AI equipment full life cycle management system
By combining Jacobi matrix bifurcation theory and generalized extreme value distribution with Chebyshev inequality correction, the nonlinear dynamic characteristics and heavy-tailed extreme events in AI equipment fault prediction are solved, enabling early and accurate fault detection and remaining life assessment, thus improving the reliability and efficiency of equipment management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU DAHESHENG INTELLIGENT TECH GRP CO LTD
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies rely on simple threshold alarms and traditional Gaussian distributions in AI equipment fault prediction, which cannot effectively handle the nonlinear dynamic characteristics and heavy-tailed extreme events of equipment operation. This results in low fault detection accuracy, unreliable remaining life estimation, and difficulty in achieving reliable maintenance and management.
The bifurcation theory of the Jacobian matrix is used to determine the stability of the equipment. The proximity of the failure is quantified by the eigenvalue of the largest real part. The instantaneous failure probability is quantified by the generalized extreme value distribution. The Chebyshev inequality is used for correction. The future failure probability is extrapolated by the adaptive weighted moving average model, and the remaining life interval is output.
It enables early and accurate fault detection and remaining life assessment, reduces resource consumption, improves the reliability of fault prediction and equipment management efficiency, and supports timely maintenance and decision-making.
Smart Images
Figure CN121998683A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of product lifecycle management technology, and in particular to an AI device full lifecycle management system. Background Technology
[0002] AI devices refer to hardware devices used to perform artificial intelligence-related tasks, typically including but not limited to computers, sensors, robots, automated equipment, and smart terminals. These devices are commonly used in fields such as machine learning, data analysis, and automated control to help complete complex computational tasks or make intelligent decisions.
[0003] Implementing full lifecycle management for AI devices is crucial. Firstly, AI devices typically require continuous maintenance and optimization to ensure long-term, efficient operation. Full lifecycle management effectively monitors device usage, performance changes, and potential problems, enabling timely repairs or updates. Secondly, with the rapid development of AI technology, devices are constantly being updated and replaced; timely lifecycle management helps ensure continuous technological innovation and stable system operation. Furthermore, effective lifecycle management can extend device lifespan and reduce operating costs.
[0004] However, existing technologies for predicting faults in AI devices largely rely on simple threshold alarms and traditional Gaussian distributions, which cannot effectively handle the nonlinear dynamic characteristics and heavy-tailed extreme events of device operation. This results in low fault detection accuracy, unreliable remaining life estimation, and difficulty in achieving reliable maintenance and management of AI devices. Summary of the Invention
[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide an AI device lifecycle management system that can solve the technical problems in the prior art where AI device fault prediction relies on simple threshold alarms and traditional Gaussian distributions, which cannot effectively handle the nonlinear dynamic characteristics and heavy-tailed extreme events of device operation, resulting in low fault detection accuracy, unreliable remaining life estimation, and difficulty in achieving reliable maintenance and management of AI devices.
[0006] This invention proposes an AI device lifecycle management system, comprising: The data acquisition module is used to collect the operating parameters of the AI device within a preset time period; The judgment module is used to determine whether the AI device is normal based on the running parameters and the bifurcation theory based on the Jacobian matrix. If it is normal, the synthesis module is called; otherwise, the acquisition module is called. The synthesis module is used to combine the eigenvalues of the largest real part of the Jacobian matrix to synthesize the proximity of the current state of the AI device to the fault state. The quantization module is used to quantify the instantaneous failure probability of AI devices under proximity conditions using a generalized extreme value distribution; The correction module is used to correct the instantaneous failure probability by incorporating Chebyshev's inequality. The extrapolation module is used to extrapolate the instantaneous failure probability at different time steps to obtain the predicted instantaneous failure probability for multiple future time steps. The output module combines the predicted instantaneous failure probabilities to output the remaining lifespan range of the AI device under a preset confidence level.
[0007] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, the bifurcation theory based on the Jacobian matrix is used to determine the inherent dynamic stability of AI devices. Initial assessments avoid full-condition remaining lifetime estimation, effectively reducing resource consumption while ensuring normal operation of the AI device. The calculation of the eigenvalue with the largest real part is used to keenly capture early signs of AI device instability (i.e., failure), thus synthesizing a failure proximity index. Subsequently, considering the heavy-tailed nature of failure data, a generalized extreme value distribution, rather than a traditional Gaussian distribution, is used in conjunction with the proximity index to accurately quantify the instantaneous probability of extreme anomalies. Chebyshev's inequality is then used to conservatively correct this probability to ensure the reliability of the warning. Finally, by extrapolating the probability over multiple future time steps, a reliable remaining lifetime prediction interval can be generated at a specified confidence level. This allows for earlier and more accurate detection of potential failures, providing a quantified and confidence-interval-based remaining lifetime assessment, which is beneficial for the timely maintenance and management of AI devices. Attached Figure Description
[0008] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0009] Figure 1 This is a schematic diagram of the structure of an AI device lifecycle management system provided in an embodiment of the present invention. Detailed Implementation
[0010] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0011] The AI device lifecycle management system provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.
[0012] Reference manual attached Figure 1 The diagram shows a structural schematic of an AI device lifecycle management system provided by an embodiment of the present invention.
[0013] This invention provides an AI device lifecycle management system, including: The data acquisition module 1 is used to collect the operating parameters of the AI device within a preset time period.
[0014] It should be noted that those skilled in the art can set the preset duration according to actual needs, and this invention does not impose any limitations on it. Optionally, the preset duration can be set as the time interval between the arrival of maintenance personnel for the actual AI equipment. This is to avoid situations where maintenance personnel cannot arrive in time after a fault is detected, leading to AI equipment downtime and affecting operation. Sufficient preparation time is allowed for maintenance.
[0015] In one possible implementation, operating parameters include throughput, error rate, output latency, and instantaneous power consumption.
[0016] These metrics include throughput (e.g., the number of inference tasks processed per second), error rate (e.g., the proportion of computational or data errors), and output latency (e.g., the response time of a task from input to output). Specifically, throughput and error rate can be extracted from the logs of AI frameworks (such as TensorFlow and PyTorch). Output latency can be recorded via timestamps. Instantaneous power consumption can be obtained through the energy management of the AI device.
[0017] It should be noted that this acquisition module collects multi-dimensional heterogeneous operating parameters such as throughput, error rate, output latency, and instantaneous power consumption in a coordinated manner. From multiple key dimensions such as performance, accuracy, efficiency, and energy consumption, it continuously senses the internal operating status of AI devices, thereby ensuring efficient and reliable AI device lifecycle management in the future.
[0018] Judgment module 2 is used to determine whether the AI device is normal based on the running parameters and the bifurcation theory based on the Jacobian matrix. If it is normal, the synthesis module is called; otherwise, the acquisition module is called.
[0019] The Jacobian matrix is a matrix composed of first-order partial derivatives. This matrix describes how a small change in each state variable affects the rate of change of all other state variables. Simply put, it's a measure of system sensitivity, revealing how all parameters within the system are interconnected and influence each other. Bifurcation theory focuses on analyzing the critical point at which system properties (such as stability) undergo sudden, qualitative changes when system parameters change slowly. Before this critical point, the system is stable, and small disturbances do not alter its normal operating state. However, once this critical point is crossed, the system's original stable state disappears or becomes unstable, and a small disturbance can cause the system to jump to a completely different behavioral pattern (e.g., "jumping" from normal operation to a failure state).
[0020] In one possible implementation, the judgment module 2 uses a combination of bifurcation theory based on the Jacobian matrix to determine whether the AI device is functioning correctly, specifically for: Calculate the rate of change of each operating parameter, where the rate of change of the operating parameter is specifically the derivative of the operating parameter with respect to time.
[0021] By combining the Jacobi matrix, we establish operating parameter description equations for operating parameters and their rates of change. The Jacobi matrix is a 4x4 matrix, where the element in the i-th row and j-th column represents the influence coefficient of a unit change in the j-th operating parameter on the rate of change of the i-th operating parameter.
[0022] The equation describing the running parameters is as follows: .
[0023] in, This represents the running parameter vector at time t. This represents the vector of the rate of change of the running parameters at time t. Let represent the Jacobian matrix at time t.
[0024] Estimate each element in the Jacobian matrix to obtain the Jacobian estimation matrix.
[0025] Specifically, the elements of the Jacobian matrix can be estimated using the recursive least squares method. First, the input data (within a preset time window) includes vectors of operating parameters at different time steps and vectors of the rate of change of these parameters at different time steps. Then, the optimization objective is determined to be minimizing the sum of squared prediction errors. ,in, This indicates taking the minimum value. Indicates time step marker, Indicates time window, and These represent the running parameter vector and the running parameter change rate vector at time marker k, respectively. Then, the elements in the Jacobian matrix are obtained through iterative solving. Here, the time step refers to the minimum time interval for data acquisition or calculation, which can be set according to actual needs.
[0026] Eigenvalue decomposition is performed on the Jacobi estimation matrix to obtain multiple eigenvalues.
[0027] Specifically, the steps for eigenvalue decomposition of the Jacobian estimation matrix include: First, calculating the characteristic polynomial of the Jacobian matrix, which is obtained by solving the equation that the determinant of the matrix obtained by subtracting a scalar multiple of the identity matrix (a scalar multiple of the identity matrix refers to a matrix obtained by multiplying a scalar by the identity matrix) is equal to zero, thereby obtaining the eigenvalues. Next, solving the characteristic equation yields all the eigenvalues of the matrix. The real parts of the eigenvalues reflect the stability of the system; therefore, by analyzing these eigenvalues, it is possible to determine whether the system is in a normal or abnormal state.
[0028] Extract the real part eigenvalues of each eigenvalue.
[0029] Determine if the eigenvalue of the largest real part is less than zero. If it is, determine that the AI device is normal according to the bifurcation theory; otherwise, determine that the AI device is abnormal.
[0030] Understandably, the maximum real part eigenvalue determines the stability trend (stable / divergent) of AI devices, which is a binary qualitative judgment (i.e., "whether there is a risk of failure"), but it cannot quantify the degree of risk (i.e., "how great the risk"), making it a low-complexity qualitative judgment. Performing a qualitative judgment first can avoid highly complex subsequent quantitative judgments, reduce computing power consumption, avoid excessive resource consumption during the management of AI devices, and reduce the impact on their computing power.
[0031] Specifically, this process estimates the Jacobian matrix, reflecting the dynamic coupling relationship between parameters, online using the recursive least squares method and performs eigenvalue decomposition. Based on bifurcation theory, the sign of the eigenvalue with the largest real part of the Jacobian matrix is used as the criterion for determining system stability. If the eigenvalue is negative, the system is considered stable, and the monitoring cycle is maintained. If the eigenvalue is non-negative, it indicates that the system has reached the instability critical point. This stability criterion replaces the lagging threshold alarm, and by performing low-complexity qualitative judgments first, it effectively filters out truly dangerous conditions that require subsequent high-complexity quantitative analysis. This significantly improves the computational efficiency and resource utilization of the entire management system while ensuring optimal early warning capabilities.
[0032] Synthesis module 3 is used to combine the eigenvalues of the largest real part of the Jacobian matrix to synthesize the proximity of the current state of the AI device to the fault state.
[0033] Among them, the eigenvalue with the largest real part is the decisive factor for system stability. Its value determines whether the system state spontaneously reverts to normal (negative value) or tends towards instability (positive value). The proximity degree maps this eigenvalue into a quantitative index to characterize how close the current state is to a failure.
[0034] In one possible implementation, the synthesis module 3 is specifically used for: Calculate the standard deviation of all the largest real part eigenvalues extracted within a preset time period.
[0035] It is understandable that each set of running data collected within the preset time period has a maximum real part eigenvalue, thus ultimately resulting in multiple maximum real part eigenvalues.
[0036] Using the maximum real part eigenvalue being zero as a fixed critical value, the standardized stability deviation of the AI device is calculated, where the standardized stability deviation is specifically the quotient of the maximum real part eigenvalue and the standard deviation.
[0037] It is understandable that the maximum real part eigenvalue being zero indicates that the AI device has entered a stable state at the bifurcation point between instability and stability. Therefore, zero is used as a fixed critical value, and the actual calculation process of the standardized stability deviation is to take the difference between the maximum real part eigenvalue and zero at the judgment time, and then take the quotient of the difference and the standard deviation as the standardized stability deviation.
[0038] Establish a noise suppression term for standardized and stable deviation to avoid misjudgment of faults caused by sudden changes in operating parameters.
[0039] Specifically, the noise suppression term is obtained by taking half the negative of the square of the Euclidean norm of the vector of the rate of change of the running parameters at the current time step, and then taking its natural exponent.
[0040] The proximity is obtained by multiplying the standardized stability deviation and the noise suppression term.
[0041] The formula for calculating proximity is as follows: .
[0042] in, This represents the standard deviation at time t. Represents the eigenvalue with the largest real part at time t. Represents the natural index. This represents the vector of the rate of change of the running parameters at time t. This indicates the degree of proximity to time t.
[0043] Specifically, when operating parameters undergo drastic changes, the magnitude of the rate of change vector increases sharply, causing the exponential term to decay rapidly to near zero, thus making the final proximity index very small. This mathematically automatically identifies and suppresses abnormal signals caused by transient disturbances, ensuring that only stability degradation exhibiting persistent and systematic deviations is identified as a real failure risk.
[0044] Specifically, the scheme first calculates the standard deviation of all maximum real part eigenvalues within a preset time period. Then, it divides the maximum real part eigenvalue at the current moment by its standard deviation to obtain a standardized stability deviation, which characterizes the degree to which the current stability deviates from the norm. To address interference caused by parameter transients, the scheme introduces an exponential noise suppression term based on the magnitude of the vector of the rate of change of the operating parameters. Finally, the standardized stability deviation is multiplied by this noise suppression term to synthesize the final proximity index.
[0045] This process eliminates inherent fluctuations across different equipment or operating conditions by dividing by the standard deviation, making the indicators universally comparable. The noise suppression term cleverly distinguishes between genuine stability degradation and transient random disturbances. When parameters change drastically, this term approaches zero, automatically masking false alarms. It can sensitively reflect a continuous quantitative scale as the system approaches the fault boundary and significantly improves early warning accuracy through a unique noise reduction design, achieving a crucial leap from qualitative judgment to robust quantitative assessment.
[0046] Quantization module 4 is used to quantify the instantaneous failure probability of AI devices under proximity conditions through generalized extreme value distribution.
[0047] The generalized extreme value distribution (GEP) is a statistical distribution model established to analyze and predict extreme values (i.e., the maximum or minimum values in data). Instantaneous failure probability specifically refers to the immediate likelihood of equipment failure given a proximity metric. Quantifying the instantaneous failure probability of equipment using the GEP is particularly suitable for handling the heavy-tailed characteristics of failure data. Compared to the traditional Gaussian distribution, the GEP is more accurate in handling extreme failure events and can effectively capture tail risks, thereby improving the accuracy of failure prediction. This method can provide more forward-looking failure probability predictions when equipment is close to failure, helping to take timely preventative measures and reduce losses from sudden failures.
[0048] In one possible implementation, quantization module 4 is specifically used for: Obtain historical operating parameters of AI devices.
[0049] Determine the sample values of the proximity of multiple fault occurrence times under historical operating parameters.
[0050] The shape parameters of the generalized extreme value distribution are estimated from the sample values of each proximity level. The shape parameters include the location parameter describing the central tendency of the proximity level, the scale parameter describing the discrete tendency of the proximity level, and the shape parameter describing the heavy-tailed distribution of the proximity level.
[0051] Among them, the proximity heavy-tailed distribution refers to the fact that although the probability of extremely high proximity sample values appearing in historical fault data is low, their actual occurrence rate is significantly higher than that predicted by the traditional Gaussian distribution.
[0052] In one possible implementation, estimating the shape parameters of the generalized extreme value distribution from each proximity sample value specifically includes: Construct the log-likelihood function for the generalized extremum distribution.
[0053] The specific formula for the log-likelihood function is: .
[0054] in, Represents the logarithmic function. Indicates the time when the i-th fault occurs. The proximity sample values, where m represents the total number of times the fault occurred. Indicates about position parameters Scale parameters and shape parameters The log-likelihood function value.
[0055] Initialize the position parameter, scale parameter, and shape parameter. The initial value of the position parameter is the median of the nearest neighbor sample value, the initial value of the scale parameter is the standard deviation of the nearest neighbor sample value, and the initial value of the shape parameter is the minimum value that is greater than zero.
[0056] Optionally, the minimum value can be set according to actual needs, specifically a value greater than zero and less than 1, such as 0.08 or 0.1.
[0057] Based on the initial values of the position parameter, scale parameter, and shape parameter, the system aims to maximize the log-likelihood function and outputs estimated values of the position parameter, scale parameter, and shape parameter.
[0058] The objective is to maximize the log-likelihood function, outputting estimates of the position, scale, and shape parameters. Specifically, Newton's method can be used to output these estimates. The steps are as follows: First, choose initial values and construct the log-likelihood function. Then, calculate the first derivative (gradient) and second derivative (Hessian matrix) of the log-likelihood function. Next, update the parameter values using Newton's method, the formula being the current parameter value minus the product of the inverse of the Hessian matrix and the gradient. Repeat this process until the parameter update values are sufficiently small, indicating convergence. The final outputs are the estimates of the position, scale, and shape parameters.
[0059] Specifically, this process estimates the shape parameters of the generalized extreme value distribution by maximizing the log-likelihood function, thereby quantifying the instantaneous failure probability of AI devices. First, a log-likelihood function of the generalized extreme value distribution is established, including location, scale, and shape parameters. Initial values are set, and the log-likelihood function is optimized using Newton's method, iteratively updating the location, scale, and shape parameters until convergence. This method obtains optimal parameters by maximizing the likelihood of the observed data, enabling the model to most accurately fit the distribution characteristics of the failure data. Newton's method efficiently finds the optimal solution for the parameters, converges faster than traditional methods, and improves the accuracy and computational efficiency of the estimation through reasonable initial value selection. This method can accurately describe the distribution of extreme values in equipment failure data, thereby improving the reliability and accuracy of failure probability prediction and providing strong support for equipment failure prevention and decision-making.
[0060] Based on the estimated shape parameters, the instantaneous failure probability is calculated using the generalized extreme value distribution.
[0061] The specific formula for calculating the instantaneous failure probability is as follows: .
[0062] in, This represents the instantaneous failure probability of the AI device at time t. , and These represent the estimated position parameters, scale parameters, and shape parameters, respectively. This indicates the degree of proximity to time t.
[0063] The estimated shape parameters include position parameter estimates, scale parameter estimates, and shape parameter estimates.
[0064] Specifically, this quantification process utilizes the generalized extreme value distribution (GEP) to accurately estimate the instantaneous failure probability of AI devices, making it particularly suitable for handling the heavy-tailed characteristics of device failure data. First, historical operating parameters of the AI device are acquired, and sample values of the proximity of multiple failure occurrence times in the historical data are analyzed. Then, three key shape parameters of the GEP are estimated from these sample values: a location parameter (describing the central tendency of proximity), a scale parameter (describing the discrete tendency of proximity), and a shape parameter (describing the heavy-tailed characteristics of failure precursors). Next, based on these estimated shape parameters, the instantaneous failure probability of the device at a given moment is calculated using the GEP model, reflecting the likelihood of failure when the device is close to a failure state. This allows for more accurate capture of extreme failure events, especially when the device is close to failure. Compared to the traditional Gaussian distribution, the GEP is better able to handle tail risks (i.e., low-probability, high-impact events), thus providing higher failure prediction accuracy and helping to take timely preventative measures to reduce the impact of sudden failures on device operation.
[0065] Correction module 5 is used to correct the instantaneous failure probability by incorporating Chebyshev's inequality.
[0066] Chebyshev's inequality is a statistical inequality used to describe the degree to which a random variable deviates from its expected value. In this scheme, Chebyshev's inequality is used to correct for instantaneous failure probabilities by applying an upper bound to the failure probability, thereby reducing prediction uncertainty. Its advantage lies in providing a relatively conservative correction value regardless of the specific form of the data distribution, avoiding over-prediction caused by extreme data or outliers. This method enhances the robustness and reliability of failure probability prediction, enabling it to work effectively even with complex and incomplete data.
[0067] In one possible implementation, the correction module 5 is specifically used for: Calculate the cumulative energy consumption of the AI device over a preset time period, where the cumulative energy consumption is the integral of the square of the Euclidean norm of the vector of rate of change of operating parameters over the preset time period.
[0068] Understandably, the greater the fluctuation in the operating parameters of AI devices, the more energy is dissipated per unit time, the greater the accumulated total energy dissipation, and the more severe the aging of the devices.
[0069] Calculate the mean and variance of cumulative energy consumption, where the mean of cumulative energy consumption is the quotient of cumulative energy consumption and preset duration.
[0070] Treating cumulative energy consumption as a random variable, the vulnerability probability of AI devices exceeding the average aging level is estimated using Chebyshev's inequality. Specifically, the vulnerability probability is the ratio of the variance of cumulative energy consumption to the square of the mean of cumulative energy consumption.
[0071] The corrected failure probability is obtained by multiplying the instantaneous failure probability by the vulnerability probability.
[0072] It should be noted that the instantaneous failure probability is calculated only based on the instantaneous state of the current operating parameters and fails to reflect the cumulative aging effect of the equipment over long-term operation. By introducing Chebyshev's inequality, the probability of vulnerability caused by aging is dynamically estimated using the statistical characteristics (mean and variance) of historical cumulative energy dissipation, thus coupling instantaneous risk with cumulative damage. This correction suppresses the failure probability of new equipment (low aging level) while amplifying the failure probability of equipment operating under high load for a long time (severe aging), thereby more accurately reflecting the true risk status of the equipment throughout its entire life cycle and avoiding misjudgments or missed alarms due to ignoring historical damage.
[0073] Specifically, this correction process adjusts the instantaneous failure probability of AI devices by incorporating Chebyshev's inequality, aiming to accurately assess failure risk throughout the device's entire lifecycle. First, the cumulative energy consumption of the AI device over a preset period is calculated as an indicator of its aging level. Next, based on the mean and variance of the cumulative energy consumption, Chebyshev's inequality is used to estimate the vulnerability probability exceeding the average aging level, which helps capture the aging effects caused by long-term operation. Finally, the instantaneous failure probability is combined with the vulnerability probability to obtain the corrected failure probability, thus more accurately reflecting the failure risk of the device at different operational stages. This correction method comprehensively considers the device's immediate state and long-term aging effects, avoiding misjudgments caused by ignoring historical damage, and improving the accuracy and reliability of failure prediction, especially when facing long-term high-load operation of the device.
[0074] Extrapolation module 6 is used to extrapolate the instantaneous failure probability at different time steps to obtain the predicted instantaneous failure probability for multiple future time steps.
[0075] In one possible implementation, the extrapolation module 6 is specifically used for: By extrapolating the instantaneous failure probability at different time steps using an adaptive weighted moving average model, the predicted instantaneous failure probability for multiple future time steps can be obtained.
[0076] Specifically, firstly, a time series of historical failure probabilities (i.e., instantaneous failure probabilities at different time steps) is constructed, and an adaptive weighted moving average model is established. The weights of the adaptive weighted moving average model are dynamically adjusted to minimize the prediction error of each historical failure probability, thereby training the adaptive weighted moving average model. Subsequently, a multi-step rolling prediction method is adopted, in which historical data within the current window are input into the trained adaptive weighted moving average model, and future prediction sequences are iteratively generated, thus obtaining the predicted instantaneous failure probabilities for multiple future time steps.
[0077] It should be noted that, compared with traditional static models, adaptive weighted moving average models can better cope with nonlinear changes in data, reduce prediction errors, and improve the accuracy of fault risk prediction.
[0078] Output module 7 is used to combine the predicted instantaneous failure probabilities and output the remaining lifespan range of the AI device under the preset confidence level.
[0079] It should be noted that those skilled in the art can set the preset confidence level according to actual needs, and this invention does not limit this.
[0080] In one possible implementation, output module 7 is specifically used for: The cumulative failure probability of the AI device before different future moments is obtained by summing up the predicted instantaneous failure probabilities. The absolute value of the difference between the cumulative failure probability and the first value is the survival probability of the AI device working normally after the same future moment.
[0081] Obtain the distribution data of cumulative failure probability.
[0082] Under a pre-set confidence level, and by combining distribution data, the upper and lower limits of the AI device's lifespan are derived by inversely calculating the survival probability: .
[0083] .
[0084] in, Indicates pre-set reliability. This indicates that the runtime of the AI device is the lower limit of the AI device's lifespan. The probability of survival while still working normally afterward. This indicates that the runtime of the AI device is the upper limit of the AI device's lifespan. The probability of survival while still working normally afterward. This indicates the lower limit of the AI device's lifespan. The cumulative failure probability of internal failures. This indicates the upper limit of the lifespan of an AI device. The cumulative failure probability of internal failures.
[0085] The remaining lifespan of the AI device is obtained by removing the running time of the AI device from the upper limit and lower limit of the AI device lifespan. The remaining lifespan range of the AI device is formed by removing the upper limit and lower limit of the AI device lifespan from the upper limit and lower limit of the AI device lifespan.
[0086] Specifically, survival probability is used to inversely estimate the remaining lifespan of equipment, determining its upper and lower limits. The already operational time is then removed from these limits to obtain the equipment's remaining lifespan range. This method provides a reliable remaining lifespan range based on dynamic predictions of failure probability and confidence levels. This helps decision-makers rationally plan equipment maintenance and replacement, avoiding negative impacts on equipment operating efficiency and resource allocation due to overly conservative or overly optimistic predictions. Through this prediction, the accuracy of equipment management can be improved, enhancing operational efficiency and equipment lifecycle management.
[0087] In one possible implementation, it also includes: The early warning module 8 is used to issue an early warning when the remaining lifespan of the AI device is less than the preset remaining lifespan of the AI device.
[0088] It should be noted that those skilled in the art have already set the minimum value of the remaining lifespan of the AI device according to actual needs, and this invention does not limit that value.
[0089] In practical applications, the principle of this solution is as follows: First, the acquisition module collects the device's operating parameters in real time, such as throughput, error rate, output latency, and power consumption. Using the Jacobian matrix and bifurcation theory, it determines whether the device is functioning normally and qualitatively assesses its stability using the eigenvalue of its maximum real part. Next, the synthesis module quantifies the proximity of the device's current state to a fault state, accurately quantifies the fault probability using a generalized extreme value distribution, and then corrects the instantaneous fault probability using Chebyshev's inequality, taking into account the device's aging effects. The extrapolation module predicts the device's future fault probability using an adaptive weighted moving average model and uses the cumulative fault probability to infer the device's remaining lifespan. Based on this, the output module provides a remaining lifespan prediction based on pre-set confidence levels, helping decision-makers optimize device maintenance plans. The entire solution, through multi-dimensional data fusion and dynamic modeling, accurately reflects the device's health status and provides effective early warnings and decision support based on reliable fault prediction, improving the accuracy and efficiency of device management and reducing the risk of AI device failure.
[0090] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, the bifurcation theory based on the Jacobian matrix is used to determine the inherent dynamic stability of AI devices. Initial assessments avoid full-condition remaining lifetime estimation, effectively reducing resource consumption while ensuring normal operation of the AI device. The calculation of the eigenvalue with the largest real part is used to keenly capture early signs of AI device instability (i.e., failure), thus synthesizing a failure proximity index. Subsequently, considering the heavy-tailed nature of failure data, a generalized extreme value distribution, rather than a traditional Gaussian distribution, is used in conjunction with the proximity index to accurately quantify the instantaneous probability of extreme anomalies. Chebyshev's inequality is then used to conservatively correct this probability to ensure the reliability of the warning. Finally, by extrapolating the probability over multiple future time steps, a reliable remaining lifetime prediction interval can be generated at a specified confidence level. This allows for earlier and more accurate detection of potential failures, providing a quantified and confidence-interval-based remaining lifetime assessment, which is beneficial for the timely maintenance and management of AI devices.
[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.
Claims
1. An AI device lifecycle management system, characterized in that, include: The data acquisition module is used to collect the operating parameters of the AI device within a preset time period; The judgment module is used to determine whether the AI device is functioning normally based on the operating parameters and the bifurcation theory based on the Jacobian matrix; if so, the synthesis module is invoked. Otherwise, invoke the acquisition module; A synthesis module is used to combine the eigenvalues of the largest real part of the Jacobian matrix to synthesize the proximity between the current state and the fault state of the AI device; The quantization module is used to quantify the instantaneous failure probability of the AI device under the said proximity level through a generalized extreme value distribution; The correction module is used to correct the instantaneous failure probability by incorporating Chebyshev's inequality. The extrapolation module is used to extrapolate the instantaneous failure probability at different time steps to obtain the predicted instantaneous failure probability for multiple future time steps. The output module is used to combine the predicted instantaneous failure probabilities to output the remaining lifespan range of the AI device under a preset confidence level.
2. The AI device lifecycle management system according to claim 1, characterized in that, The operating parameters include throughput, error rate, output latency, and instantaneous power consumption.
3. The AI device lifecycle management system according to claim 1, characterized in that, The judgment module, which combines bifurcation theory based on the Jacobian matrix to determine whether the AI device is functioning correctly, is specifically used for: Calculate the rate of change of each of the operating parameters, wherein the rate of change of the operating parameter is specifically the derivative of the operating parameter with respect to time; Based on the Jacobi matrix, an operating parameter description equation is established for the operating parameter and the rate of change of the operating parameter. The Jacobi matrix is a four-row, four-column matrix, and the element in the i-th row and j-th column of the Jacobi matrix represents the influence coefficient of the unit change of the j-th operating parameter on the rate of change of the i-th operating parameter. Estimate each element in the Jacobian matrix to obtain the Jacobian estimation matrix; The Jacobian estimation matrix is decomposed into eigenvalues to obtain multiple eigenvalues; Extract the real part eigenvalues of each of the aforementioned eigenvalues; Determine whether the eigenvalue of the largest real part is less than zero; if so, determine that the AI device is normal according to the bifurcation theory; otherwise, determine that the AI device is abnormal.
4. The AI device lifecycle management system according to claim 1, characterized in that, The synthesis module is specifically used for: Calculate the standard deviation of all the maximum real part eigenvalues extracted within the preset time period; Using the maximum real part eigenvalue being zero as a fixed critical value, the standardized stability deviation of the AI device is calculated, wherein the standardized stability deviation is specifically the quotient of the maximum real part eigenvalue and the standard deviation; Establish a noise suppression term for the standardized stability deviation to avoid misjudgment of faults caused by jumps in the operating parameters; The proximity degree is obtained by multiplying the standardized stable deviation and the noise suppression term.
5. The AI device lifecycle management system according to claim 1, characterized in that, The quantization module is specifically used for: Obtain the historical operating parameters of the AI device; Determine the sample values of the proximity of multiple fault occurrence times under the historical operating parameters; The shape parameters of the generalized extreme value distribution are estimated from each of the proximity sample values, wherein the shape parameters include a location parameter describing the central tendency of proximity, a scale parameter describing the discrete tendency of proximity, and a shape parameter describing the heavy-tailed distribution of proximity. Based on the estimated shape parameters, the instantaneous fault probability is calculated using the generalized extreme value distribution.
6. The AI device lifecycle management system according to claim 5, characterized in that, The estimation of the shape parameters of the generalized extreme value distribution from each of the proximity sample values specifically includes: Establish the log-likelihood function for the generalized extreme value distribution; Initialize the position parameter, the scale parameter, and the shape parameter, wherein the initial value of the position parameter is the median of each proximity sample value, the initial value of the scale parameter is the standard deviation of each proximity sample value, and the initial value of the shape parameter is a local minimum value greater than zero. Based on the initial values of the position parameter, the initial values of the scale parameter, and the initial values of the shape parameter, with the objective of maximizing the log-likelihood function, the estimated values of the position parameter, the estimated values of the scale parameter, and the estimated values of the shape parameter are output.
7. The AI device lifecycle management system according to claim 1, characterized in that, The correction module is specifically used for: Calculate the cumulative energy consumption of the AI device within the preset time period, wherein the cumulative energy consumption is the integral value of the square of the Euclidean norm of the rate of change vector of the operating parameters within the preset time period; Calculate the cumulative energy consumption mean and cumulative energy consumption variance, wherein the cumulative energy consumption mean is the quotient of the cumulative energy consumption and the preset duration; Treating the cumulative energy consumption as a random variable, the vulnerability probability of the AI device exceeding the average aging level is estimated using the Chebyshev inequality, wherein the vulnerability probability is specifically the ratio of the variance of the cumulative energy consumption to the square of the mean of the cumulative energy consumption. The instantaneous failure probability is multiplied by the vulnerability probability to obtain the corrected failure probability.
8. The AI device lifecycle management system according to claim 1, characterized in that, The extrapolation module is specifically used for: The instantaneous failure probability at different time steps is extrapolated using an adaptive weighted moving average model to obtain the predicted instantaneous failure probability for multiple future time steps.
9. The AI device lifecycle management system according to claim 1, characterized in that, The output module is specifically used for: The predicted instantaneous failure probabilities are accumulated to obtain the cumulative failure probability of the AI device before different future moments, wherein the absolute value of the difference between the cumulative failure probability and the value one is the survival probability of the AI device working normally after the same future moment. Obtain the distribution data of the cumulative failure probability; Under the preset confidence level, combined with the distribution data, the upper limit and lower limit of the AI device lifespan are obtained by inverse calculation using the survival probability. The remaining lifespan of the AI device is obtained by removing the running time of the AI device from the upper limit and the lower limit of the AI device lifespan, respectively. The remaining lifespan range of the AI device is formed by removing the upper limit and the lower limit of the AI device lifespan.
10. The AI device lifecycle management system according to claim 9, characterized in that, Also includes: The early warning module is used to issue an early warning when the remaining lifespan of the AI device is less than the preset remaining lifespan of the AI device.