, by analyzing interference factors in electromagnetic compatibility characteristic variations, combined with time series features and prediction models, interference factors are classified and their priority order is dynamically adjusted, ensuring the scientific and practical nature of system stability assessment. The original signal is processed using a median filter to eliminate measurement noise and normalized to generate a normalized signal sequence. Statistics are then calculated using time series windowing to capture signal variation patterns. Potential fault areas are identified based on the characteristic difference values ​​of adjacent windows and a baseline threshold. Clock signal jitter and impedance variation parameters are collected at a basic sampling frequency. For example, during normal operation, the jitter amplitude remains within 20% of the baseline value, but may increase to over 50% after long-term aging. Median filtering uses a five-point sliding window to remove random noise, and normalization eliminates outliers based on statistical features to form a normalized signal sequence. The time series window length is dynamically adjusted based on the frequency of signal variation. For example, it is set to 60 seconds with a 50% overlap during normal operation and shortened to 20 seconds during abnormal conditions to improve response speed. The mean calculated within the window reflects performance, the variance measures stability, and the kurtosis identifies abnormal fluctuations. For example, an increase in kurtosis from 3 to 6 during aging indicates the incipient fault. The feature difference is measured by the Euclidean distance between adjacent windows. The baseline threshold is determined by twice the standard deviation of the historical data. If the difference value exceeds the threshold for five consecutive periods, such as a short-term fluctuation increase of 30% or a medium-term slope increase of 2 times, it is marked as a potential fault area, providing a basis for subsequent classification.

[0081] S1061. Extract time series features based on the signal change rules and use gradient screening to filter dominant variables to construct training samples. Then use a multi-layer network structure to predict the quality degradation trend and generate acceleration indicators and danger levels to provide data support for interference factor classification.

[0082] In the embodiment of the present application, timing features are extracted from the normalized signal sequence, focusing on the acceleration characteristics of signal changes, such as the gradual increase of jitter or sudden change of impedance. The gradient screening method calculates the gradient value of each feature, and after sorting, selects the top 20% of the dominant variables, such as signal quality parameters and reflection coefficients. Each training sample contains 10 continuous window data to preserve timing integrity. A multi-layer long short-term memory network is used to predict degradation trends. The network consists of three layers and 64 hidden layer neurons. The timing dependencies are captured through memory units. The training data covers multiple fault evolution cycles and marks key turning points. Verification shows that the accuracy of early warning for fast faults 48 hours in advance is 90%, and the accuracy of early warning for slow faults 96 hours in advance is 85%. Based on the prediction results, the acceleration index is extracted, and the five levels of danger are divided into five levels based on the degradation rate and cumulative effect. For example, the acceleration of a high-speed interface area reaches 0.95, which is classified as a high-risk level, laying the foundation for the classification of interference factors.

[0083] S1062. Classify the interference factors based on the initial priority sequence and quantify the fault priority through a scoring function, and then analyze the coupling effect between the fault points to dynamically update the priority sequence.

[0084] In the implementation of this application, the interference factors are associated with the initial sequence of priorities and classified into instantaneous interference, trend interference and cumulative interference. For example, the aging of the core power supply area belongs to cumulative interference, and the degradation of the surrounding signal lines is a trend interference. The priority scoring function integrates the degradation rate weight of 0.4, the impact range weight of 0.35 and the maintenance complexity weight of 0.25. For example, in a certain case, the scores of the three fault points are 0.82, 0.75 and 0.68 respectively to determine the maintenance order. Further analysis of the coupling effect shows that when the coupling coefficient of adjacent fault points exceeds 0.7, such as a dual fault point coupling strength of 0.85, the combined impact increases the performance degradation rate by 60%, and the priority is increased by 0.15. The priority sequence is dynamically updated to respond to changes in the fault situation to ensure that high-priority fault points are handled first.

[0085] In the embodiments of the present application, multi-level analysis and prediction are used to accurately identify interference factors and optimize classification. Clock signal jitter is a key indicator reflecting the health status of the circuit. Short-term fluctuations are random, and long-term trends are gradual. Dynamic window adjustment improves the sensitivity of feature extraction. For example, a server motherboard uses this method to identify the coupling effect of the aging of the power supply area accelerating the degradation of surrounding signal lines, and timely intervention avoids cascading failures. The high predictive ability of the model combined with the practicality of the scoring system ensures the efficient use of maintenance resources and improved system stability. The window parameters or network structure can be adjusted according to the actual scenario to enhance adaptability, and the specific implementation is optimized by technical personnel.

[0086] S107. Adjust the abnormality judgment standard threshold of aging pattern recognition according to the interference factor classification and compare the matching degree of the priority sequence and system stability before and after the adjustment to identify high-priority aging abnormality areas and generate electrical characteristic adjustment instructions, and formulate a real-time maintenance plan for the mainboard circuit.

[0087] In an embodiment of the present application, the abnormality judgment threshold is optimized by interference factor classification, and priority changes and stability assessments are combined to accurately locate high-risk areas and generate maintenance instructions to ensure the long-term reliable operation of the mainboard circuit.

[0088] S1071. Standardize the original interference data according to the interference classification level and eliminate noise through median filtering to generate a first interference feature matrix. Then calculate the aging severity index and optimize the threshold adjustment function using a recursive iterative method to obtain a first threshold adjustment matrix according to the adjustment step size.

[0089] In an embodiment of the present application, voltage, current and impedance parameters are extracted from the raw data with a sampling rate of 100MHz based on interference classification such as power supply noise, clock crosstalk and electromagnetic coupling. The data is normalized to the interval [-1,1] by standardization processing, and a 5-point median filter is used to remove burst noise. For example, when the power supply noise amplitude fluctuates within ±5% of the reference voltage, it can be effectively smoothed. The first interference feature matrix is ​​generated for the abnormal area segmentation, and the aging severity index is calculated. The signal degradation rate and cumulative effect are integrated. For example, when the degradation rate exceeds 0.3% / hour and lasts for 24 hours, it is marked as severe. The recursive iterative method optimizes the threshold adjustment function, with an initial step size of 10% of the reference threshold. Convergence is judged by a change of less than 1% for 3 consecutive iterations. The step size is dynamically adjusted to generate the first threshold adjustment matrix, providing a basis for priority optimization.

[0090] In some optional embodiments, a priority change curve is constructed based on the first threshold adjustment matrix. The stability index is used to calculate the difference between the sequence before and after the threshold adjustment to establish a matching quantization function, resulting in a second threshold adjustment matrix. Subsequently, electrical parameter variation patterns are extracted and mapped using a recursive neural network to generate a first parameter optimization sequence and formulate maintenance instructions. The first threshold adjustment matrix is ​​used to plot the priority change curve. The stability index is calculated based on signal quality improvement and system reliability. For example, after adjustment, the priority of a high-speed interface area increases from 0.75 to 0.92, resulting in a 25% improvement in stability. The matching quantization function evaluates the adjustment effect using the sequence correlation coefficient to generate a second threshold adjustment matrix. Electrical parameter variation patterns, such as the relationship between timing jitter and impedance matching, are extracted. A three-layer recursive neural network with 32 hidden units is used to map features. The input includes key parameters and the output is a parameter correction vector. The weights of each component are determined by sensitivity analysis. Based on the first parameter optimization sequence, electrical characteristic corrections, such as clock buffer voltage adjustment and terminal impedance optimization, are encoded using an instruction template library. A second parameter optimization sequence is generated and prioritized. Dependencies are described using a directed acyclic graph to ensure a reasonable execution order.

[0091] S1072. Extract coordinate information for the abnormal area and classify and mark the abnormal type and severity through the area positioning algorithm to generate a first abnormal area table, then construct a circuit parameter adjustment matrix and verify the adjustment plan, generate a first maintenance instruction set and optimize the maintenance schedule.

[0092] In this embodiment, the coordinates of abnormal regions are extracted from a first interference signature matrix. The circuit board is divided into a 10×10 grid. Anomaly types, such as power supply interference or signal integrity degradation, are marked using a region location algorithm. Severities exceeding 0.8 are classified as high risk. For example, a certain interface region's index reaches 0.95. A circuit parameter adjustment matrix is ​​constructed based on the first abnormal region table. The voltage adjustment range is limited to ±10%, the current variation does not exceed 15%, and the impedance deviation is controlled within 10%, with a step size of 1% of the nominal value. A support vector regression model is used to validate the adjustment plan. The radial basis kernel function parameters are optimized through cross-validation and validated when the comprehensive score exceeds 0.9. A first maintenance instruction set is generated, containing target parameters and adjustment steps. Conflict detection ensures correct sequencing. Maintenance time windows are divided based on the anomaly rate: areas with a severity of 0.9 or higher are maintained every 24 hours, and those between 0.7 and 0.9 are maintained every 72 hours. Dynamic programming optimizes resource allocation. For example, in a data center, tasks for coupled regions can be merged to reduce maintenance time from 6 hours to 4 hours. Finally, a dynamic maintenance scheduling plan is generated.

[0093] In the embodiments of this application, adaptive threshold adjustment and parameter optimization significantly improve anomaly identification accuracy and maintenance efficiency. For example, a server motherboard power adjustment instruction has a priority of 0.95, which is executed first to avoid affecting signal quality and improve stable operation time by 40%. Coupling effect analysis merges related instructions, reducing repeated operations and shortening the average repair time by 30%, providing an efficient solution for multi-point aging scenarios. The iteration step size or network structure can be adjusted according to actual needs, and the specific implementation is flexibly set by technical personnel.

[0094] The above is only a preferred implementation of the embodiment of the present application. It should be pointed out that ordinary technicians in this technical field can make several improvements and supplements without departing from the principles of the embodiment of the present application. These improvements and supplements should also be regarded as the scope of protection of the embodiment of the present application.

Claims

1. An industrial computer predictive maintenance method based on artificial intelligence, characterized in that: The method comprises: Acquiring operating data of the mainboard circuit, the operating data including at least real-time monitoring values ​​of clock signal jitter, signal integrity parameters, and electromagnetic compatibility characteristics; Analyzing the operating data to obtain at least one signal abnormality group in an aging mode; Determining a refined expression of the abnormal signal of each of the abnormal signal groups; Identifying, based on refined expressions of abnormal signals of each of the signal abnormality groups, an acceleration point distribution of a first signal abnormality group; the first signal abnormality group is any one of the at least one signal abnormality group, and a clock jitter offset rate of the abnormal signal of the refined expression of the first signal abnormality group exceeds a preset offset threshold; According to the statistical characteristics of the abnormal signals corresponding to the distribution of acceleration points of each aging mode in different time windows, the evolution law of the abnormal signals in different time windows is obtained; Long short-term memory network is used to predict the evolution law in time series and determine the initial priority sequence of maintenance sorting. According to the priority initial sequence of the maintenance sorting, high-priority aging abnormality areas are obtained, electrical characteristic adjustment instructions are generated, and a real-time maintenance plan for the mainboard circuit is determined.

2. The method according to claim 1, characterized in that The method of obtaining a high-priority aging abnormality area based on the maintenance sorting priority initial sequence, generating an electrical characteristic adjustment instruction, and determining a real-time maintenance plan for the mainboard circuit includes: Identifying interference factors that affect system stability assessment from the coupling characteristics, and classifying the interference factors based on the initial priority sequence; According to the classification results of the interference factors, the abnormality judgment threshold of the aging pattern recognition is adjusted; Compare the matching degree between the change in priority sequence before and after adjustment and the improvement in system stability to identify high-priority aging abnormal areas; An electrical characteristic adjustment instruction is generated according to the high-priority aging abnormality area, and a real-time maintenance plan for the mainboard circuit is determined.

3. The method according to claim 1, characterized in that The determining of a refined expression of the abnormal signal of each of the abnormal signal groups includes: Extracting coupling features with electromagnetic compatibility characteristic variations of the corresponding signal abnormality group from the signal integrity parameters of each signal abnormality group; the features of any of the abnormal signal groups are obtained by weighted analysis of the interaction between clock signal jitter and microcrack propagation within the signal abnormality group; The dimension of the coupling feature data of any of the abnormal signal groups is compressed in combination with the features of any of the abnormal signal groups to obtain a refined expression of the abnormal signal of each of the abnormal signal groups.

4. The method according to claim 1, wherein The analyzing the operating data to obtain at least one signal abnormality group in the aging mode includes: Obtain real-time monitoring values ​​of clock signal jitter, signal integrity parameters, and electromagnetic compatibility characteristics from the motherboard circuit operation data to form multi-dimensional time series data; Performing feature extraction on the multidimensional time series data to obtain a preliminary distribution of abnormal signals related to aging; A preliminary distribution cluster analysis of abnormal signals is performed to obtain signal anomaly groups under the aging mode. A preset convolution kernel is used to perform a weighted analysis on the interactive influence of clock signal jitter and microcrack extension within the group to obtain the abnormal signal grouping characteristics.

5. The method according to claim 4, characterized in that The feature extraction of the multi-dimensional time series data to obtain a preliminary distribution of abnormal signals related to aging includes: Jitter data is collected based on the sampling rate of the clock signal during operation of the mainboard circuit, and a first time series data set is obtained by using the signal integrity transmission rate and the electromagnetic field intensity threshold; Normalizing the voltage fluctuation amplitude and signal attenuation degree in the first time series data set using a maximum and minimum value normalization method to obtain a second time series data set; Performing time series clustering on the second time series data set using a radial basis kernel function in a support vector machine clustering algorithm to obtain a first abnormal feature cluster; An aging feature vector is constructed for the first abnormal feature cluster, and the aging feature vector is classified using a multi-layer perceptron structure in a convolutional neural network. An aging correlation matrix is ​​calculated based on the signal integrity transmission rate to obtain a circuit aging distribution map.

6. The method according to claim 4, characterized in that The preliminary distribution cluster analysis of abnormal signals is performed to obtain signal abnormality groups under the aging mode. A preset convolution kernel is used to perform a weighted analysis on the interactive effects of clock signal jitter and microcrack extension within the group to obtain abnormal signal group features, including: Constructing an original characteristic matrix based on the clock jitter amplitude and the microcrack length, and normalizing the original characteristic matrix using a maximum and minimum value normalization method to obtain a first normalized characteristic matrix; Extracting time series statistical features through a sliding time window according to the first standardized feature matrix, and grouping the time series statistical features using a hierarchical clustering method to obtain a first feature grouping set; For the first feature grouping set, a radial basis function in a support vector machine is used to perform mapping transformation on the crack stress distribution data, and a second feature grouping set is obtained according to the signal clustering distance; The Gaussian convolution kernel weight matrix is ​​loaded into the second feature group set, a multi-scale convolution transform is performed on the clock signal jitter data, and weighted superposition is performed on the crack growth rate data to obtain abnormal signal grouping features.

7. The method according to claim 3, characterized in that Extracting coupling features corresponding to electromagnetic compatibility characteristic variations of the signal abnormality group from the signal integrity parameters of each signal abnormality group, and compressing the dimension of the coupling feature data of any signal abnormality group in combination with the features of any signal abnormality group to obtain a refined expression of the abnormal signal of each signal abnormality group includes: constructing an original feature matrix according to the signal integrity and the electromagnetic interference intensity, and normalizing the original feature matrix using a maximum and minimum value normalization method to obtain a first standardized feature matrix; Extracting compression features based on the waveform distortion degree and parameter variation amplitude in the first standardized feature matrix through a coding layer based on a multi-layer autoencoder to obtain a first compressed feature matrix; Calculating coupling metrics based on the signal transmission stability and anti-interference capability in the first compressed feature matrix, and grouping the coupling metrics using a density-based adaptive clustering method to obtain a second compressed feature matrix; constructing a covariance representation of electromagnetic radiation intensity for the second compressed feature matrix, extracting main feature directions using an eigenvalue decomposition method, and obtaining a first principal component matrix; The information gain of each feature dimension is calculated according to the first principal component matrix, the key dimensions are extracted by a feature selection method based on mutual information, a coupling feature vector is constructed for the abnormal frequency distribution, and a refined expression of the abnormal signal of each of the signal abnormality groups is obtained.

8. The method according to claim 1, characterized in that The identifying the acceleration point distribution of the first signal abnormality group according to the refined expression of the abnormal signals of each of the signal abnormality groups includes: Eliminating measurement noise from the clock jitter data by a median filtering method, and performing synchronization processing on the jitter data by timestamp alignment to obtain a first jitter sequence; calculating offset rates of adjacent sampling points for the first jitter sequence, normalizing the offset rates according to a preset threshold range, marking abnormal points using a three-times standard deviation principle, and obtaining a first rate sequence; A mapping relationship between crack length and jitter rate is established using a radial basis kernel function according to the first rate sequence, and kernel function parameters are determined through cross-validation to obtain a first trend function; The local rate of change is calculated for the first trend function, and the trend slope characteristics are extracted through a sliding differential window. The Gaussian process regression model is loaded according to the stress intensity factor. The mean function and covariance function are used to construct a nonlinear mapping, and the local acceleration area is extracted. The acceleration points are spatially grouped using the density clustering method. The damage evolution fluctuations are analyzed according to the degree of signal degradation, and the distribution of acceleration points is identified.

9. The method according to claim 1, characterized in that According to the statistical characteristics of the abnormal signals corresponding to the distribution of acceleration points of each aging mode in different time windows, the evolution law of the abnormal signals in different time windows is obtained; Long short-term memory network is used to predict the evolution law in time series and determine the initial priority sequence of maintenance sorting, including: The original data sequence is processed by a median filter method, and a first data sequence is obtained by a standard deviation normalization operation; Constructing a fixed-length sliding window according to the first data sequence, and calculating feature statistics through the sliding window to obtain a first feature sequence, wherein the feature statistics include a mean, a variance, and a kurtosis value; Calculating the difference between adjacent windows for the first feature sequence, marking the mutation points according to the comparison result of the difference between the adjacent windows and the reference threshold to obtain a second feature sequence; Extracting time series features in segments according to the second feature sequence, selecting key variables using a gradient-based feature screening method to obtain a first training matrix, wherein the key variables are determined by a time series change trend; Loading the long short-term memory network structure according to the first training matrix, recording the temporal dependency through the hidden layer state vector, establishing a prediction model, and obtaining a first prediction sequence; Aging acceleration indicators are extracted for the first prediction sequence, the hazard level is calculated through a quantitative evaluation method of fault severity, and an initial priority sequence is constructed according to maintenance resource constraints.

10. The method according to any one of claims 2 to 9, characterized in that: According to the classification results of the interference factors, the abnormality judgment standard threshold of the aging pattern recognition is adjusted; the matching degree of the priority sequence change before and after the adjustment and the system stability improvement is compared to identify the high-priority aging abnormality area; Generating an electrical characteristic adjustment instruction according to the high-priority aging abnormality area and determining a real-time maintenance plan for the mainboard circuit includes: Obtain original interference data according to the interference classification level, perform noise elimination on the original interference data through the median filtering method, and obtain the interference feature matrix; Calculating an aging severity index for the interference feature matrix, optimizing parameters of a threshold adjustment function using a recursive iterative method, determining an adjustment step size according to the aging severity index, and obtaining a first threshold adjustment matrix; constructing a priority change curve based on the first threshold adjustment matrix, calculating the sequence difference before and after the threshold adjustment using a stability index, and establishing a matching quantization function for the sequence difference to obtain a second threshold adjustment matrix; The electrical parameter change pattern is extracted from the second threshold adjustment matrix, and a recursive neural network is used to perform feature mapping on the electrical parameter change pattern. Electrical characteristic adjustment instructions are generated based on the feature mapping results, and priorities are assigned to the instructions. A scheduling sequence is generated based on the dependency relationship between the instructions to determine the real-time maintenance plan for the mainboard circuit.

Images