AI-based server hardware failure prediction system

Abstract

This invention discloses an AI-based server hardware fault prediction system, relating to the fields of artificial intelligence and server operation and maintenance. It constructs a multi-dimensional energy spectrum feature matrix containing frequency domain energy density distribution, generates a degradation spatiotemporal coordinate system and node topology graph, obtains a set of correlation locking probabilities by calculating the energy gradient between nodes, performs folding fitting on the feature matrix to construct a micro-degradation field mapping function, generates a Lévy morphological degradation evolution curve, further calculates the fault energy convergence rate, and combines the correlation locking probabilities to form a fault risk heatmap. By extracting features from abnormal gradient regions in the risk heatmap, a micro-degradation early warning sequence is obtained, fault prediction coordinates are generated, and a comprehensive risk score is calculated to output the server hardware risk level. This invention enables early identification of hardware micro-degradation, energy accumulation trends, and abrupt changes, improving the accuracy and foresight of fault prediction and ensuring the stable operation of data centers.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and server operation and maintenance, specifically to an AI-based server hardware fault prediction system. Background Technology

[0002] Traditional server hardware fault prediction relies heavily on single, discrete metrics such as temperature, fan speed, and voltage. This fails to characterize the cross-temporal correlations of subtle perturbations like CPU leakage current drift, SSD latency spikes, and power supply ripple phase noise. Especially in high-density clusters, the weak degradation behaviors of different hardware models exhibit non-linear diffusion, making early warning difficult. Furthermore, existing models cannot integrate multi-source weak noise and identify abrupt changes in degradation paths, resulting in latent faults under thermal shock and high-concurrency loads remaining undetected. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based server hardware fault prediction system to address the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI-based server hardware fault prediction system, comprising:

[0005] Feature construction module: Collects multi-source operating signals generated by the server during operation, and calculates a multi-dimensional energy spectrum feature matrix containing frequency domain energy density distribution through a segmented wavelet convolution fusion algorithm;

[0006] Topology generation module: Based on the multidimensional energy spectrum feature matrix, a degenerate spatiotemporal coordinate system is constructed to form a node topology graph, and the energy gradient between nodes is calculated to obtain the association locking probability set;

[0007] Evolution curve generation module: Based on the associated locking probability set, the feature matrix is ​​folded and fitted to construct a micro-degeneration field mapping function and generate the Lévy morphological degradation evolution curve;

[0008] Risk map generation module: Calculates the fault energy accumulation rate of hardware components based on the Lévy morphological degradation evolution curve, and couples and corrects it with the associated lock-in probability to generate a fault risk heat map;

[0009] Early warning sequence generation module: Extracts features from abnormal gradient regions in the fault risk heatmap to form a micro-degradation early warning sequence;

[0010] Risk level output module: Generates fault prediction coordinates based on micro-degradation early warning sequence and outputs the server hardware risk level.

[0011] Preferably, the step of calculating the multidimensional energy spectrum feature matrix containing the frequency domain energy density distribution using the piecewise wavelet convolution fusion algorithm includes:

[0012] Time synchronization processing is performed on the CPU leakage current drift sequence, fan micro-disturbance speed sequence, and power supply ripple phase noise sequence continuously recorded during server operation;

[0013] The synchronized operating signals are segmented according to a preset window, and wavelet basis function convolution is performed on each data segment to obtain the frequency domain energy density distribution sequence of each signal at different scales.

[0014] Based on the frequency domain energy density distribution sequence, cross-dimensional feature alignment is performed on different signals within the same time period, and a unified energy spectrum fragment is generated using a multi-layer convolutional fusion network.

[0015] All energy spectrum fragments are stacked and matrixed in chronological order to form the multidimensional energy spectrum feature matrix.

[0016] Preferably, the step of constructing a degenerate spatiotemporal coordinate system based on the multidimensional energy spectrum feature matrix to form a node topology graph includes:

[0017] The energy density vectors of each time period in the multidimensional energy spectrum feature matrix are extracted in chronological order, and the continuous time length is used as the vertical axis and the energy density change amplitude is used as the horizontal axis to construct a basic framework of degradation spatiotemporal coordinates to reflect the hardware degradation trend.

[0018] Based on the aforementioned degenerate spatiotemporal coordinate framework, the rate of change between energy density vectors in adjacent time periods is calculated, and a degenerate gradient sequence is obtained through continuous difference. Each degenerate gradient is used as the attribute value of a potential node.

[0019] Based on the relationship between the attribute values ​​of the nodes, nodes with adjacent attribute values ​​and continuous rate of change are connected by lines, and the degree of gradient proximity between nodes is used as the connection weight to form a primary node topology.

[0020] A layer-by-layer weighted filtering operation is performed on the primary node topology to remove weakly connected lines with weights below a preset threshold, resulting in a node topology that truly reflects the energy density evolution path.

[0021] Preferably, the step of calculating the energy gradient between nodes to obtain the set of associated locking probabilities includes:

[0022] The energy density attribute values ​​of any adjacent nodes in the node topology graph are interpolated, and the initial value of the energy gradient is obtained by dividing the difference by the time interval between nodes.

[0023] All node pairs are sorted according to the initial value of the energy gradient, and node pairs with continuous gradient changes and consistent directions are merged to form an energy gradient sequence.

[0024] Based on the energy gradient sequence, an interval partitioning model is established according to the distribution range of gradient values ​​in the sequence, and the correlation strength value between nodes is calculated by dividing the number of occurrences of each gradient interval by the total length of the gradient sequence.

[0025] The association strength value is normalized and coupled with the connection weights in the topology. The normalized coupling result is used as the association locking probability, and the probability values ​​of all node pairs are used to form the association locking probability set.

[0026] Preferably, the step of generating the Lévy morphological degradation evolution curve includes:

[0027] The feature vectors of each time period in the multidimensional energy spectrum feature matrix are sorted according to the probability magnitude in the associated locking probability set, and then folded and superimposed in a way that the probability is from high to low.

[0028] The folded feature structure is input into the microdegenerate field mapping function, and the nonlinear coupling strength between feature vectors is calculated in a progressive manner, and the coupling strength is used as the estimate of the local change rate of the microdegenerate field.

[0029] Based on the local rate of change sequence, Lévy distribution characteristic parameters are constructed according to the joint conditions of jump probability and mutation amplitude, and the characteristic parameters are used to drive the time series to generate Lévy morphological degradation and evolution curves.

[0030] Preferably, the step of constructing Lévy distribution characteristic parameters based on the local rate of change sequence according to the joint condition of jump probability and abrupt change magnitude includes:

[0031] The local rate of change sequence output by the microdegenerate field mapping function is filtered by amplitude. Time points where the absolute value of the rate of change exceeds a preset mutation threshold are marked as jump events, and the jump probability is calculated based on the proportion of jump events.

[0032] Statistical analysis is performed on the absolute values ​​of the rate of change of all jump events, and the range of abrupt change amplitude is constructed using the maximum, minimum and mean values, so that each jump event can find a corresponding amplitude level in the range.

[0033] The jump probability and the occurrence ratio of each amplitude level are jointly merged to generate the stability parameter, skewness parameter and scaling parameter of the Lévy distribution in a joint probability manner.

[0034] The Lévy distribution characteristic parameters are input into the Lévy morphological degradation model to drive the time series and generate Lévy morphological degradation evolution curves that conform to the hardware degradation jump characteristics.

[0035] Preferably, the step of calculating the fault energy accumulation rate of the hardware component based on the Lévy morphological degradation evolution curve includes:

[0036] The degradation amplitude of Lévy morphology degradation evolution curve at each time point is extracted, and the absolute value of the degradation amplitude is accumulated in chronological order to form a primary energy accumulation sequence to characterize the energy accumulation trend.

[0037] The primary energy accumulation sequence is smoothed, and the average growth rate within the window is calculated using a sliding window method, so that the growth rate can reflect the intensity of energy accumulation in the local time range.

[0038] Based on the changing trend of the growth rate, an energy convergence gradient sequence is constructed, and the length and amplitude of the continuous rising interval in the gradient sequence are used as the estimation factors of the energy convergence rate.

[0039] The estimated factors are combined in a weighted manner to calculate the fault energy accumulation rate.

[0040] Preferably, the step of generating the fault risk heatmap includes:

[0041] The fault energy convergence rate was normalized according to time sequence.

[0042] The normalized fault energy accumulation rate is matched with the correlation lock probability of the corresponding time period, and the joint impact value of the two is calculated by multiplicative coupling.

[0043] All joint impact values ​​are mapped in two dimensions according to the physical location of the hardware components, and a fault risk heatmap is formed by assigning color gradients to different locations.

[0044] Preferably, the step of extracting features from the abnormal gradient regions of the fault risk heatmap to form a micro-degradation early warning sequence includes:

[0045] Gradient calculation is performed at each location in the fault risk heatmap. The initial value of the spatial gradient is obtained by dividing the difference of the risk values ​​of adjacent pixels by the spatial distance. Regions whose initial gradient values ​​exceed the preset gradient threshold are marked as abnormal gradient regions.

[0046] For each abnormal gradient region, a connected component analysis is performed. The boundary range of continuous regions is identified by a point-by-point search method, and the maximum gradient value, average gradient value and area within the region are extracted as feature factors characterizing the degree of local degradation.

[0047] The feature factors are arranged in chronological order, and a micro-degradation intensity sequence is constructed using a window accumulation method;

[0048] Based on the continuous change pattern of the microdegradation intensity sequence, the intensity increase interval is extracted as the microdegradation early warning sequence.

[0049] Preferably, the steps of generating fault prediction coordinates and outputting the server hardware risk level based on the micro-degradation early warning sequence include:

[0050] Each warning value in the micro-degradation warning sequence is extracted in chronological order, and the absolute magnitude of the warning value is used as the vertical axis and the corresponding time position is used as the horizontal axis to construct a basic dataset for fault prediction coordinates.

[0051] Coordinate fitting is performed on the fault prediction coordinate dataset, and degradation path curves are generated by comparing the changes in longitudinal coordinates over time periods.

[0052] A set of risk level evaluation factors is constructed based on the maximum value, average value, and frequency of slope change in the degradation path curve;

[0053] The risk level assessment factors are aggregated in a weighted manner to calculate a comprehensive risk score, and the server hardware risk level is divided into low risk, medium risk and high risk levels according to the score range.

[0054] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0055] 1. This invention achieves deep modeling of server hardware degradation processes across multiple scales, time periods, and feature dimensions by introducing a multi-dimensional energy spectrum feature matrix, a set of correlation locking probabilities, and Lévy morphological degradation evolution curves. Compared to traditional methods that rely solely on a single physical indicator or fixed threshold monitoring, this invention can capture the cumulative effects of weak degradation perturbations, abrupt jump behaviors, and energy accumulation trends. Furthermore, through feature folding fitting and nonlinear coupling calculations, it obtains a local rate of change sequence reflecting micro-degradation dynamics, enabling early latent anomalies in the hardware degradation process to be identified before explicit failures occur, significantly improving prediction forwarding capability and stability.

[0056] 2. This invention further realizes a full-link prediction framework from physical degradation signal → energy behavior model → spatial risk expression → temporal evolution trend → graded risk output by constructing fault energy convergence rate, fault risk heat map, abnormal gradient region characteristic factors, micro-degradation early warning sequence, and fault prediction coordinates based on weighted evaluation factors. This framework can not only accurately locate risk areas and quantify degradation speed and fluctuation patterns, but also output clear risk levels that can be used for scheduling, load reduction, or maintenance decisions. This transforms server hardware fault prediction from passive response to proactive prediction, effectively improving the reliability and maintainability of data center operations. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0058] Figure 1 This is a flowchart of the system modules of the present invention.

[0059] Figure 2 This is a flowchart of the method for generating the Lévy morphological degradation evolution curve according to the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] For examples, please refer to Figure 1 , 2 As shown, the AI-based server hardware fault prediction system described in this embodiment includes:

[0062] Feature construction module: Collects multi-source operating signals generated by the server during operation, and calculates a multi-dimensional energy spectrum feature matrix containing the frequency domain energy density distribution through a segmented wavelet convolution fusion algorithm.

[0063] In this embodiment, the server generates multiple operating signals from different sources in real time during task execution. Since different hardware components have different sampling frequencies—for example, the sampling frequency of the CPU leakage current drift sequence may be 100 times per second, the sampling frequency of the fan micro-disturbance speed sequence may be 50 times per second, and the sampling frequency of the power supply ripple phase noise sequence may be 200 times per second—these operating signals do not have a natural correspondence on the time axis.

[0064] To ensure that subsequent multi-source signals can be processed together, time synchronization processing needs to be performed on the above three sequences. Time synchronization methods include:

[0065] Using the highest sampling frequency as the target time resolution, other sequences are extended to a unified time scale using linear interpolation.

[0066] The interpolated sequences are rearranged according to a unified timestamp to form a set of synchronization signals with a consistent length and a consistent time step.

[0067] For missing values ​​caused by sampling, a weighted average of nearby time points is used to fill in the missing values, so that all sequences exist continuously on a unified time axis.

[0068] The three types of signals, after synchronization processing, can achieve time-point alignment, providing an input basis for subsequent segmented processing.

[0069] The synchronized signal sequence is divided into several data segments of fixed length. The size of the preset window is determined by the continuous time length, for example, setting each data segment to 2 seconds to ensure that each data segment contains a continuous CPU leakage current drift sequence, fan micro-disturbance speed sequence, and power supply ripple phase noise sequence.

[0070] Each data segment needs to be frequency-domain expanded using wavelet basis function convolution. The wavelet basis function used is the Daubechies wavelet, which has compact support properties. It is constructed by iteratively generating a scaling function to reflect the high-frequency micro-perturbation characteristics of the signal. The convolution operation for each data segment includes the following steps:

[0071] The wavelet basis functions are stretched or compressed at different scales to give them different frequency domain resolution capabilities;

[0072] The stretched wavelet basis function is multiplied point by point with each segment of the synchronization signal and then summed to obtain the corresponding energy response value for each scale;

[0073] Arrange the energy responses at all scales in chronological order to obtain the frequency domain energy density distribution sequences of the CPU leakage current drift sequence, fan micro-perturbation speed sequence, and power supply ripple phase noise sequence at different scales.

[0074] These energy density distributions can reflect the characteristics of weakly degenerate perturbations at different scales.

[0075] The dimensionality of the frequency domain energy density distribution sequence varies depending on the hardware type. To achieve cross-signal fusion, cross-dimensional feature alignment is required. Feature alignment methods include:

[0076] Statistical analysis of the energy density range of various signals across all scales;

[0077] By using a linear interval mapping method, the three types of energy densities are mapped to a unified interval, enabling them to enter the same feature space;

[0078] The mapped signals are aligned in chronological order so that the energy densities of the central processing unit, fan, and power supply can be matched in the same time period.

[0079] Next, the aligned three types of features are input into a multi-layer convolutional fusion network. This multi-layer convolutional fusion network is constructed in the following manner:

[0080] The first convolutional layer is responsible for extracting local coupling features between different signals;

[0081] The second convolutional layer is responsible for extracting the trend of signal change in the time direction;

[0082] The third convolutional layer extracts cross-scale correlation features by expanding the convolutional receptive field;

[0083] Finally, the outputs of all convolutional layers are merged in chronological order to obtain a unified energy spectrum fragment. This unified energy spectrum fragment can fully describe the energy evolution structure of the hardware over the same time period. The unified energy spectrum fragments are stacked sequentially according to the actual acquisition order. The stacking steps include:

[0084] Arrange adjacent segments in chronological order;

[0085] Fill each segment into the corresponding position in the matrix according to a fixed row and column structure;

[0086] Smoothing is applied to the time boundaries between segments to ensure that there are no abrupt changes in the matrix at the segment boundaries.

[0087] The resulting multidimensional energy spectrum feature matrix can be used to represent the energy evolution structure of all hardware components during continuous operation, providing basic data for modeling subsequent degradation behavior.

[0088] Topology generation module: Based on the multidimensional energy spectrum feature matrix, a degenerate spatiotemporal coordinate system is constructed to form a node topology graph, and the energy gradient between nodes is calculated to obtain the set of association locking probabilities.

[0089] The multidimensional energy spectrum feature matrix consists of a sequence of energy density vectors over a continuous time period. Each energy density vector reflects the energy change characteristics of the server hardware within the corresponding time period. To describe the hardware degradation trend, the energy density vectors arranged in time within the matrix need to be extracted one by one, and the vertical coordinates need to be determined according to their time positions, so that the time length increases continuously with a fixed step size.

[0090] To characterize the direction of energy evolution, the magnitude of change of each energy density vector is used as the horizontal axis. The magnitude of change is described by comparing the differences in energy density values ​​between adjacent time periods, using absolute differences to depict the degree of energy fluctuation. A coordinate framework constructed based on the vertical and horizontal axes is used to reflect the path of hardware energy change over time, thus forming the basic framework of the degradation spatiotemporal coordinate system. This framework provides a spatial representation benchmark for subsequent gradient analysis and node construction.

[0091] To reveal the rate of change in hardware degradation trends, it is necessary to calculate the rate of change of energy density vectors between adjacent time periods. The rate of change is determined by subtracting the energy density vector of the previous time period from the current energy density vector, and then dividing the difference by the fixed time interval between the two time periods. This calculation result represents the amount of energy change per unit time.

[0092] The above calculations are performed on the entire time series of the energy spectrum feature matrix using a continuous difference method, so that all adjacent time points generate corresponding rates of change, thus forming a degradation gradient sequence. Each gradient value in the degradation gradient sequence indicates the rate of change of the hardware energy state during that time period. Each gradient value is used as an attribute value of a potential node, maintaining the correspondence between attribute values ​​and time positions, providing basic data for constructing node connections.

[0093] Among all potential nodes, those with adjacent attribute values ​​and consistent trends represent continuous changes in hardware energy. To identify this continuity, attribute values ​​are arranged chronologically, and the differences in attribute values ​​between adjacent nodes are compared. When the difference is within a preset range, the two nodes are connected by a line, so that the line represents the continuity of energy changes.

[0094] The weights of connections between nodes are determined by gradient proximity. Gradient proximity is the reciprocal of the difference in attribute values ​​between two adjacent nodes; the smaller the difference, the higher the proximity and the greater the weight; conversely, the greater the difference, the lower the proximity and the smaller the weight. By assigning weights to each connection, a primary node topology that reflects the continuity of the degradation trend can be formed. This structure includes the directionality and association strength between nodes.

[0095] To ensure that the node topology accurately reflects the hardware energy evolution path, weakly connected links need to be removed from the primary node topology. Weakly connected links are defined as links with weights below a preset threshold. This preset threshold is determined through statistical analysis of all link weights. Specifically, the average weight of all links is calculated, and this average is multiplied by a set weight ratio coefficient to obtain the threshold. The ratio coefficient is between 0.5 and 0.8 to ensure that the retained links all have practical significance.

[0096] The screening method includes: checking all connections layer by layer in descending order of weight, and deleting all connections with weights below a threshold. The remaining connections after screening constitute a node topology graph that truly represents the energy density evolution path, ensuring that the topology structure retains only node associations with practical degenerative significance. The resulting topology graph will serve as the input for subsequent association locking probability calculations.

[0097] In this embodiment, each node in the node topology graph has an energy density attribute value and corresponds to a specific time position. To obtain the rate of energy change between nodes, it is necessary to perform interpolation processing on the energy density attribute values ​​of any adjacent nodes. The interpolation processing method is as follows: subtract the energy density attribute value of the preceding node from the energy density attribute value of the following node, and divide the difference by the time interval between the corresponding time positions of the two nodes, so that the result represents the amount of energy density change per unit time. This result is defined as the initial value of the energy gradient. The initial value of the energy gradient is used to reflect the trend of energy evolution between adjacent nodes, providing a basic metric for the subsequent gradient sequence construction.

[0098] To identify the continuous trend of hardware energy changes, the initial energy gradient values ​​of all node pairs need to be sorted according to their magnitude. After sorting, node pairs with consistent gradient directions are merged. Consistent direction means that the initial energy gradient values ​​all show an increasing trend or all show a decreasing trend. The method for merging continuous sequences is to combine node pairs with the same gradient direction after sorting and whose numerical changes are within a set range, in chronological order, so that the combined node pairs form an energy gradient sequence that represents the stable direction of energy evolution. This energy gradient sequence is used to depict the continuity of the degradation trend.

[0099] To quantify the correlation between nodes, an interval partitioning model needs to be established. The model is constructed as follows: The maximum and minimum values ​​of all gradient values ​​in the energy gradient sequence are statistically analyzed, and this range is divided into at least five equally spaced gradient intervals, each representing a different level of energy evolution intensity. Each gradient value falls into its corresponding interval based on its magnitude. Subsequently, the correlation strength value corresponding to each interval is obtained by recording the frequency of occurrence of each interval and dividing this frequency by the total length of the energy gradient sequence. A higher correlation strength value indicates greater consistency in energy evolution among nodes. This calculation method ensures that the correlation strength is quantifiable and verifiable.

[0100] The final degree of association between nodes depends not only on the association strength value but also on the node's connection weight in the topology. To obtain a probability value that reflects both influences, a normalized coupling operation is required. The normalized coupling method involves normalizing both the association strength value and the connection weight to their respective maximum values, ensuring both are within the range of 0 to 1. Then, the two normalized results are combined in a weighted manner, with the weighting rule: association strength accounts for 0.6 and connection weight accounts for 0.4, ensuring the coupling result reflects both energy evolution trends and topological stability. The weighted result is the association-lock probability. Summarizing the association-lock probabilities of all node pairs according to node correspondence constitutes the association-lock probability set. This set will be used in the subsequent construction of the degenerate mapping function.

[0101] Evolution curve generation module: Based on the associated locking probability set, the feature matrix is ​​folded and fitted to construct a micro-degeneration field mapping function and generate the Lévy morphological degradation evolution curve.

[0102] Each time period in the multidimensional energy spectrum feature matrix contains a corresponding feature vector. To strengthen features with high correlation to degradation, each probability value in the correlation locking probability set needs to be paired with its corresponding feature vector and arranged in descending order of probability. After sorting, the sorted feature vectors are then superimposed sequentially. The superposition method is to superimpose corresponding elements of different feature vectors, so that high-probability features contribute more to the folded structure. This folded superposition structure serves as the input basis for the micro-degradation field mapping function, making subsequent calculations more prominent on the core degradation features.

[0103] The folded feature structure is input into a micro-degradation field mapping function. This mapping function consists of multiple nonlinear transformations, each layer performing a nonlinear combination of the input features. The calculation method involves calculating the degree of difference between feature vectors layer by layer and generating the corresponding nonlinear coupling strength through weighted superposition. A greater coupling strength indicates a more significant degree of degradation impact. By recording the coupling strength at each time interval, a local rate of change sequence of the micro-degradation field can be obtained, reflecting the rate of change at different time intervals during the hardware degradation process.

[0104] Local rate-of-change sequences reflect the speed of hardware degradation. To construct Lévy distribution characteristic parameters that can describe abrupt changes, jump behavior analysis needs to be performed on the rate-of-change sequences. Jump behavior analysis includes determining jump probabilities and constructing abrupt change amplitudes. Jump probability represents the proportion of anomalous fluctuations in the rate of change to the total sequence length, while abrupt change amplitude represents the strength distribution of these anomalous fluctuations. Combining jump probability and abrupt change amplitude distributions, stability, skewness, and scaling parameters of the Lévy distribution can be generated, enabling these parameters to accurately describe the abrupt change characteristics of the rate of change. Inputting these characteristic parameters into the degradation sequence generation process can drive the time series to form a Lévy morphological degradation evolution curve, thereby reflecting the random jump characteristics of hardware degradation.

[0105] Each rate of change value in the local rate of change sequence is used to determine whether a jump event has occurred. The mutation threshold is determined by sorting all rate of change values ​​and selecting the top 10% as the mutation threshold, ensuring that this threshold represents high-amplitude anomalies in the rate of change. Any time point where the absolute value of the rate of change exceeds this threshold is marked as a jump event. The jump probability is obtained by dividing the number of jump events by the total sequence length, allowing the jump probability to quantify the frequency of mutations during hardware degradation.

[0106] To characterize the distribution of jump amplitudes, it is necessary to statistically analyze the absolute values ​​of the rates of change for all jump events. By recording the maximum, minimum, and mean values, this range can be divided into several intervals of abrupt change amplitudes. The interval division method is as follows: divide the range from the minimum to the maximum value into at least three intervals with a fixed width, ensuring that each jump event can find a matching level within each interval. The abrupt change amplitude levels are used to characterize the intensity distribution of jump events, providing a classification basis for subsequent joint probability calculations.

[0107] To ensure the Lévy distribution accurately describes degradation behavior, the statistical results of jump probabilities and mutation magnitude levels need to be jointly merged. The joint merging method involves using jump probabilities as macro-level weights and the proportion of occurrence at each mutation magnitude level as local weights, generating Lévy distribution parameters through a weighted combination. The stability parameter is dominated by jump probabilities, the skewness parameter is determined by the asymmetry of the magnitude levels, and the scale parameter is determined by the average mutation magnitude. The Lévy distribution parameters generated in this way can simultaneously reflect both mutation frequency and mutation intensity.

[0108] The generated stability, skewness, and scale parameters are input into the Lévy morphological degradation model, enabling the model to drive random jumps in the time series according to the Lévy distribution characteristics. The driving method is as follows: in each time period, the fluctuation amplitude is determined by the stability parameter, the jump direction by the skewness parameter, and the jump intensity by the scale parameter. By continuously executing the driving process, a Lévy morphological degradation evolution curve that conforms to the actual behavior of hardware degradation can be generated, providing basic data for subsequent calculations of fault energy accumulation rate.

[0109] Risk map generation module: Calculates the fault energy accumulation rate of hardware components based on the Lévy morphological degradation evolution curve, and couples and corrects it with the associated lock-in probability to generate a fault risk heat map.

[0110] In this embodiment, each time period of the Lévy morphological degradation evolution curve includes a degradation magnitude value. The degradation magnitude describes the intensity of hardware degradation within that time period. To characterize the overall energy accumulation trend, the degradation magnitude needs to be processed into absolute values ​​so that both positive and negative degradation can be included in the measurement range. Subsequently, the degradation magnitudes after absolute values ​​are accumulated in chronological order. For example, if the absolute values ​​of the degradation magnitudes in time periods 1, 2, and 3 are 2, 3, and 5 respectively, the accumulated results are 2, 5, and 10. The accumulated results constitute the primary energy accumulation sequence, which reflects the intensity of energy aggregation in the time dimension and provides input data for subsequent smoothing calculations.

[0111] To reduce the interference of transient disturbances on energy analysis, the primary energy accumulation sequence needs to be smoothed. The smoothing method is the sliding window calculation method.

[0112] The length of the sliding window can be set to five consecutive time periods. Starting from the beginning of the sequence, the average growth rate of the accumulated energy value is calculated within each window. The average growth rate is obtained by subtracting the initial value from the final value within the window and then dividing by the window span. For example, if the accumulated energy increases from 10 to 20 within a window, the average growth rate is (20 - 10) divided by 5, resulting in 2. This growth rate is used to represent the intensity of local energy accumulation. The results from all windows constitute a growth rate sequence.

[0113] The growth rate sequence is used to reflect the accumulation and change of energy over time. To further identify regions where energy continues to rise, it is necessary to calculate the energy accumulation gradient. The energy accumulation gradient is obtained by the difference between adjacent growth rate values. For example, if the growth rate increases from 2 to 3, the gradient is 1.

[0114] After constructing the gradient sequence, identify the continuously rising intervals. The length of the continuously rising interval describes the time span of the continuous energy growth, and the magnitude of the continuously rising interval is obtained by subtracting the minimum gradient value from the maximum gradient value within the interval, which is used to represent the strength of the growth change.

[0115] Length and amplitude are used as estimation factors to calculate the final fault energy convergence rate.

[0116] The fault energy accumulation rate needs to consider both duration and intensity. Multiply the rise interval length by 0.4 and the rise amplitude by 0.6, then add them together to obtain the final fault energy accumulation rate. For example, if the rise interval length is 4 and the amplitude is 5, then the energy accumulation rate is (4 x 0.4) plus (5 x 0.6), resulting in 4.6. The energy accumulation rate is used to characterize the severity of energy accumulation before a hardware component fails.

[0117] To facilitate subsequent probabilistic coupling, the energy convergence rate needs to be normalized to the range of 0 to 1. The normalization method is as follows: subtract the minimum convergence rate from the current convergence rate, and then divide by the difference between the maximum and minimum convergence rates, ensuring that the relative magnitudes are maintained after normalization. The normalized sequence is used for subsequent one-to-one alignment with the associated locking probabilities.

[0118] Each time period has an associated lock-in probability. The normalized fault energy convergence rate is aligned by time, and the two are multiplied to obtain the joint impact value. The multiplicative coupling method can simultaneously enhance the suppression effect of low-probability, low-energy regions and amplify the risk expression of high-probability, high-energy regions, making the joint impact value more reflective of the overall risk level of fault occurrence.

[0119] The combined impact value is mapped to a two-dimensional plane according to the physical layout of hardware components. For example, the combined impact values ​​corresponding to the locations of the processor, memory, power supply unit, etc., are filled into a two-dimensional matrix according to the motherboard coordinate system. The color gradient from light to dark corresponds to the combined impact value from low to high. Dark areas represent high-risk areas. The final color distribution forms a fault risk heatmap, which is used to visually indicate the locations where faults are likely to occur.

[0120] Early warning sequence generation module: Extracts features from abnormal gradient regions in the fault risk heatmap to form a micro-degradation early warning sequence.

[0121] The fault risk heatmap consists of multiple pixels, each corresponding to a risk value. To identify regions of sudden risk changes, it is necessary to calculate the magnitude of risk change between each pixel and its neighboring pixels. The initial value of the spatial gradient is calculated as follows: subtract the risk values ​​of the adjacent pixels from the risk value of a given pixel, and then divide this difference by the spatial distance between the two pixels. For example, if the risk values ​​of two pixels are 0.8 and 0.5 respectively, and the spatial distance is 1, then the initial value of the spatial gradient is 0.3.

[0122] To distinguish between normal fluctuations and abrupt changes, a preset gradient threshold needs to be set. The threshold is determined by sorting all initial spatial gradient values ​​and selecting the top 15% as the threshold, ensuring that the threshold reflects the intensity of significant abrupt changes. Pixel regions with initial spatial gradient values ​​exceeding this threshold are marked as aberrant gradient regions, providing a basis for subsequent region analysis.

[0123] Abnormal gradient regions typically consist of multiple spatially adjacent pixels. To accurately identify the shape and extent of these regions, connectivity analysis is required. The connectivity analysis process involves starting from the initial pixel of the abnormal gradient region and checking each pixel in its four cardinal directions (up, down, left, and right) to see if they simultaneously meet the abnormality criteria. If they do, they are merged into the same region. This process continues until all consecutive abnormal pixels have been identified.

[0124] After completing the region boundary identification, the initial values ​​of all spatial gradients within the region are statistically analyzed. The maximum gradient value is extracted to represent the strongest local abrupt change, the average of all gradient values ​​is calculated to describe the overall degradation intensity, and the region area is represented by the number of pixels contained within it, reflecting the extent of degradation diffusion. These three feature factors together constitute the basis for the degradation description of the region.

[0125] As time progresses, the fault risk heatmap will be continuously updated. The feature factors of the abnormal gradient region obtained from each moment need to be arranged sequentially in chronological order so that the degree of degradation at different time points can be continuously expressed.

[0126] To enhance the ability of the sequence to express degradation trends, a window accumulation method is needed to construct a micro-degradation intensity sequence. The sliding window length can be set to three consecutive time periods. The degradation intensity within the window is obtained by multiplying the maximum gradient value of the region by the area of ​​the region and then dividing by the window span, thus measuring the overall intensity of local degradation. For example, if the maximum gradient value within the window is 0.4 and the area of ​​the region is 6, then the window intensity is (0.4 multiplied by 6) divided by 3, resulting in 0.8. The continuous arrangement of window intensities forms the micro-degradation intensity sequence.

[0127] Micro-degradation intensity sequences are used to describe degradation propagation trends. To capture early signs of impending hardware failure, it is necessary to identify intervals of increasing intensity. An interval of increasing intensity is determined by a positive difference between the intensity values ​​of two adjacent windows, indicating an upward degradation trend.

[0128] All intervals with continuous positive growth are extracted to form a micro-degradation early warning sequence. The early warning sequence consists of continuously rising window intensity values, which can provide early warning of hardware components entering the accelerated degradation stage and provide key basis for the construction of subsequent fault prediction coordinates.

[0129] Risk level output module: Generates fault prediction coordinates based on micro-degradation early warning sequence and outputs the server hardware risk level.

[0130] The micro-degradation early warning sequence consists of early warning values ​​over consecutive time periods, with each value corresponding to the potential degradation intensity of the hardware within that time period. To generate coordinate data suitable for trend analysis, each early warning value needs to be extracted chronologically, with its absolute magnitude used as the vertical axis to quantify the degree of degradation. Simultaneously, the time period number to which the early warning value belongs is used as the horizontal axis, arranging all coordinate points in a progressive chronological order. The resulting coordinate set constitutes the basic dataset for fault prediction coordinates, used for subsequent curve fitting and risk level calculation.

[0131] After obtaining the basic dataset of predicted coordinates, it needs to be fitted to identify degradation trends. The fitting method involves calculating the difference in vertical coordinates between adjacent time periods and dividing this difference by the time interval of the horizontal coordinates to obtain a local slope representing the rate of degradation growth. For example, if the time interval is 1 and the vertical coordinate increases from 3 to 5, the local slope is 2. Arranging all local slopes sequentially in chronological order forms the degradation path curve. The degradation path curve reflects the rate of hardware condition evolution over time, providing a basis for constructing subsequent risk level assessment factors.

[0132] The local slope in the degradation path curve is used to characterize the degree of degradation acceleration. To construct the multidimensional evaluation factors required for risk level assessment, three types of values ​​need to be extracted from the curve:

[0133] The maximum slope is used to represent the strongest acceleration behavior during the degradation process;

[0134] The average slope is used to represent the persistence of the overall degradation trend;

[0135] The slope change frequency refers to the proportion of the number of times the slope sign changes relative to the total length of the curve, and is used to describe the degree of fluctuation in the degradation state.

[0136] These three types of evaluation factors together constitute a risk level evaluation factor set, which is used to comprehensively reflect the intensity, persistence, and instability of hardware degradation.

[0137] To arrive at the final risk assessment, a weighted combination of the three evaluation factors is required. The weighting method is as follows: multiply the maximum slope by 0.5, the average slope by 0.3, and the slope change frequency by 0.2. The three results are then summed to obtain the comprehensive risk score. For example, when the maximum slope is 4, the average slope is 2, and the change frequency is 0.3, the comprehensive risk score is (4 × 0.5) plus (2 × 0.3) plus (0.3 × 0.2), resulting in a score of 2.66.

[0138] A comprehensive risk score is used to classify risk levels. A score below 1.5 is defined as low risk, a score between 1.5 and 3 is defined as medium risk, and a score above 3 is defined as high risk. The final server hardware risk level is used for subsequent fault prediction and maintenance decisions.

[0139] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An AI-based server hardware failure prediction system, characterized by: include: Feature construction module: Collects multi-source operating signals generated by the server during operation, and calculates a multi-dimensional energy spectrum feature matrix containing frequency domain energy density distribution through a segmented wavelet convolution fusion algorithm; Topology generation module: Based on the multidimensional energy spectrum feature matrix, a degenerate spatiotemporal coordinate system is constructed to form a node topology graph, and the energy gradient between nodes is calculated to obtain the association locking probability set; Evolution curve generation module: Based on the associated locking probability set, the feature matrix is ​​folded and fitted to construct a micro-degeneration field mapping function and generate the Lévy morphological degradation evolution curve; Risk map generation module: Calculates the fault energy accumulation rate of hardware components based on the Lévy morphological degradation evolution curve, and couples and corrects it with the associated lock-in probability to generate a fault risk heat map; Early warning sequence generation module: Extracts features from abnormal gradient regions in the fault risk heatmap to form a micro-degradation early warning sequence; Risk level output module: Generates fault prediction coordinates based on micro-degradation early warning sequence and outputs the server hardware risk level. 2.The AI-based server hardware failure prediction system of claim 1, wherein: The step of calculating the multidimensional energy spectrum feature matrix containing the frequency domain energy density distribution using the piecewise wavelet convolution fusion algorithm includes: Time synchronization processing is performed on the CPU leakage current drift sequence, fan micro-disturbance speed sequence, and power supply ripple phase noise sequence continuously recorded during server operation; The synchronized operating signals are segmented according to a preset window, and wavelet basis function convolution is performed on each data segment to obtain the frequency domain energy density distribution sequence of each signal at different scales. Based on the frequency domain energy density distribution sequence, cross-dimensional feature alignment is performed on different signals within the same time period, and a unified energy spectrum fragment is generated using a multi-layer convolutional fusion network. All energy spectrum fragments are stacked and matrixed in chronological order to form the multidimensional energy spectrum feature matrix. 3.The AI-based server hardware failure prediction system of claim 1, wherein: The steps for constructing a degenerate spatiotemporal coordinate system and forming a node topology graph based on the multidimensional energy spectrum feature matrix include: The energy density vectors of each time period in the multidimensional energy spectrum feature matrix are extracted in chronological order, and the continuous time length is used as the vertical axis and the energy density change amplitude is used as the horizontal axis to construct a basic framework of degradation spatiotemporal coordinates to reflect the hardware degradation trend. Based on the aforementioned degenerate spatiotemporal coordinate framework, the rate of change between energy density vectors in adjacent time periods is calculated, and a degenerate gradient sequence is obtained through continuous difference. Each degenerate gradient is used as the attribute value of a potential node. Based on the relationship between the attribute values ​​of the nodes, nodes with adjacent attribute values ​​and continuous rate of change are connected by lines, and the degree of gradient proximity between nodes is used as the connection weight to form a primary node topology. A layer-by-layer weighted filtering operation is performed on the primary node topology to remove weakly connected lines with weights below a preset threshold, resulting in a node topology that truly reflects the energy density evolution path. 4.The AI-based server hardware failure prediction system of claim 3, wherein: The steps for calculating the energy gradient between nodes to obtain the set of associated locking probabilities include: The energy density attribute values ​​of any adjacent nodes in the node topology graph are interpolated, and the initial value of the energy gradient is obtained by dividing the difference by the time interval between nodes. All node pairs are sorted according to the initial value of the energy gradient, and node pairs with continuous gradient changes and consistent directions are merged to form an energy gradient sequence. Based on the energy gradient sequence, an interval partitioning model is established according to the distribution range of gradient values ​​in the sequence, and the correlation strength value between nodes is calculated by dividing the number of occurrences of each gradient interval by the total length of the gradient sequence. The association strength value is normalized and coupled with the connection weights in the topology. The normalized coupling result is used as the association locking probability, and the probability values ​​of all node pairs are used to form the association locking probability set. 5.The AI-based server hardware failure prediction system of claim 1, wherein: The steps for generating the Lévy morphological degradation evolution curve include: The feature vectors of each time period in the multidimensional energy spectrum feature matrix are sorted according to the probability magnitude in the associated locking probability set, and then folded and superimposed in a way that the probability is from high to low. The folded feature structure is input into the microdegenerate field mapping function, and the nonlinear coupling strength between feature vectors is calculated in a progressive manner, and the coupling strength is used as the estimate of the local change rate of the microdegenerate field. Based on the local rate of change sequence, Lévy distribution characteristic parameters are constructed according to the joint conditions of jump probability and mutation amplitude, and the characteristic parameters are used to drive the time series to generate Lévy morphological degradation and evolution curves. 6.The AI-based server hardware failure prediction system of claim 5, wherein: The steps for constructing Lévy distribution characteristic parameters based on the local rate of change sequence according to the joint condition of jump probability and abrupt change amplitude include: The local rate of change sequence output by the microdegenerate field mapping function is filtered by amplitude. Time points where the absolute value of the rate of change exceeds a preset mutation threshold are marked as jump events, and the jump probability is calculated based on the proportion of jump events. Statistical analysis is performed on the absolute values ​​of the rate of change of all jump events, and the range of abrupt change amplitude is constructed using the maximum, minimum and mean values, so that each jump event can find a corresponding amplitude level in the range. The jump probability and the occurrence ratio of each amplitude level are jointly merged to generate the stability parameter, skewness parameter and scaling parameter of the Lévy distribution in a joint probability manner. The Lévy distribution characteristic parameters are input into the Lévy morphological degradation model to drive the time series and generate Lévy morphological degradation evolution curves that conform to the hardware degradation jump characteristics. 7.The AI-based server hardware failure prediction system of claim 1, wherein: The steps for calculating the fault energy accumulation rate of the hardware component based on the Lévy morphological degradation evolution curve include: The degradation amplitude of Lévy morphology degradation evolution curve at each time point is extracted, and the absolute value of the degradation amplitude is accumulated in chronological order to form a primary energy accumulation sequence to characterize the energy accumulation trend. The primary energy accumulation sequence is smoothed, and the average growth rate within the window is calculated using a sliding window method, so that the growth rate can reflect the intensity of energy accumulation in the local time range. Based on the changing trend of the growth rate, an energy convergence gradient sequence is constructed, and the length and amplitude of the continuous rising interval in the gradient sequence are used as the estimation factors of the energy convergence rate. The estimated factors are combined in a weighted manner to calculate the fault energy accumulation rate. 8.The AI-based server hardware failure prediction system of claim 7, wherein: The steps for generating the fault risk heatmap include: The fault energy convergence rate was normalized according to time sequence. The normalized fault energy accumulation rate is matched with the correlation lock probability of the corresponding time period, and the joint impact value of the two is calculated by multiplicative coupling. All joint impact values ​​are mapped in two dimensions according to the physical location of the hardware components, and a fault risk heatmap is formed by assigning color gradients to different locations.

9. The AI-based server hardware fault prediction system according to claim 1, characterized in that: The step of extracting features from the abnormal gradient regions of the fault risk heatmap to form a micro-degradation early warning sequence includes: Gradient calculation is performed at each location in the fault risk heatmap. The initial value of the spatial gradient is obtained by dividing the difference of the risk values ​​of adjacent pixels by the spatial distance. Regions whose initial gradient values ​​exceed the preset gradient threshold are marked as abnormal gradient regions. For each abnormal gradient region, a connected component analysis is performed. The boundary range of continuous regions is identified by a point-by-point search method, and the maximum gradient value, average gradient value and area within the region are extracted as feature factors characterizing the degree of local degradation. The feature factors are arranged in chronological order, and a micro-degradation intensity sequence is constructed using a window accumulation method; Based on the continuous change pattern of the microdegradation intensity sequence, the intensity increase interval is extracted as the microdegradation early warning sequence.

10. The AI-based server hardware fault prediction system according to claim 1, characterized in that: The steps for generating fault prediction coordinates and outputting the server hardware risk level based on the micro-degradation early warning sequence include: Each warning value in the micro-degradation warning sequence is extracted in chronological order, and the absolute magnitude of the warning value is used as the vertical axis and the corresponding time position is used as the horizontal axis to construct a basic dataset for fault prediction coordinates. Coordinate fitting is performed on the fault prediction coordinate dataset, and degradation path curves are generated by comparing the changes in longitudinal coordinates over time periods. A set of risk level evaluation factors is constructed based on the maximum value, average value, and frequency of slope change in the degradation path curve; The risk level assessment factors are aggregated in a weighted manner to calculate a comprehensive risk score, and the server hardware risk level is divided into low risk, medium risk and high risk levels according to the score range.

Images