A Reliability Analysis Method and Apparatus for Intelligent Systems Addressing Soft Faults

Abstract

This invention discloses a method and apparatus for reliability analysis of intelligent systems oriented towards soft errors. The method includes: dividing the intelligent system into different functional modules; establishing a soft error-oriented reliability model (FMRM) for each functional module; converting the reliability model into an evaluation model to calculate the reliability of each functional module; obtaining a joint function describing the correlation of failure rates between modules based on the failure rates of different functional modules; establishing a probabilistic topological relationship for the entire intelligent system based on the reliability probabilities of the functional modules and the joint function, establishing an overall reliability evaluation model (MIR-BN), and obtaining the overall reliability result of the intelligent system through probabilistic reasoning; and identifying key reliability modules in the intelligent system through reverse reasoning. The method of this invention adopts a hierarchical approach to analyze and evaluate the reliability of intelligent systems. The model has a clear hierarchy and is scalable, enabling both qualitative and quantitative analysis, thereby improving the accuracy of reliability analysis and evaluation.

Description

Technical Field

[0001] This invention relates to the field of reliability analysis technology, and in particular to a method and apparatus for reliability analysis of intelligent systems oriented towards soft errors. Background Technology

[0002] The single-event effect (SEE) is a physical phenomenon where high-energy charged particles from radiation bombard sensitive nodes in microelectronic devices, such as microprocessors and semiconductor memories, causing a unit flip of the information logic bits. SEE can not only damage devices but also easily lead to errors in program execution, program crashes, and even irreversible serious consequences. For example, in 2011, China's first Mars probe, "Yinghuo-1," malfunctioned and failed to complete its Mars exploration mission. The culprit was a soft error caused by a SEE triggered by space radiation. In 2016, Sun Microsystems in the United States also suffered tens of thousands of dollars in economic losses due to a soft error in its memory, affecting the normal operation of numerous servers across the country. In 2017, China's first dark matter detector, "Wukong," was also temporarily "blinded" for nearly 19 hours due to a soft error caused by a SEE, resulting in significant losses.

[0003] Soft errors are recoverable errors caused by radiation, interference, etc., and do not damage the hardware itself. They are transient logic errors caused by the flipping of one or more bits in a hardware unit, resulting in some information bits changing from 0 to 1 or 1 to 0. Some bit flips can be automatically recovered, while others require techniques such as restarting or error recovery. Although soft errors do not cause permanent hardware damage, they can change values ​​in storage components or the state of the processor and are propagating, thus interfering with the normal operation of the system, causing errors in output or operation, and in severe cases, even causing system crashes.

[0004] Therefore, to avoid system malfunctions or crashes due to soft errors, it is necessary to analyze and evaluate the soft errors that occur in the system and their potential impact. By analyzing and evaluating system reliability, vulnerable modules can be identified, and necessary technical measures can be taken to prevent system failures after operation. Therefore, evaluating the reliability of intelligent systems is of great significance.

[0005] Currently, there are three main methods for evaluating system reliability: The first is fault simulation-based methods, which mimic the occurrence of faults / errors in the target system to accurately reflect their characteristics, including behavioral features and temporal and spatial distribution, ensuring the reliability evaluation results accurately reflect the system's reliability. The second is fault-free simulation methods, which do not require fault injection. They directly use architecture time simulators or performance simulators to simulate the benchmark program, and then analyze the operation of each system component (e.g., utilization) to assess the likelihood of system failure due to soft errors. The third is formal methods, which typically extract information from the original design of the target system, construct a formal model, describe faults using formal languages, and finally use formal verification techniques to verify whether certain reliability properties of the system are still maintained in the system model after fault injection. Formal modeling methods can be broadly categorized into semi-formal and formal methods. Semi-formal methods include AADL (Architecture Analysis and Design language) and MARTE (Modeling and Analysis of Real-Time and Embedded systems), while formal methods include Z language, Petri nets, Markov models, and Bayesian networks (BN). Z language is the most widely used formal description language. Compared to other formal methods, Z language offers advantages such as strong descriptive power, model scalability, and modular modeling, making it a commonly used method for formal software modeling.

[0006] Basis Networks (BNs) are probabilistic networks represented by directed acyclic graphs (DAGs) consisting of nodes and directed edges connecting them. Nodes represent variables, and directed edges represent dependencies between nodes. BNs are suitable for expressing and analyzing uncertainties, possessing the ability to describe the polymorphism and nondeterministic logical relationships of events, and effectively addressing the shortcomings of traditional reliability analysis methods. BNs can predict the probability of unknown variables and, based on variables with determined states, deduce the update probability of known given variables through probabilistic reasoning. Therefore, applying BNs to the reliability assessment of intelligent systems can more clearly express the state relationships between the various functional modules of the intelligent system. Furthermore, based on the obtained reliability of each functional module and Bayes' theorem, the reliability of the intelligent system and its key reliability modules can also be calculated. Summary of the Invention

[0007] The purpose of this invention is to provide a method and apparatus for reliability analysis of intelligent systems oriented towards soft errors. This method can accurately describe the dynamic characteristics of intelligent systems and their functional modules in relation to time, and can effectively analyze and evaluate the reliability of the overall intelligent system and its individual functional modules, providing an effective method for identifying vulnerable modules. Furthermore, this invention can also be applied to functional components such as CPUs, registers, and memory, enabling reliability assessment of these components from the perspective of their microarchitecture.

[0008] The technical solution for achieving the objective of this invention is: a reliability analysis method for intelligent systems oriented towards soft errors, the method comprising:

[0009] For the intelligent system to be analyzed and evaluated, it is divided into different functional modules (MIFM) according to its structure; the functional modules include CPU, registers, memory, cache, high-speed cache, I / O and bus;

[0010] Establish corresponding models for the functional modules to obtain reliability models (FMRM) for different functional modules;

[0011] The reliability models of each functional module are converted into corresponding reliability assessment models (MREMs) to obtain the reliability assessment results of different functional modules.

[0012] Based on the failure rate of each functional module, the failure correlation between modules is analyzed, and the joint function describing the failure rate correlation between modules is estimated.

[0013] Based on the failure rate and correlation joint function of each functional module, a system overall reliability assessment model MIR-BN is established based on Bayesian network to obtain the overall system reliability assessment result. Then, based on the system overall reliability assessment model, the key reliability modules of the system are obtained by back reasoning.

[0014] Furthermore, corresponding models are established for the functional modules to obtain reliability models for different functional modules, which are determined in the following manner:

[0015] Determine the underlying architecture of each functional module and divide the functional modules into different functional components;

[0016] Determine the parameters in the reliability model, including the failure rate of the functional modules, the state space of the functional modules, and all the state transition relationships of the functional modules;

[0017] The reliability model of a functional module is constructed through the following steps;

[0018] Step 2.1: Determine the modules included in the functional module and their failure rates. Taking the CPU as an example, the CPU can be divided into modules such as the Prefetch Unit (PFU), Memory Protection Unit (MPU), Load / Store Unit (LSU), and Data Processing Unit (DPU). The inherent failure rate of the functional module can be determined by the following formula:

[0019] λ CPU =1-(1-λ) PFU )*(1-λ MPU )*(1-λ LSU )*(1-λ DPU )

[0020] Where, λ PFU ,λ MPU ,λ LSU ,λ DPU These represent the inherent failure rates of the prefetch unit, memory protection unit, load storage unit, and data processing unit, respectively.

[0021] Step 2.2, establish the reliability model of the functional module, represented by triples:

[0022] FMRM = (FRate, States, STR)

[0023] Here, FRate represents the failure rate of the functional module, and the FRate of each functional module can be obtained by the calculation method of the inherent failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state and recovery state; STR represents all state transition relationships existing in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state.

[0024] Furthermore, the reliability assessment model for the functional module is determined in the following manner:

[0025] Based on the established functional module reliability model, the corresponding FMRM model is converted into the corresponding MREM evaluation model according to the conversion rules.

[0026] Step 3.1: The basic structure of the functional module reliability assessment model is determined by the following quadruple formula:

[0027] MREM=(S,S in ,A,t)

[0028] Where S represents the state space of MREM, which is the set of all possible states of the functional modules, S = {s1, s2, ..., s} n}, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability ∈ S; t represents the moment when the reliability is evaluated;

[0029] Step 3.2: Perform model conversion according to the conversion rules. To ensure that the entire conversion process between the established functional module reliability model and the MREM evaluation model is equivalent, i.e., to maintain consistency in the description of reliability before and after conversion, it is necessary to analyze and compare the elements contained in both the FMRM model and the MREM model, and establish equivalence mapping rules for the common elements in both models.

[0030] Furthermore, after converting the reliability models of each functional module into corresponding reliability assessment models (MREMs), the process also includes:

[0031] Step 4.1: Evaluate the reliability of the functional modules using state transition equations;

[0032] The elements included in the state transition equation—the probability vectors of each state the functional module is in at a given moment and the probability matrix of possible state transitions for the functional module—are determined by the following formula:

[0033] P(t')=P(t)*A

[0034] Where P(t) represents the state probability vector of the functional module at time t, P(t') = (P N (t'),P R (t'),P F (t')) represents the state probability vector of the functional module at time t', P N (t'),P R (t') and P F (t') represents the probability that the functional module is in the normal state (NS), recovery state (RS), and failure state (FS) at time t', respectively; t' represents the next time step after time t, and A represents the state transition probability matrix of the functional module;

[0035] The relationship between the failure rate of a functional module and time is determined by the following formula:

[0036] λ(t) = exp(λ*t)

[0037] Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0;

[0038] Step 4.2: By solving the state transition equation, the failure probability of the functional module at time t can be obtained. The reliability of the functional module at time t is determined by the following formula:

[0039] R MIFM (t)=1-P F (t)

[0040] Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure rate of the functional module at time t.

[0041] Furthermore, the joint function describing the correlation of failure rates among modules is determined by the following method:

[0042] The relationship function of the failure rate of the functional module over time is used as the failure distribution function of the functional module, and the joint function describing the correlation of failure rates between modules is used as the correlation joint function between modules.

[0043] The joint function describing the correlation of failure rates between modules is determined in the following way:

[0044] Step 5.1: Determine the candidate joint functions, which are Copula joint functions, including the Frank function, Clayton function, and Gumbel function;

[0045] Step 5.2: Determine the parameters in the joint function to be selected, and sample the parameters in the joint function using the MH algorithm; use the normal distribution as the proposal density function, sample from the proposal density function, and generate a sequence α according to the Markov chain. (0) →α (1) →...→α (n) Initialize the parameters in the joint function, given an initial value α for each parameter. (0) And let u = α (0) Based on the prior information of the parameters, a sample α is drawn from the proposed density function. (1) ~μ(N,σ 2 The acceptance probability r of the MH algorithm is calculated based on the posterior distribution, and r is determined by the following formula:

[0046]

[0047] Where p(·|·) is the posterior distribution, and J(·|·) is the proposal density distribution; if r > u, then accept α. (1) Otherwise, α (1) =α (0) The extracted parameter values ​​are used as new initial values ​​for sampling. The sampling is then performed iteratively according to the parameter estimation method described above until the loss rate distribution converges to the target distribution. Sampling stops when the Markov chain reaches stationarity.

[0048] Step 5.3: Determine the joint function describing the failure correlation between modules; select the joint function by measuring the squared Euclidean distance. The smaller the squared Euclidean distance, the more accurately the selected joint function describes the failure correlation between modules.

[0049] Furthermore, the overall system reliability assessment model is determined by the following method:

[0050] Using the failure of the intelligent system to be evaluated as the central node and the failure of the functional module as the edge node, the edge node is connected to the central node through a connecting arc, thus obtaining the reliability model of the intelligent system.

[0051] Furthermore, a reliability assessment of the system is performed based on the aforementioned intelligent system reliability model, including:

[0052] Obtain the failure probability of the functional module, and determine the qualitative assessment result corresponding to the failure probability based on the intelligent system reliability model;

[0053] The logical relationships between nodes in the intelligent system reliability model are determined, and the importance of each functional module is determined based on the qualitative results and logical relationships, thus identifying key reliability modules.

[0054] A reliability analysis device for intelligent systems oriented towards soft errors, the device comprising:

[0055] The modeling module is used to model the various functional modules in the intelligent system divided according to the structure, and to model the intelligent system as a whole based on Bayesian networks; it is also used to establish reliability evaluation models for the functional modules.

[0056] The conversion module is used to convert the reliability models of each functional module into corresponding reliability assessment models (MREM models).

[0057] The correlation analysis module is used to analyze the failure correlation between the various functional modules and determine the failure joint function between any two modules.

[0058] The evaluation module is used to extract the relevant reliability constraints in the reliability model of the functional module, and also to extract the reliability constraints in the evaluation model of the functional module and the overall system reliability evaluation model.

[0059] The evaluation module is also used to calculate the reliability of functional modules and the reliability of the intelligent system. When evaluating the reliability of functional modules, since the transition probability matrix is ​​time-dependent, a specific moment needs to be defined. The corresponding module transition probability matrix is ​​then calculated, and the state probability distribution of the module is obtained by solving the state transition probability equation. Finally, the reliability of the functional module is calculated. Because the failure probability of the functional module used for inference calculations is based on the reliability evaluation method described above, the reliability evaluation of the intelligent system using Bayesian networks also requires defining a specific moment. The system's reliability and key reliability modules are then obtained using the inference formula of the Bayesian network.

[0060] Furthermore, the modeling module is specifically used for:

[0061] Determine the reliability modeling elements for functional modules;

[0062] Establish reliability models for functional modules and corresponding reliability assessment models.

[0063] The conversion module is specifically used for:

[0064] The reliability model of the aforementioned functional modules is analyzed;

[0065] Based on the aforementioned transformation rules, the reliability constraints required for the MREM evaluation model are extracted from the reliability model of the parsed functional modules.

[0066] The correlation analysis module is specifically used for:

[0067] The failure correlations between each pair of the aforementioned functional modules are analyzed;

[0068] Determine the joint failure function between each pair of modules;

[0069] Furthermore, the evaluation module is specifically used for:

[0070] The reliability of the functional module is evaluated based on the reliability evaluation model of the functional module.

[0071] The reliability assessment model for the aforementioned functional modules is constructed through the following steps;

[0072] Step 6.1: Determine the modules included in the functional module and determine their failure rate;

[0073] Step 6.2, establish the reliability model of the functional module, represented by triples:

[0074] FMRM = (FRate, States, STR)

[0075] Wherein, FRate represents the failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state and recovery state; STR represents all state transition relationships in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state.

[0076] Step 6.3: The basic structure of the functional module reliability assessment model is determined by the following quadruple formula:

[0077] MREM=(S,S in ,A,t)

[0078] Where S represents the state space of MREM, which is the set of all possible states of the functional modules, S = {s1, s2, ..., s} n}, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability ∈ S; t represents the moment when the reliability is evaluated;

[0079] Step 6.4: Analyze and compare the elements contained in the FMRM model and the MREM model, and establish the equivalent mapping rules for the same elements in the two models.

[0080] Step 6.5: Evaluate the reliability of the functional modules using the state transition equations;

[0081] The elements included in the state transition equation—the probability vectors of each state the functional module is in at a given moment and the probability matrix of possible state transitions for the functional module—are determined by the following formula:

[0082] P(t')=P(t)*A

[0083] Where P(t) represents the state probability vector of the functional module at time t, P(t') = (P N (t'),P R (t'),P F(t')) represents the state probability vector of the functional module at time t', P N (t'),P R (t') and P F (t') represents the probability that the functional module is in the normal recovery state and the failure state at time t', respectively; t' represents the next time after time t, and A represents the state transition probability matrix of the functional module;

[0084] The relationship between the failure rate of a functional module and time is determined by the following formula:

[0085] λ(t) = exp(λ*t)

[0086] Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0;

[0087] Step 6.6: By solving the state transition equation, the failure probability of the functional module at time t can be obtained. The reliability of the functional module at time t is determined by the following formula:

[0088] R MIFM (t)=1-P F (t)

[0089] Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure rate of the functional module at time t.

[0090] Furthermore, the evaluation module is also used for:

[0091] Using the failure of the intelligent system to be evaluated as the central node and the failure of the functional module as the edge node, the edge node is connected to the central node through connecting arcs to construct a reliability model of the intelligent system.

[0092] Obtain the failure probability of the functional module, and determine the qualitative assessment result corresponding to the failure probability based on the intelligent system reliability model;

[0093] The logical relationships between nodes in the intelligent system reliability model are determined, and the importance of each functional module is determined based on the qualitative results and logical relationships, thus identifying key reliability modules.

[0094] Compared with existing technologies, the significant advantages of this invention are: by dividing the system into different functional modules according to its functions and adopting a hierarchical approach, the reliability of the intelligent system is analyzed and evaluated. The reliability analysis method proposed in this invention has a clear hierarchical model, is scalable, and can perform both qualitative and quantitative analysis, thereby improving the accuracy of reliability analysis and evaluation.

[0095] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating a reliability analysis method for intelligent systems oriented towards soft errors in one embodiment.

[0097] Figure 2 Figure (a) is a typical state transition diagram of a functional module, and Figure (b) is a state transition diagram without considering instantaneous states.

[0098] Figure 3 This is a schematic diagram of an overall reliability analysis model structure provided in one embodiment.

[0099] Figure 4 This is a schematic diagram of the structure of an intelligent system reliability analysis device for soft errors provided in one embodiment. Detailed Implementation

[0100] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0101] To address the shortcomings of existing reliability analysis and assessment techniques, such as their inability to describe the dynamic characteristics of system hardware and their poor assessment accuracy, this invention provides a method and apparatus for intelligent system reliability analysis oriented towards soft errors. In the solution provided by this invention, the established functional module reliability model and assessment method fully consider the correlation between system hardware reliability and time, and the model can be dynamically expanded. Furthermore, the overall intelligent system reliability assessment model established based on Bayesian networks can perform not only qualitative analysis but also quantitative analysis, thus improving the accuracy of system reliability analysis and assessment.

[0102] The following is a combination of... Figure 1 The possible system architectures applicable to the embodiments of the present invention are described.

[0103] refer to Figure 1 This illustration shows a flowchart of a reliability analysis method for intelligent systems oriented towards soft errors, applicable to embodiments of the present invention. The method provided by this embodiment includes the following steps:

[0104] Step S101: Determine the typical architecture of the intelligent system and divide it into different functional modules.

[0105] Specifically, before conducting reliability analysis and evaluation of an intelligent system, it is necessary to first clarify its hardware composition. The hardware subsystem of any system consists of a processor and various functional modules. Each functional module is composed of different constraints depending on its function. In the solution proposed in this embodiment of the invention, the intelligent system is divided into five functional modules based on different functions, including:

[0106] CPU (Central Processing Unit), Register (REG), Memory (MEM), Cache, I / O (Input and Output), and Bus.

[0107] Step S102: Establish corresponding reliability models for different functional modules.

[0108] The reliability model of the functional module provided in this embodiment of the invention is determined using the following method:

[0109] Mobile intelligence is divided into different functional modules according to its functions. A hierarchical modeling approach is adopted. First, the reliability of the functional modules is modeled. By extracting their reliability constraints, a corresponding reliability model is established.

[0110] It should be noted that the reliability model of the functional modules is based on the Z language.

[0111] Step 1, the reliability model (FMRM) of the functional module can be determined by the following formula:

[0112] FMRM = (FRate, States, STR)

[0113] In the formula, FRate (Failure Rate) represents the failure rate of the functional module. The FRate of each functional module can be obtained by the calculation method of the inherent failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state and recovery state; STR (State Transfer Relationship) represents all state transition relationships that exist in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state.

[0114] Step 2: Describe the modeling elements in the model.

[0115] The failure rate FRate in the FMRM model is described as follows:

[0116] A functional module is typically composed of different modules. Taking the CPU as an example, a CPU can be divided into modules such as the Prefetch Unit (PFU), Memory Protection Unit (MPU), Load / Store Unit (LSU), and Data Processing Unit (DPU). The failure rates of these modules are defined as λ. PFU , λ MPU , λ LSU , λ DPU So, what is the CPU failure rate λ? CPU It can be represented as:

[0117] λ CPU =1-(1-λ) PFU )*(1-λ MPU )*(1-λ LSU )*(1-λ DPU )

[0118] The state space in the FMRM model is described as follows:

[0119] Each module's state space includes a normal state (NS), a recovery state (RS), an error state (ES), and a failure state (FS). NS is the initial state; ES is a transient state, representing the transition state when NS fails. A typical state transition diagram for a functional module is shown below. Figure 2 As shown, the ellipse represents the state, and the connecting line represents the state transition relationship and its transition probability.

[0120] The state transition relation STR in the FMRM model is described as follows:

[0121] The state transition relationship of the functional modules is also changed. Figure 2 This means that the state transition includes the source state sState, the target state tState, and the transition probability parameter TR. Among them, sState and tState must be contained in the state space of the Mobile Intelligent System Functional Module (MIFM).

[0122] Step S103: According to the conversion rules, the established functional module reliability model is converted into the corresponding reliability assessment model MREM to obtain the reliability data of each functional module.

[0123] In this embodiment of the invention, the MREM reliability assessment model is established using the following steps:

[0124] Step 1: Determine the MREM reliability assessment model. Specifically, the MREM reliability assessment model is constructed using the method shown in the following formula:

[0125] MREM=(S,S in ,A,t)

[0126] In the formula, S represents the state space of MREM, which is the set of all possible states of the functional module, S = {s1, s2, ..., s}. n}, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability of ∈S; t represents the moment when the reliability is evaluated.

[0127] Step 2: Based on the transformation rules, the reliability model of the functional module is converted into the corresponding MREM evaluation model. Specifically, to ensure the equivalence of the entire transformation process—that is, to maintain consistency in the description of reliability before and after the transformation—it is necessary to analyze and compare the elements contained in both (MREM model and FMRM model) and establish equivalent mapping rules for the common elements in both. The mapping transformation rules are shown in Table 1:

[0128] Table 1 Element mapping rules between FMRM and MREM

[0129]

[0130] Step 3: Based on the established MREM evaluation model, the reliability of the functional modules is evaluated using the state transition equation.

[0131] In this embodiment of the invention, the state transition equation is determined using the method shown in the following equation:

[0132] P(t')=P(t)*A

[0133] In the formula, P(t')=(P N (t'),P R (t'),P F (t')) represents the state probability vector of the functional module at time t', P N (t'), P R (t') and P F(t') represents the probability that the functional module is in the normal state (NS), recovery state (RS), and failure state (FS) at time t', respectively; t' represents the next time step after time t, and A represents the state transition probability matrix of the functional module.

[0134] Because the exponential model has a high degree of fit to the true failure rate curve, it has better advantages compared to other models describing failure rates. Therefore, this invention defines the relationship between the failure rate of a functional module and time as follows:

[0135] λ(t) = exp(λ*t)

[0136] Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0.

[0137] Therefore, based on the state transition diagram and the corresponding transition probabilities, the state transition matrix A can be determined as shown in the following equation:

[0138]

[0139] In the formula, λ Detected λ represents the probability that the fault is detected. Recovery This indicates the probability of a successful repair.

[0140] Therefore, the probability transition equation for the functional module can be determined from the three formulas in step 3:

[0141]

[0142] Solving this equation yields the state probability distribution of the functional module at time t, where the failure probability at time t is P. F Therefore, the reliability of the functional module at time t is given by the following formula:

[0143] R MIFM (t)=1-P F (t)

[0144] Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure probability of the functional module at time t.

[0145] Step S104: Based on the failure rates of each functional module, estimate the failure distribution function of the module and the joint function describing their correlation.

[0146] Specifically, the relationship function of the failure rate of the functional module over time is used as the failure distribution function of the functional module, and the joint function describing the correlation between failure rates of modules is used as the correlation joint function between modules.

[0147] The joint function describing the correlation of failure rates between modules can be determined by the following steps:

[0148] Step 1: Determine the parameters in the joint function to be selected. The MH (Metropolis Hastings) algorithm is used to sample the parameters in the joint function. A normal distribution is used as the proposed density function, and sampling is performed from the proposed density function. Based on the Markov chain, a sequence α can be generated. (0) →α (1) →...→α (n) Initialize the parameters in the joint function, given an initial value α for each parameter. (0) And let u = α (0) Based on the prior information of the parameters, a sample α is drawn from the proposed density function. (1) ~μ(N,σ 2 The acceptance probability r of the MH algorithm is calculated based on the posterior distribution, and r can be determined by the method shown in equation (9):

[0149]

[0150] In the formula, p(·|·) is the posterior distribution, and J(·|·) is the proposal density distribution; if r > u, then α is accepted. (1) Otherwise, α (1) =α (0) The extracted parameter values ​​are used as new initial values ​​for sampling. The sampling is then performed iteratively according to the parameter estimation method described above, converging from the failure rate distribution to the target distribution until the Markov chain reaches stationarity.

[0151] Step 2: Determine the joint function describing the failure correlation between modules. The joint function is selected based on the magnitude of the squared Euclidean distance; the smaller the squared Euclidean distance, the more accurately the selected joint function describes the failure correlation between modules. The squared Euclidean distance is described by the difference between the empirical distribution function and the estimated distribution function of the joint function. The empirical distribution function is determined by the following method:

[0152]

[0153] In the formula, I is the indicator function, that is, when F1(x) i When ≤u), I[F1(x i ≤u)]=1, otherwise I[F1(x i≤u)]=0; the square Euclidean distance can be determined by the following method:

[0154]

[0155] In the formula, C(u) i ,v i ) represents the value of the joint distribution function. This represents the value of the empirical distribution function.

[0156] Step S105: Based on the obtained reliability probabilities of each functional module, establish an overall reliability assessment model based on a Bayesian network to obtain the overall reliability probability of the system.

[0157] Specifically, there are many ways to build an overall reliability model of an intelligent system based on Bayesian networks. The following is an example of a better approach.

[0158] According to the overall reliability model of the intelligent system, it includes:

[0159] By taking the failure of the intelligent system to be evaluated as the central node and the failure of the functional modules as the edge nodes, and connecting the edge nodes with the central node through connecting arcs, the reliability model of the intelligent system can be obtained.

[0160] See Figure 3 This is a schematic diagram of a reliability model structure provided in an embodiment of the present invention. Figure 3 In the model shown, the functional modules affecting the reliability of the intelligent system include: CPU, registers, memory, cache, I / O, and bus. CPU-related reliability states include CPU failure and CPU normal operation; register-related reliability states include register failure and register normal operation; memory-related reliability states include memory failure and memory normal operation; cache-related reliability states include cache failure and cache normal operation; I / O-related reliability states include I / O failure and I / O normal operation; and bus-related reliability states include bus failure and bus normal operation. Using the failure of the intelligent system as the central node and the failures of the aforementioned functional modules as edge nodes, connecting the edge nodes to the central node via connecting arcs yields the following result: Figure 3 The diagram shows the reliability model structure of the intelligent system.

[0161] Furthermore, after establishing a reliability assessment model for the intelligent system, the overall reliability of the system is assessed based on the failure probability of the functional modules and the logical relationships between nodes.

[0162] In the solution provided in this embodiment of the invention, MIR-BN = (I,E,Θ) is used for description. Here, node i∈I represents a functional module variable; edge e∈E represents the failure probability dependency between the functional module and the whole; and Θ represents the parameters of the conditional probability table.

[0163] Assume M = {m1, m2, ..., m} n} represents a functional module variable, parent(m i ) represents node m i Given the set of all parent node variables, the joint probability distribution of the Bayesian network model can be written as follows:

[0164]

[0165] In the formula, P(m1,m2,...,m n P(m) represents the joint probability of the Bayesian network model; i parent(m i )) represents node m i The conditional probability of the set of all corresponding parent node variables.

[0166] Furthermore, the overall reliability assessment model of the intelligent system mainly uses the expression of the joint probability distribution described in equation (12) above for probabilistic reasoning.

[0167] Assume that the failure probabilities of each functional module are λ. CPU ,λ REG ,λ MEM ,λ Cache ,λ IO ,λ BUS Therefore, the reliability of the intelligent system can be determined according to the following formula:

[0168]

[0169] Step S106: Based on the established overall reliability assessment model, the key reliability modules of the system are obtained through reverse reasoning.

[0170] A reliability-critical module refers to a functional module whose reliability has the greatest impact on the overall system; its failure directly affects the failure of the entire system. Furthermore, improving the reliability of this functional module has the most significant impact on the overall system reliability.

[0171] Specifically, assuming a functional module has two states: failed and normal, the contribution of each functional module to the overall system failure in the failed state can be determined by the following formula:

[0172] I Pr =P(MIS=1|xi =1)-P(MIS=1|x i =0),i∈{CPU,REG,MEM,Cache,IO,BUS}

[0173] In the formula, I Pr This represents the probability importance, i.e., the degree of impact of the functional module on the overall system failure under a failure state; P(MIS=1|x i =1) represents the probability of system failure under the condition that this functional module fails; P(MIS=1|x i =0) indicates the probability of system failure under normal conditions of this module.

[0174] Furthermore, when the probability importance differs very little, the importance of functional modules can be further analyzed through key importance.

[0175] Specifically, key importance can be determined using the following formula:

[0176] I Cr (x i )=P(x i =1)[P(MIS=1|x i =1)-P(MIS=1|x i =0)] / P(MIS=1)

[0177] =P(x i =1) / P(MIS=1)×I Pr (x i ),i∈{CPU,REG,MEM,Cache,IO,BUS}

[0178] In the formula, I Cr Criticality refers to the rate of change in the system failure rate caused by the rate of change in the probability of failure of a functional module. It is a key indicator for judging the impact of a certain functional module on the overall system; P(x i =1) represents the probability that the functional module is in a failed state; P(MIS=1) represents the probability that the system is in a failed state.

[0179] Among them, I Pr and I Cr The modules that appear at the top of the list are the key modules for the reliability of the intelligent system.

[0180] The solution provided in this invention divides the system into different functional modules according to its functions, and adopts a hierarchical approach to model and evaluate the reliability of the intelligent system. The reliability analysis method provided by this invention has a clear model hierarchy, is scalable, and can perform both qualitative and quantitative analysis, thereby improving the accuracy of reliability analysis and evaluation.

[0181] Figure 4 This illustration shows a schematic diagram of a method and apparatus for soft-fault-oriented intelligent system reliability analysis according to an embodiment of the present invention. The apparatus has the function of implementing the aforementioned method for soft-fault-oriented intelligent system reliability analysis and evaluation. This function can be implemented in hardware or by hardware executing corresponding software. The apparatus may include: a modeling module 401, a conversion module 402, a correlation analysis module 403, and an evaluation module 404.

[0182] Modeling module 401 is used to construct reliability models for each functional module in an intelligent system divided according to its structure.

[0183] The modeling module is also used to establish a reliability assessment model for the functional modules;

[0184] The conversion module 402 is used to convert the reliability models of each functional module into corresponding MREM reliability assessment models.

[0185] The correlation analysis module 403 is used to analyze the failure correlation between the various functional modules and determine the failure joint function between each pair of modules;

[0186] The evaluation module 404 is used to extract the relevant reliability constraints in the reliability model of the functional module, and also to extract the reliability constraints in the evaluation model of the functional module and the overall system reliability evaluation model.

[0187] Evaluation module 404 is also used to calculate the reliability of functional modules and the reliability of the intelligent system. When evaluating the reliability of functional modules, since the transition probability matrix is ​​time-dependent, it is necessary to define a specific moment and then calculate the corresponding module transition probability matrix. After that, the state probability distribution of the module is obtained by solving the state transition probability equation, and then the reliability of the functional module is calculated. Since the failure probability of the functional module used for inference calculation is obtained based on the above-mentioned reliability evaluation method of functional modules, it is also necessary to define a specific moment when using Bayesian networks to evaluate the reliability of the intelligent system. Then, the reliability of the system and the key reliability modules are obtained by using the inference formula of Bayesian networks.

[0188] Optionally, modeling module 401 is specifically used for:

[0189] Determine the reliability modeling elements for functional modules;

[0190] Establish reliability models for functional modules and corresponding reliability assessment models.

[0191] Optionally, the conversion module 402 is specifically used for:

[0192] The reliability model of the aforementioned functional modules is analyzed;

[0193] Based on the aforementioned transformation rules, the reliability constraints required for the MREM evaluation model are extracted from the reliability model of the parsed functional modules.

[0194] Optionally, the correlation analysis module 403 is specifically used for:

[0195] The failure correlations between each pair of the functional modules are analyzed;

[0196] Determine the joint failure function between the two modules;

[0197] Optionally, evaluation module 404 is specifically used for:

[0198] The reliability of the functional modules is evaluated based on the reliability evaluation model of the functional modules described above.

[0199] The reliability of the intelligent system is assessed based on the aforementioned intelligent system reliability assessment model.

[0200] Optionally, the reliability assessment model for the functional module is constructed using the following steps:

[0201] Step 1: Determine the modules included in the functional module and their failure rates. Taking the CPU as an example, the CPU can be divided into modules such as the Prefetch Unit (PFU), Memory Protection Unit (MPU), Load / Store Unit (LSU), and Data Processing Unit (DPU). The inherent failure rate of the functional module can be determined by the following formula:

[0202] λ CPU =1-(1-λ) PFU )*(1-λ MPU )*(1-λ LSU )*(1-λ DPU )

[0203] Where, λ PFU , λ MPU , λ LSU , λDPU These represent the inherent failure rates of the prefetch unit, memory protection unit, load storage unit, and data processing unit, respectively.

[0204] Step 2, the reliability model of the functional module is represented by triples:

[0205] FMRM = (FRate, States, STR)

[0206] Here, FRate (Failure Rate) represents the failure rate of a functional module, and the FRate of each functional module can be obtained by the calculation method of the inherent failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state, and recovery state; STR (State Transfer Relationship) represents all state transition relationships that exist in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state.

[0207] Step 3, the basic structure of the functional module MREM reliability assessment model can be determined by the quadruple shown in equation (18):

[0208] MREM=(S,S in ,A,t)

[0209] Where S represents the state space of MREM, which is the set of all possible states of the functional modules, S = {s1, s2, ..., s} n}, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability of ∈S; t represents the moment when the reliability is evaluated.

[0210] Based on the established functional module reliability model, the corresponding functional module reliability model FMRM is converted into the corresponding MREM evaluation model according to the conversion rules.

[0211] Step 4: Evaluate the reliability of the functional modules based on the state transition equations described above;

[0212] The state transition equation includes the probability vectors of each state of the functional module at a certain moment and the possible state transition probability matrix of the functional module, which can be determined by the method shown in the following formula:

[0213] P(t')=P(t)*A

[0214] Where, P(t')=(P N (t'),P R (t'),P F (t')) represents the state probability vector of the functional module at time t', t' represents the next time step after time t, and A represents the state transition probability matrix of the functional module;

[0215] The relationship between the failure rate of a functional module and time can be determined by the following formula:

[0216] λ(t) = exp(λ*t)

[0217] Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0.

[0218] By solving the state transition probability equation, the failure probability of the functional module at time t can be obtained. Therefore, the reliability of the functional module at time t can be determined by the following formula:

[0219] R MIFM (t)=1-P F (t)

[0220] Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure probability of the functional module at time t.

[0221] Optionally, the reliability assessment model for intelligent systems is determined using the following methods:

[0222] By taking the failure of the intelligent system to be evaluated as the central node and the failure of the functional modules as the edge nodes, and connecting the edge nodes with the central node through connecting arcs, the reliability model of the intelligent system can be obtained.

[0223] Obtain the failure probability of the functional module, and determine the qualitative assessment result corresponding to the failure probability based on the intelligent system reliability model;

[0224] The logical relationships between nodes in the intelligent system reliability model are determined, and the importance of each functional module is determined based on the qualitative results and logical relationships, thus identifying key reliability modules.

[0225] The solution provided in this invention analyzes and evaluates the reliability of the intelligent system by dividing the system into different functional modules according to its functions and adopting a hierarchical approach. The reliability analysis method proposed in this invention has a clear hierarchical model, is scalable, and can perform both qualitative and quantitative analysis, thereby improving the accuracy of reliability analysis and evaluation.

[0226] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0227] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the service building apparatus and service loading apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0228] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention.

Claims

1. A reliability analysis method for intelligent systems oriented towards soft errors, characterized in that, The method includes: For the intelligent system to be analyzed and evaluated, it is divided into different functional modules (MIFM) according to its structure; the functional modules include CPU, registers, memory, cache, high-speed cache, I / O and bus; Establish corresponding models for the functional modules to obtain reliability models (FMRM) for different functional modules; The reliability models of each functional module are converted into corresponding reliability assessment models (MREMs) to obtain the reliability assessment results of different functional modules. Based on the failure rate of each functional module, the failure correlation between modules is analyzed, and the joint function describing the failure rate correlation between modules is estimated. Based on the failure rate and correlation joint function of each functional module, a system overall reliability assessment model MIR-BN is established based on Bayesian network to obtain the overall system reliability assessment result. Then, based on the system overall reliability assessment model, the key reliability modules of the system are obtained by back reasoning. Establish corresponding models for the functional modules to obtain reliability models for different functional modules, using the following method: Determine the underlying architecture of each functional module and divide the functional modules into different functional components; Determine the parameters in the reliability model, including the failure rate of the functional modules, the state space of the functional modules, and all the state transition relationships of the functional modules; The reliability model of a functional module is constructed through the following steps; Step 2.1: Determine the modules included in the functional module and determine their failure rate; Step 2.2, establish the reliability model of the functional module, represented by triples: FMRM = (FRate, States, STR) Wherein, FRate represents the failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state and recovery state; STR represents all state transition relationships in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state. The reliability assessment model for the functional modules is determined in the following manner: Based on the established functional module reliability model, the corresponding FMRM model is converted into the corresponding evaluation model according to the conversion rules. Step 3.1: The basic structure of the functional module reliability assessment model is determined by the following quadruple formula: MREM=(S,S in ,A,t) Where S represents the state space of MREM, which is the set of all possible states of the functional modules, S = {s1, s2, ..., s} n }, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability ∈ S; t represents the moment when the reliability is evaluated; Step 3.2: Analyze and compare the elements contained in the FMRM model and the MREM model, and establish the equivalent mapping rules for the same elements in the two models.

2. The reliability analysis method for intelligent systems oriented towards soft errors according to claim 1, characterized in that, After converting the reliability models of each functional module into corresponding reliability assessment models (MREMs), the following is also included: Step 4.1: Evaluate the reliability of the functional modules using state transition equations; The elements included in the state transition equation—the probability vectors of each state the functional module is in at a given moment and the probability matrix of possible state transitions for the functional module—are determined by the following formula: P(t')=P(t)*A Where P(t) represents the state probability vector of the functional module at time t, P(t') = (P N (t'),P R (t'),P F (t')) represents the state probability vector of the functional module at time t', P N (t'),P R (t') and P F (t') represents the probability that the functional module is in the normal recovery state and the failure state at time t', respectively; t' represents the next time after time t, and A represents the state transition probability matrix of the functional module; The relationship between the failure rate of a functional module and time is determined by the following formula: λ(t) = exp(λ*t) Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0; Step 4.2: By solving the state transition equation, the failure probability of the functional module at time t can be obtained. The reliability of the functional module at time t is determined by the following formula: R MIFM (t)=1-P F (t) Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure rate of the functional module at time t.

3. The reliability analysis method for intelligent systems oriented towards soft errors according to claim 2, characterized in that, The joint function describing the correlation of failure rates between modules is determined by the following method: The relationship function of the failure rate of the functional module over time is used as the failure distribution function of the functional module, and the joint function describing the correlation of failure rates between modules is used as the correlation joint function between modules. The joint function describing the correlation of failure rates between modules is determined in the following way: Step 5.1: Determine the candidate joint functions, which are Copula joint functions, including the Frank function, Clayton function, and Gumbel function; Step 5.2: Determine the parameters in the joint function to be selected, and sample the parameters in the joint function using the MH algorithm; use the normal distribution as the proposal density function, sample from the proposal density function, and generate a sequence α according to the Markov chain. (0) →α (1) →...→α (n) Initialize the parameters in the joint function, given an initial value α for each parameter. (0) And let u = α (0) Based on the prior information of the parameters, a sample α is drawn from the proposed density function. (1) ~μ(N,σ 2 The acceptance probability r of the MH algorithm is calculated based on the posterior distribution, and r is determined by the following formula: Where p(·|·) is the posterior distribution, and J(·|·) is the proposal density distribution; if r > u, then accept α. (1) Otherwise, α (1) =α (0) The extracted parameter values ​​are used as new initial values ​​for sampling. The sampling is then performed iteratively according to the parameter estimation method described above until the loss rate distribution converges to the target distribution. Sampling stops when the Markov chain reaches stationarity. Step 5.3: Determine the joint function describing the failure correlation between modules; select the joint function by measuring the squared Euclidean distance. The smaller the squared Euclidean distance, the more accurately the selected joint function describes the failure correlation between modules.

4. The reliability analysis method for intelligent systems oriented towards soft errors according to claim 3, characterized in that, The joint function mentioned in step 5.1 is the Copula joint function, which includes the Frank function, Clayton function, and Gumbel function.

5. The reliability analysis method for intelligent systems oriented towards soft errors according to claim 1, characterized in that, The overall system reliability assessment model is determined by the following method: Using the failure of the intelligent system to be evaluated as the central node and the failure of the functional module as the edge node, the edge node is connected to the central node through a connecting arc, thus obtaining the reliability model of the intelligent system. Furthermore, a reliability assessment of the system is performed based on the aforementioned intelligent system reliability model, including: Obtain the failure probability of the functional module, and determine the qualitative assessment result corresponding to the failure probability based on the intelligent system reliability model; The logical relationships between nodes in the intelligent system reliability model are determined, and the importance of each functional module is determined based on the qualitative evaluation results and logical relationships, thereby identifying key reliability modules.

6. A reliability analysis device for intelligent systems oriented towards soft errors based on the method of any one of claims 1 to 5, characterized in that, The device includes: The modeling module is used to model the various functional modules in the intelligent system divided according to the structure, and to model the intelligent system as a whole based on Bayesian networks; it is also used to establish reliability evaluation models for the functional modules. The conversion module is used to convert the reliability models of each functional module into corresponding reliability assessment models (MREM models). The correlation analysis module is used to analyze the failure correlation between the various functional modules and determine the failure joint function between any two modules. The evaluation module is used to extract relevant reliability constraints from the reliability model of the functional module, as well as from the evaluation model of the functional module and the overall system reliability evaluation model; it is also used to calculate the reliability of the functional module and the reliability of the intelligent system.

7. The apparatus according to claim 6, characterized in that, The evaluation module is specifically used for: The reliability of the functional module is evaluated based on the reliability evaluation model of the functional module. The reliability assessment model for the aforementioned functional modules is constructed through the following steps; Step 6.1: Determine the modules included in the functional module and determine their failure rate; Step 6.2, establish the reliability model of the functional module, represented by triples: FMRM = (FRate, States, STR) Wherein, FRate represents the failure rate of the functional module; States represents the state space of the functional module, which contains all possible states of the functional module, including but not limited to normal state, failure state, fault state and recovery state; STR represents all state transition relationships in the functional module, including but not limited to the relationship from normal state to failure state, from normal state to fault state, and from fault state to recovery state. Step 6.3: The basic structure of the functional module reliability assessment model is determined by the following quadruple formula: MREM=(S,S in ,A,t) Where S represents the state space of MREM, which is the set of all possible states of the functional modules, S = {s1, s2, ..., s} n }, s n S represents the nth state of MREM. in ∈S represents the initial state of the MREM; A = [a ij ] represents the state transition probability matrix of MREM, where a ij Indicates that MREM is from state s i ∈S transitions to state s j The probability ∈ S; t represents the moment when the reliability is evaluated; Step 6.4: Analyze and compare the elements contained in the FMRM model and the MREM model, and establish the equivalent mapping rules for the same elements in the two models. Step 6.5: Evaluate the reliability of the functional modules using the state transition equations; The elements included in the state transition equation—the probability vectors of each state the functional module is in at a given moment and the probability matrix of possible state transitions for the functional module—are determined by the following formula: P(t')=P(t)*A Where P(t) represents the state probability vector of the functional module at time t, P(t') = (P N (t'),P R (t'),P F (t')) represents the state probability vector of the functional module at time t', P N (t'),P R (t') and P F (t') represents the probability that the functional module is in the normal recovery state and the failure state at time t', respectively; t' represents the next time after time t, and A represents the state transition probability matrix of the functional module; The relationship between the failure rate of a functional module and time is determined by the following formula: λ(t)==exp(λ*t) Where λ(t) represents the failure rate of the functional module at time t, and λ represents the inherent failure rate of the functional module when t = 0; Step 6.6: By solving the state transition equation, the failure probability of the functional module at time t can be obtained. The reliability of the functional module at time t is determined by the following formula: R MIFM (t)=1-P F (t) Among them, R MIFM (t) represents the reliability of the functional module at time t, P F (t) represents the failure rate of the functional module at time t.

8. The apparatus according to claim 7, characterized in that, The evaluation module is also used for: Using the failure of the intelligent system to be evaluated as the central node and the failure of the functional module as the edge node, the edge node is connected to the central node through connecting arcs to construct a reliability model of the intelligent system. Obtain the failure probability of the functional module, and determine the qualitative assessment result corresponding to the failure probability based on the intelligent system reliability model; The logical relationships between nodes in the intelligent system reliability model are determined, and the importance of each functional module is determined based on the qualitative evaluation results and logical relationships, thus identifying key reliability modules.

Images