Improved FMEA device considering uncertainty and failure mode correlation

Abstract

The invention provides an improved FMEA device considering uncertainty and failure mode correlation, and the device is characterized in that the device comprises an SVTraNN score conversion module which is used for obtaining an SVTraNN evaluation matrix; the expert weight calculation module is used for calculating to obtain an expert weight; the expert evaluation summarization module is used for calculating to obtain an SVTraNN decision matrix; the subjective weight calculation module is used for calculating the final subjective weight of each risk factor; the objective weight calculation module is used for calculating the objective weight of each risk factor; the comprehensive weight calculation module is used for calculating the comprehensive weight of each risk factor; the comprehensive score calculation module is used for calculating a comprehensive score of each risk mode; the risk sorting module is used for obtaining a risk sorting result of each risk mode; and the risk level module is used for obtaining the risk level of each risk mode. In a word, the method can effectively evaluate the risk degree of the risk mode.

Description

Technical Field

[0001] The invention belongs to the field of reliability and risk analysis methods, and in particular relates to an improved FMEA device which takes into account the relationship between uncertainty and failure mode. Background Art

[0002] Failure Mode and Effects Analysis (FMEA) is a forward-looking reliability and risk analysis technology. Before failure occurs, it analyzes the causes and effects of each failure mode to identify the weak and critical links of the system, so as to help decision makers set priorities and preventive measures for risk management activities. Today, FMEA has been widely used in reliability and risk management research in aerospace, equipment maintenance, medical and other fields.

[0003] The traditional FMEA method uses discrete values ​​of 1-10 to evaluate the severity S, occurrence O, and detection D of the failure mode, and then multiplies the three to obtain the risk priority number RiskPriorityNumber, or RPN, which is used to measure the risk level of each failure mode. By comparing RPN, failure modes can be prioritized to assist risk management decisions. However, the completeness and robustness of the traditional FMEA method in the process of use have been questioned by many scholars, which can be summarized into four points: (1) The evaluation information of experts on failure modes is fuzzy and uncertain, and it is difficult to express it through discrete values; (2) The weights of evaluation experts and risk factors are not considered, and the differences between experts and the different effects of risk factors on the research objects are not fully considered; (3) The calculation of RPN lacks robustness and scientificity. RPN is sensitive to changes in risk factor evaluation, and different risk factor evaluation combinations may have the same RPN value but different meanings; (4) The relationship between failure modes is not considered. Therefore, the traditional FMEA method still has a lot of room for improvement. Summary of the invention

[0004] The present invention is made to solve the above problems, and aims to provide an improved FMEA device which takes into account the relationship between uncertainty and failure mode.

[0005] The present invention provides an improved FMEA device that considers the relationship between uncertainty and failure mode, which is used to obtain the risk ranking results and risk levels of each artificial intelligence safety potential risk mode according to the scoring data of each expert on each artificial intelligence safety potential risk mode and the scoring data of the importance of risk factors, and has the following characteristics, including: an SVTraNN scoring conversion module, which stores a preset SVTraNN language variable conversion table, and is used to obtain the SVTraNN evaluation matrix corresponding to each expert according to the SVTraNN language variable conversion table and the scoring data; an expert weight calculation module, which is used to calculate the expert weight of each expert on the risk factor according to the SVTraNNs score corresponding to the risk factor in all SVTraNNs evaluation matrices; an expert evaluation summary module, which is used to calculate the SVTraNN decision matrix according to all SVTraNNs evaluation matrices and expert weights; a subjective weight calculation module, which stores the SVTraNN language variable conversion table and the SVTraN-FUCOM planning model, and is used to calculate the expert weight of each expert on the risk factor according to the SVTraNNs score corresponding to the risk factor in all SVTraNNs evaluation matrices. The language variable conversion table, SVTraN-FUCOM planning model and all risk factor importance score data are used to calculate the final subjective weights corresponding to each risk factor; the objective weight calculation module stores the difference construction target planning model, which is used to calculate the objective weights corresponding to each risk factor according to the SVTraNN decision matrix and the difference construction target planning model; the comprehensive weight calculation module is used to calculate the comprehensive weights corresponding to each risk factor according to all the final subjective weights and objective weights; the comprehensive score calculation module is used to calculate the comprehensive score of each artificial intelligence safety potential risk mode according to the SVTraN-CODAS method, the SVTraNN decision matrix and all comprehensive weights; the risk ranking module is used to obtain the risk ranking results of each artificial intelligence safety potential risk mode according to all comprehensive scores; the risk level module is used to perform clustering and stratification according to the comprehensive score, scoring data and comprehensive weight to obtain the risk level of each artificial intelligence safety potential risk mode, among which the higher the risk ranking result, the higher the risk of the corresponding artificial intelligence safety potential risk mode.

[0006] In the improved FMEA device that takes into account the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: wherein, the expert weight calculation module includes: an average similarity calculation unit, which is used to calculate the average similarity corresponding to each risk factor according to the SVTraNN evaluation matrix of each expert; a confidence level calculation unit, which is used to calculate the confidence level corresponding to each expert for each risk factor according to the average similarity between experts; an expert weight calculation unit, which is used to calculate the expert weight corresponding to each expert for each risk factor according to the confidence level of all experts.

[0007] The improved FMEA device considering the relationship between uncertainty and failure mode provided by the present invention may also have the following characteristics: wherein the average similarity AS between expert k and expert l for risk factor S is S (e k ,e l ) is calculated as: f=<(a f ,b f ,c f ,d f );T f ,I f ,F f >, g=<(a g ,b g ,c g ,d g );T g ,I g ,F g >, Where m is the total number of potential risk patterns for AI security, f and g are two SVTraNNs, is the SVTraNN corresponding to the i-th AI security potential risk pattern in the SVTraNN evaluation matrix corresponding to expert k, is the SVTraNN corresponding to the i-th AI security potential risk mode in the SVTraNN evaluation matrix corresponding to expert l, x fT 、x gT ,y fT ,y gT Corresponding to the true membership, x fI 、x gI ,y fI ,y gI Corresponding to the uncertain membership, x fF 、x gF ,y fF ,y gF Corresponding to the pseudo-membership, the confidence level CL of expert k under risk factor S S (e k ) is calculated as: Where t is the total number of experts, and the expert weight ws corresponding to expert k under risk factor S is k The calculation expression is:

[0008] The improved FMEA device considering the relationship between uncertainty and failure mode provided by the present invention may also have the following features: wherein the SVTraNN evaluation matrix H corresponding to the expert k is k The expression is: The first, second, and third columns correspond to the scores of risk factors S, O, and D, respectively. The expression of the SVTraNN decision matrix is: Where i = 1,…,m, m is the total number of potential risk modes of AI security, and are the scores of the i-th AI security potential risk model in the SVTraNN evaluation matrix corresponding to the k-th expert under risk factors S, O, and D, respectively, and ws k ,wo k and wd k are the expert weights corresponding to the kth expert under risk factors S, O, and D respectively.

[0009] In the improved FMEA device that considers the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: wherein, the subjective weight calculation module includes: an SVTraNN conversion unit, which is used to convert the risk factor importance score data according to the SVTraNN language variable conversion table to obtain the SVTraNN data corresponding to each expert; a subjective weight solving unit, which is used to solve the SVTraN-FUCOM planning model through the SVTraN-FUCOM planning model according to the SVTraNN data corresponding to each expert, and obtain the subjective weight of each expert for each risk factor; a final subjective weight calculation unit, which is used to calculate the average value of the subjective weights corresponding to all experts for each risk factor, and use the calculation result as the final subjective weight corresponding to the risk factor.

[0010] In the improved FMEA device considering the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: when solving the subjective weight of expert e1, the corresponding expression of the SVTraN-FUCOM planning model is: minχ, st

[0011]

[0012] In the formula is the SVTraNN data corresponding to the score of the i-th expert on the importance of the j-th risk factor, is the SVTraNN data corresponding to the subjective weight of the i-th expert on the j-th risk factor, is the subjective weight of the i-th expert on the j-th risk factor.

[0013] In the improved FMEA device that considers the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: wherein, the objective weight calculation module includes: a score matrix generation unit, which is used to perform SVTraNN score calculation and normalization on the SVTraNN decision matrix to obtain the SVTraNN normalized score matrix; a parameter calculation unit, which is used to calculate the standard deviation corresponding to each risk factor and the Pearson correlation coefficient between each risk factor according to the SVTraNN normalized score matrix; an objective weight solving unit, which is used to solve the difference construction target planning model according to the standard deviation and the Pearson correlation coefficient to obtain the objective weight corresponding to each risk factor.

[0014] In the improved FMEA device considering the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: wherein the calculation expression of SVTraNN score calculation is: Where <(a f ,b f ,c f ,d f );T f ,I f ,F f > is the expression corresponding to the element f in the SVTraNN decision matrix, ξ(f) is the result of the SVTraNN score calculation for the element f, and the expression of the difference construction target planning model is: minλ b +0.5λ c , st ωo j >0,j=1,2,3.where ωo j is the objective weight of the j-th risk factor, σ j is the standard deviation of all normalized scores of the jth risk factor, r jl is the Pearson correlation coefficient between the j-th risk factor and all normalized scores of the l-th risk factor.

[0015] In the improved FMEA device that considers the relationship between uncertainty and failure mode provided by the present invention, it can also have the following characteristics: wherein, the comprehensive score calculation module includes: a normalization unit, which is used to normalize the SVTraNN decision matrix by the normalization method of RAFSI to obtain the SVTraNN normalized decision matrix; a weighting unit, which is used to obtain a weighted normalized matrix according to the comprehensive weight and the SVTraNN normalized decision matrix; a negative ideal solution unit, which is used to obtain the negative ideal solution of each risk factor according to the SVTraNN normalized decision matrix and the corresponding SVTraNN score; a distance calculation unit, which is used to calculate the Euclidean distance and Manhattan distance corresponding to each artificial intelligence safety potential risk mode according to the SVTraNN normalized decision matrix and the negative ideal solution; a relevant evaluation matrix construction unit, which is used to construct a relevant evaluation matrix according to all Euclidean distances and Manhattan distances; a comprehensive score calculation unit, which is used to calculate the comprehensive score of each artificial intelligence safety potential risk mode according to the relevant evaluation matrix.

[0016] The improved FMEA device considering the relationship between uncertainty and failure mode provided by the present invention may also have the following features: wherein the expression of the weighted normalization matrix is: V m×3 =[v ij ] m×3 , Where V m×3 is the weighted normalization matrix, v ij is the weighted normalization matrix V m×3 The element in the i-th row and j-th column, ω j is the comprehensive weight corresponding to the j-th risk factor, is the Euclidean distance E corresponding to the i-th AI security potential risk mode in the i-th row and j-th column of the SVTraNN normalized decision matrix i The calculation expression is: In the formula and is the true membership degree corresponding to the i-th AI safety potential risk model under the j-th risk factor, and is the uncertainty membership corresponding to the i-th AI safety potential risk mode under the j-th risk factor, and is the pseudo-membership corresponding to the ith AI safety potential risk model under the jth risk factor, and is the true membership degree corresponding to the negative ideal solution of the j-th risk factor, and is the uncertain membership corresponding to the negative ideal solution of the j-th risk factor, and is the pseudo-membership degree corresponding to the negative ideal solution of the j-th risk factor, and the Manhattan distance T corresponding to the i-th AI safety potential risk pattern i The calculation expression is: The expression of the relevant evaluation matrix is: RA = [o ik ] m×m , o ik =(E i -E k )+Ψ(E i -E k )*(T i -T k ), Where τ is the preset value, and the comprehensive score corresponding to the ith AI security potential risk mode is O i The calculation expression is: Where m is the total number of potential risk modes for AI security.

[0017] Functions and Effects of the Invention

[0018] According to the improved FMEA device that considers the relationship between uncertainty and failure mode involved in the present invention, first, the scoring data is converted into SVTraNN, and the uncertainty and fuzziness of the expert evaluation information are considered; second, the expert weight is calculated by the confidence level, which solves the shortcoming that the traditional FMEA does not consider the differences of experts; third, the final subjective weight of the risk factor is solved according to the SVTraN-FUCOM planning model, and the objective weight is solved according to the difference. The problem that the traditional FMEA method is highly subjective and does not consider the differences in the impact of different risk factors on the research object; fourth, the comprehensive weight of the risk factor is calculated using the objective weight and the final subjective weight, and the comprehensive score is further calculated and ranked, so as to obtain the risk ranking result, while solving the shortcomings of the traditional FMEA method RPN calculation lacking robustness and scientificity, and adding consideration to the relationship between risk modes. Therefore, the improved FMEA device that considers the relationship between uncertainty and failure mode of the present invention can effectively evaluate the risk level of risk mode. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a block diagram of an improved FMEA device in an embodiment of the present invention;

[0020] Figure 2 is a schematic diagram of the calculation results of the relevant evaluation matrix and the comprehensive score in the embodiment of the present invention;

[0021] Figure 3is a flow chart of generating risk ranking results by improving the FMEA device in an embodiment of the present invention;

[0022] Figure 4 It is a schematic diagram of the results of ISODATA clustering in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments are combined with the accompanying drawings to specifically illustrate the improved FMEA device of the present invention that takes into account the relationship between uncertainty and failure mode.

[0024] In this embodiment, an improved FMEA device that takes into account the relationship between uncertainty and failure mode is provided, hereinafter referred to as the improved FMEA device, which is used to obtain the risk ranking results and risk levels of each potential artificial intelligence safety risk mode based on the scoring data of each expert on each potential artificial intelligence safety risk mode and the scoring data of the importance of risk factors.

[0025] In this embodiment, the experts include an expert team consisting of university scholars e1, senior risk strategy analysts e2, artificial intelligence risk assessment experts e3, and senior artificial intelligence product managers e4.

[0026] There are 19 potential AI security risk modes in this embodiment, as shown in the following table:

[0027]

[0028]

[0029] The first column in the above table is the risk modules, the second column is the potential risk modes of AI safety, and the third to fifth columns are the risk causes, risk consequences and numbers corresponding to the potential risk modes of AI safety.

[0030] According to the risk factors S, O, and D, the scoring criteria are established as shown in the following table:

[0031]

[0032] The first column in the above table is the various scores, and the second to fourth columns are the corresponding descriptions of risk factor S, i.e. severity, risk factor O, i.e. occurrence, and risk factor D, i.e. detection, under each score.

[0033] The scoring data generated by each expert for each potential risk model of AI security is shown in the following table:

[0034]

[0035]

[0036] The first column in the above table is the number of each potential risk mode of artificial intelligence security, and the second to thirteenth columns are the scores of each expert on the risk factors S, O, and D respectively.

[0037] The risk factor importance rating criteria of this embodiment are shown in the following table:

[0038] score Importance 1 Almost no impact 2 Slightly affected 3，4 A lower degree of impact 5，6 More important 7，8 Big impact 9 Play a decisive role 10 Crucial

[0039] The corresponding risk factor importance score data of each expert is shown in the following table:

[0040] Risk Factors <![CDATA[e1]]> <![CDATA[e2]]> <![CDATA[e3]]> <![CDATA[e4]]> S 9 8 9 8 O 8 7 7 6 D 6 6 5 5

[0041] Figure 1 is a block diagram of an improved FMEA device in an embodiment of the present invention.

[0042] like Figure 1 As shown, the improved FMEA device 100 includes an SVTraNN score conversion module 11, an expert weight calculation module 12, an expert evaluation summary module 13, a subjective weight calculation module 14, an objective weight calculation module 15, a comprehensive weight calculation module 16, a comprehensive score calculation module 17, a risk ranking module 18, a risk level module 19 and a control module 20 for controlling the operation of the above modules.

[0043] The SVTraNN score conversion module 11 stores a preset SVTraNN language variable conversion table, and is used to obtain the SVTraNN evaluation matrix corresponding to each expert according to the SVTraNN language variable conversion table and the score data.

[0044] Among them, the SVTraNN evaluation matrix H corresponding to expert k k The expression is:

[0045]

[0046] The first, second, and third columns in the formula correspond to the scores of risk factors S, O, and D, respectively.

[0047] The SVTraNN language variable conversion table for this implementation is as follows:

[0048] score SVTraNNs Score 1 <(0,0.05,0.05,0.1)；0.1,0.9,0.9> 0.0440 2 <(0.1,0.15,0.15,0.2)；0.1,0.8,0.75> 0.1256 3 <(0.2,0.25,0.25,0.3)；0.25,0.7,0.6> 0.2917 4 <(0.3,0.35,0.35,0.4)；0.4,0.6,0.5> 0.3789 5 <(0.4,0.45,0.45,0.5)；0.5,0.5,0.4> 0.4766 6 <(0.5,0.55,0.55,0.6)；0.6,0.4,0.3> 0.5755 7 <(0.6,0.65,0.65,0.7)；0.7,0.3,0.2> 0.6750 8 <(0.7,0.75,0.75,0.8)；0.8,0.2,0.1> 0.7748 9 <(0.8,0.85,0.85,0.9)；0.9,0.1,0.1> 0.8747 10 <(0.9,0.95,0.95,1)；1,0.1,0> 0.9746

[0049] The first column in the table above is the score, and the second column is the SVTraNNs data corresponding to the score. The third column is the score corresponding to the SVTraNNs data.

[0050] The expert weight calculation module 12 is used to calculate the expert weight of each expert for each risk factor according to the SVTraNNs score corresponding to the risk factor in all SVTraNNs evaluation matrices.

[0051] The expert weight calculation module 12 includes an average similarity calculation unit 121 , a confidence level calculation unit 122 and an expert weight calculation unit 123 .

[0052] The average similarity calculation unit 121 is used to calculate the average similarity corresponding to each risk factor according to the SVTraNN evaluation matrix of each pair of experts.

[0053] Among them, the average similarity AS between expert k and expert l on risk factor S is S (e k ,e l ) is calculated as:

[0054]

[0055] f=<(a f ,b f ,c f ,d f );T f ,I f ,F f >,

[0056] g=<(a g ,b g ,c g ,d g );T g ,I g ,F g >,

[0057]

[0058] Where m is the total number of potential risk patterns for AI security, f and g are two SVTraNNs, is the SVTraNN corresponding to the i-th AI security potential risk pattern in the SVTraNN evaluation matrix corresponding to expert k, is the SVTraNN corresponding to the i-th AI security potential risk mode in the SVTraNN evaluation matrix corresponding to expert l, x fT 、x gT ,y fT ,y gT Corresponding to the true membership, x fI 、x gI ,y fI ,y gI Corresponding to the uncertain membership, xfF 、x gF ,y fF ,y gF Corresponding pseudo-membership.

[0059] In this embodiment, x fT 、x gT ,y fT ,y gT For example, the formula is defined as:

[0060]

[0061] In this embodiment, the average similarity between experts under risk factor S is shown in the following table:

[0062]

[0063]

[0064] The confidence level calculation unit 122 is used to calculate the confidence level corresponding to each expert for each risk factor according to the average similarity between two experts.

[0065] Among them, the confidence level CL corresponding to expert k under risk factor S is S (e k ) is calculated as:

[0066]

[0067] Where t is the total number of experts.

[0068] The confidence levels of the experts under the risk factor S in this embodiment are shown in the following table:

[0069] Confidence Level <![CDATA[e1]]> <![CDATA[e2]]> <![CDATA[e3]]> <![CDATA[e4]]> <![CDATA[CL S (and k )]]> 0.8463 0.7919 0.8599 0.8532

[0070] The first row in the above table contains the experts, and the second row contains the corresponding confidence levels of the experts.

[0071] The expert weight calculation unit 123 is used to calculate the expert weight corresponding to each expert for each risk factor according to the confidence level of all experts.

[0072] Among them, the expert weight ws corresponding to expert k under risk factor S k The calculation expression is:

[0073]

[0074] The expert weights of the experts under each risk factor in this embodiment are shown in the following table:

[0075] Expert Weight <![CDATA[e1]]> <![CDATA[e2]]> <![CDATA[e3]]> <![CDATA[e4]]> <![CDATA[ws k ]]> 0.2525 0.2363 0.2566 0.2546 <![CDATA[wo k ]]> 0.2533 0.2411 0.2543 0.2513 <![CDATA[wd k ]]> 0.2402 0.2515 0.2567 0.2516

[0076] The first row in the above table is for each expert, and the second to fourth rows are the expert weights corresponding to each expert under risk factor S, risk factor O, and risk factor D respectively.

[0077] The expert evaluation summary module 13 is used to calculate the SVTraNN decision matrix according to all SVTraNNs evaluation matrices and expert weights.

[0078] Among them, the expression of SVTraNN decision matrix is:

[0079]

[0080]

[0081] Where i = 1,…,m, m is the total number of potential risk modes of AI security, and are the scores of the i-th AI security potential risk model in the SVTraNN evaluation matrix corresponding to the k-th expert under risk factors S, O, and D, respectively, and ws k ,wo k and wd k are the expert weights corresponding to the kth expert under risk factors S, O, and D respectively.

[0082] The calculation results of the SVTraNN decision matrix in this embodiment are as follows:

[0083]

[0084] The subjective weight calculation module 14 stores the SVTraNN language variable conversion table and the SVTraN-FUCOM planning model, and is used to calculate the final subjective weight corresponding to each risk factor based on the SVTraNN language variable conversion table, the SVTraN-FUCOM planning model and all risk factor importance score data.

[0085] The subjective weight calculation module 14 includes an SVTraNN conversion unit 141 , a subjective weight solution unit 142 and a final subjective weight calculation unit 143 .

[0086] The SVTraNN conversion unit 141 is used to convert the risk factor importance score data according to the SVTraNN language variable conversion table to obtain the SVTraNN data corresponding to each expert. For example, the scores of expert e1 on risk factors S, O, and D are converted into <(0.8, 0.85, 0.85, 0.9); 0.9, 0.1, 0.1>, <(0.7, 0.75, 0.75, 0.8); 0.8, 0.2, 0.1>, <(0.5, 0.55, 0.55, 0.6); 0.6, 0.4, 0.3>.

[0087] The subjective weight solving unit 142 is used to solve the SVTraN-FUCOM planning model according to the SVTraNN data corresponding to each expert through the SVTraN-FUCOM planning model, and obtain the subjective weight of each expert for each risk factor.

[0088] When solving the subjective weight of expert e1, the corresponding expression of the SVTraN-FUCOM planning model is:

[0089] min

[0090] st

[0091]

[0092]

[0093] In the formula is the SVTraNN data corresponding to the score of the i-th expert on the importance of the j-th risk factor, is the SVTraNN data corresponding to the subjective weight of the i-th expert on the j-th risk factor, is the subjective weight of the i-th expert on the j-th risk factor.

[0094] In this embodiment, the subjective weights of experts e1, e2, e3 and e4 are calculated to be {0.3916, 0.3479, 0.2605}, {0.3845, 0.3333, 0.2821}, {0.4359, 0.3333, 0.2308} and {0.4067, 0.3175, 0.2758} respectively.

[0095] The final subjective weight calculation unit 143 is used to calculate the average of the subjective weights corresponding to all experts for each risk factor, and use the calculation result as the final subjective weight corresponding to the risk factor.

[0096] The final subjective weight calculated in this embodiment is {0.4047, 0.3330, 0.2623}.

[0097] The objective weight calculation module 15 stores a difference-constructed target programming model, and is used to calculate the objective weight corresponding to each risk factor according to the SVTraNN decision matrix and the difference-constructed target programming model.

[0098] The objective weight calculation module 15 includes a score matrix generation unit 151 , a parameter calculation unit 152 and an objective weight solution unit 153 .

[0099] The score matrix generating unit 151 is used to calculate and normalize the SVTraNN score on the SVTraNN decision matrix to obtain a SVTraNN normalized score matrix.

[0100] Among them, the calculation expression for SVTraNN normalized score calculation is:

[0101]

[0102] Where <(a f ,b f ,c f ,d f );T f ,I f ,F f > is the expression corresponding to the SVTraNN element f, and ξ(f) is the result obtained by calculating the SVTraNN score of the element f.

[0103] The SVTraNN normalized score matrix calculated in this embodiment is shown in the following table:

[0104]

[0105] The first column in the above table shows the potential risk modes of various artificial intelligence security. The second to fourth columns are the normalized results of the SVTraNN scores corresponding to the potential risk modes of artificial intelligence security under risk factors S, O, and D respectively.

[0106] The parameter calculation unit 152 is used to calculate the standard deviation corresponding to each risk factor and the Pearson correlation coefficient between any two risk factors according to the SVTraNN normalized score matrix.

[0107] The objective weight solving unit 153 is used to solve the difference construction target programming model according to the standard deviation and the Pearson correlation coefficient to obtain the objective weight corresponding to each risk factor.

[0108] Among them, the expression of the difference construction target planning model is:

[0109] minλ b +0.5λ c

[0110] st

[0111]

[0112] ωo j >0,j=1,2,3.

[0113] Where ωo j is the objective weight of the j-th risk factor, σ j is the standard deviation of all normalized scores of the jth risk factor, r jl is the Pearson correlation coefficient between all normalized scores of the j-th risk factor and the l-th risk factor. j |j=1,2,3}={0.2563,0.2381,0.1892}, Pearson correlation coefficient r 12 =0.3569, r 13 =0.7691, r 23 =0.2639.

[0114] In this embodiment, the objective weight calculated is {ωo j |j=1,2,3}={0.3404,0.3749,0.2846}.

[0115] The comprehensive weight calculation module 16 is used to calculate the comprehensive weight corresponding to each risk factor according to all the final subjective weights and objective weights. The calculation formula is:

[0116]

[0117] In this embodiment, the comprehensive weight calculated is {ω j |j=1,2,3}={0.4085,0.3702,0.2213}.

[0118] The comprehensive score calculation module 17 is used to calculate the comprehensive score of each artificial intelligence safety potential risk mode according to the SVTraN-CODAS method, the SVTraNN decision matrix and all comprehensive weights.

[0119] The comprehensive score calculation module 17 includes a normalization unit 171 , a weighting unit 172 , a negative ideal solution unit 173 , a distance calculation unit 174 , a correlation evaluation matrix construction unit 175 and a comprehensive score calculation unit 176 .

[0120] The specific implementation method of the SVTraN-CODAS method in this embodiment is as follows:

[0121] The normalization unit 171 is used to normalize the SVTraNN decision matrix by using the RAFSI normalization method to obtain the SVTraNN normalized decision matrix.

[0122] In this embodiment, the SVTraNN score matrix is ​​used as a criterion to determine the ideal solution under each risk factor. and non-ideal solutions The results are as follows:

[0123]

[0124] Secondly, through the function H m×3 All elements of are mapped to the standard interval [h1,h 2c ], take h1=<(0.49,0.64,0.69,0.82); 0.2,0.8,0.75>, h 2c =<(1.79,2.06,2.14,2.27);0.7,0.3,0.2>, ξ(h1)=0.8410, ξ(h 2c )=7.5089,ξ(h 2c ) is about 9 times of ξ(h1), and the result is expressed as G m×3 =[g ij ] m×3 .

[0125] Next, calculate h1 and h 2c The harmonic mean HM and arithmetic mean AM of

[0126] Finally, get the SVTraNN normalized decision matrix For the largest indicators, For the smallest indicator, In FMEA, the three risk factors are the largest indicators.

[0127] The weighting unit 172 is used to obtain a weighted normalized matrix according to the comprehensive weight and the SVTraNN normalized decision matrix.

[0128] Among them, the expression of the weighted normalization matrix is:

[0129] V m×3 =[v ij ] m×3 ,

[0130]

[0131] Where V m×3 is the weighted normalization matrix, vij is the weighted normalization matrix V m×3 The element in the i-th row and j-th column of j is the comprehensive weight corresponding to the j-th risk factor, is the element in the i-th row and j-th column of the SVTraNN normalized decision matrix.

[0132] The weighted normalization matrix calculated in this embodiment is:

[0133]

[0134] The negative ideal solution unit 173 is used to obtain the negative ideal solution of each risk factor according to the SVTraNN normalized decision matrix and the corresponding SVTraNN score. In this embodiment, the negative ideal solution is composed of the SVTraNN data corresponding to the lowest score in each column, and the negative ideal solution ns = {<(0.09, 0.10, 0.10, 0.11); 0.11, 0.89, 0.88>, <(0.08, 0.09, 0.09, 0.10); 0.10, 0.90, 0.89>, <(0.05, 0.05, 0.05, 0.06); 0.06, 0.94, 0.93>}.

[0135] The distance calculation unit 174 is used to calculate the Euclidean distance and Manhattan distance corresponding to each artificial intelligence safety potential risk mode according to the SVTraNN normalized decision matrix and the negative ideal solution.

[0136] Among them, the Euclidean distance E corresponding to the i-th AI security potential risk mode is i The calculation expression is:

[0137]

[0138] In the formula and is the true membership degree corresponding to the i-th AI safety potential risk model under the j-th risk factor, and is the uncertainty membership corresponding to the i-th AI safety potential risk mode under the j-th risk factor, and is the pseudo-membership corresponding to the ith AI safety potential risk model under the jth risk factor, and is the true membership degree corresponding to the negative ideal solution of the j-th risk factor, and is the uncertain membership corresponding to the negative ideal solution of the j-th risk factor, and is the pseudo-membership corresponding to the negative ideal solution of the j-th risk factor.

[0139] Manhattan distance T corresponding to the i-th AI safety potential risk pattern i The calculation expression is:

[0140]

[0141] The calculation results of the Euclidean distance and Manhattan distance of each potential risk model of artificial intelligence security are shown in the following table:

[0142]

[0143]

[0144] The correlation evaluation matrix construction unit 175 is used to construct a correlation evaluation matrix according to all Euclidean distances and Manhattan distances.

[0145] Among them, the expression of the relevant evaluation matrix is:

[0146] RA=[o ik ] m×m ,

[0147] o ik =(E i -E k )+Ψ(E i -E k )*(T i -T k ),

[0148]

[0149] Wherein τ is a preset value. In this embodiment, τ is set to 0.03.

[0150] The comprehensive score calculation unit 176 is used to calculate the comprehensive score of each artificial intelligence safety potential risk mode according to the relevant evaluation matrix.

[0151] Among them, the comprehensive score O corresponding to the i-th AI security potential risk mode i The calculation expression is:

[0152]

[0153] Where m is the total number of potential risk modes for AI security.

[0154] Figure 2 It is a schematic diagram of the calculation results of the relevant evaluation matrix and the comprehensive score in the embodiment of the present invention.

[0155] like Figure 2As shown, the second to twentieth rows in the table are the relevant evaluation values ​​of artificial intelligence potential safety risk modes FM1 to FM19 and other modes, and the twenty-first row in the table is the comprehensive score of each artificial intelligence potential safety risk mode FM1 to FM19.

[0156] The risk ranking module 18 is used to obtain the risk ranking results of each artificial intelligence security potential risk mode according to all comprehensive scores. In this embodiment, the comprehensive scores are sorted from large to small, and the higher the risk ranking result, the higher the risk of the corresponding artificial intelligence security potential risk mode.

[0157] The risk ranking results in this embodiment are shown in the following table:

[0158] serial number Risk Model <![CDATA[O i ]]> Sorting <![CDATA[FM1]]> Over-collection of training data -1.771 17 <![CDATA[FM2]]> Training data collection is not compliant -1.277 14 <![CDATA[FM3]]> Training data poisoning -1.116 12 <![CDATA[FM4]]> Membership Inference Attack -1.552 15 <![CDATA[FM5]]> Training data leakage 1.253 9 <![CDATA[FM6]]> Training data forgery -1.051 10 <![CDATA[FM7]]> Algorithm design flaws 1.559 8 <![CDATA[FM8]]> Algorithm backdoor attack 2.319 1 <![CDATA[FM9]]> Weak algorithm robustness 1.946 4 <![CDATA[FM 10 ]]> Algorithmic unexplainability -1.101 11 <![CDATA[FM 11 ]]> Algorithmic bias and discrimination 1.604 7 <![CDATA[FM 12 ]]> Adversarial Example Attack 1.842 5 <![CDATA[FM 13 ]]> Model Stealing Attack 2.179 2 <![CDATA[FM 14 ]]> Model Reversal Attack -1.120 13 <![CDATA[FM 15 ]]> Model feedback misleading -2.267 18 <![CDATA[FM 16 ]]> Software Framework -1.683 16 <![CDATA[FM 17 ]]> Computing support facilities -3.338 19 <![CDATA[FM 18 ]]> Generative Model Abuse 1.965 3 <![CDATA[FM 19 ]]> Social Engineering Attacks 1.611 6

[0159] The first column in the table above is the number of each AI security potential risk model, the second column is the content of each AI security potential risk model, and the third and fourth columns are the comprehensive scores and corresponding ranking numbers of each AI security potential risk model. It can be seen that the five risk models of "algorithm backdoor attack", "model stealing attack", "generated model abuse", "algorithm weak robustness" and "adversarial sample attack" are relatively more likely to cause highly harmful security risks to human society.

[0160] The following describes the process of obtaining risk ranking results using the improved FMEA device 100 in conjunction with the accompanying drawings.

[0161] Figure 3 It is a flow chart of generating risk ranking results by improving the FMEA device in an embodiment of the present invention.

[0162] like Figure 3 As shown, the improved FMEA device 100 is used to obtain the risk ranking result, including the following steps:

[0163] Step S1, using the SVTraNN score conversion module 11 to obtain the SVTraNN evaluation matrix corresponding to each expert according to the SVTraNN language variable conversion table and the score data.

[0164] Step S2, using the expert weight calculation module 12 to calculate the expert weight of each expert for each risk factor according to the SVTraNNs score corresponding to the risk factor in all SVTraNNs evaluation matrices.

[0165] Step S3, using the expert evaluation summary module 13 to calculate the SVTraNN decision matrix according to all SVTraNNs evaluation matrices and expert weights.

[0166] Step S4, using the subjective weight calculation module 14 to calculate the final subjective weight corresponding to each risk factor according to the SVTraNN language variable conversion table, the SVTraN-FUCOM planning model and all risk factor importance score data.

[0167] Step S5, using the objective weight calculation module 15 to construct a target planning model according to the SVTraNN decision matrix and the difference, and calculate the objective weight corresponding to each risk factor.

[0168] Step S6: using the comprehensive weight calculation module 16 to calculate the comprehensive weight corresponding to each risk factor according to all the final subjective weights and objective weights.

[0169] Step S7, using the comprehensive score calculation module 17 to calculate the comprehensive score of each artificial intelligence security potential risk mode according to the SVTraN-CODAS method, the SVTraNN decision matrix and all comprehensive weights.

[0170] Step S8, using the risk ranking module 18 to obtain the risk ranking results of each artificial intelligence security potential risk mode based on all comprehensive scores.

[0171] The risk level module 19 is used to perform clustering and stratification through ISODATA according to the comprehensive score, scoring data and comprehensive weight to obtain the risk level of each potential risk mode of artificial intelligence security.

[0172] In this embodiment, the risk level module 19 firstly uses the comprehensive score as the first characteristic value; secondly, the scoring data is averaged and integrated to obtain the data corresponding to each risk factor, namely and Next, WARPN is calculated based on the comprehensive weight i As the second eigenvalue, its calculation expression is Then, the first eigenvalue and the second eigenvalue are normalized respectively, and the calculation expression is: Then generate the input of ISODATA clustering, which is the two-dimensional feature data

[0173] The various data calculated in the above process are shown in the following table:

[0174]

[0175] The first column in the table above is the number of the potential risk model for AI security, and the second to seventh columns are the corresponding WARPN i , and

[0176] When clustering ISODATA, the risk level module 19 uses the silhouette coefficient, CH score and Davidson-Botting index DBI as measurement indicators to obtain the best clustering results.

[0177] Figure 4 It is a schematic diagram of the results of ISODATA clustering in an embodiment of the present invention.

[0178] like Figure 4 As shown, the horizontal axis is The value of The value of the result is 0.7563, CH score is 209.1, and DBI is 0.1868, indicating that the clustering result has significant superiority. Finally, three risk levels are obtained, namely "high risk", "medium risk" and "low risk".

[0179] The "high risk" mode includes algorithm backdoor attacks on FM8 and model stealing attacks on FM 13 , Generative Model Abuse of FM 18 , Algorithm weak robustness FM9, Adversarial sample attack FM 12 , Social Engineering Attack FM 19 , Algorithmic Bias Discrimination FM 11 , algorithm design vulnerability FM7, training data leakage FM5.

[0180] The "medium risk" mode includes training data forgery FM6 and algorithm unexplainability FM 10 , training data poisoning FM3, model reverse attack FM 14 , training data collection non-compliant FM2, member reasoning attack FM4, software framework FM 16 , training data over-collection FM1, model feedback misleading FM 15 .

[0181] The “low risk” model includes hashing support facilities FM 17 .

[0182] The control module 20 stores a control program for controlling the operation of each module.

[0183] Functions and Effects of the Embodiments

[0184] According to the improved FMEA device that considers the relationship between uncertainty and failure mode involved in this embodiment, first, the scoring data is converted into SVTraNN, and the uncertainty and fuzziness of the expert evaluation information are considered; second, the expert weight is calculated by the confidence level, which solves the shortcoming that the traditional FMEA does not consider the differences of experts; third, the final subjective weight is solved according to the SVTraN-FUCOM planning model, and the objective weight is solved according to the difference. The problem that the traditional FMEA method is highly subjective and does not consider the differences in the impact of different risk factors on the research object is solved; fourth, the comprehensive weight is calculated using the objective weight and the final subjective weight, and the comprehensive score is further calculated and ranked according to the SVTraN-CODAS method to obtain the risk ranking result, while solving the shortcomings of the traditional FMEA method RPN calculation lacking robustness and scientificity, and adding consideration to the relationship between risk modes. In short, this method can effectively evaluate the risk level of risk modes.

[0185] In addition, in this embodiment, the risk level of the risk model is obtained according to ISODATA clustering through the risk level module, so that the manager can implement corresponding solutions for risk models of different risk levels.

[0186] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An improved FMEA device that considers the relationship between uncertainty and failure mode, which is used to obtain the risk ranking results and risk levels of each potential risk mode of artificial intelligence safety according to the scoring data of each expert on each potential risk mode of artificial intelligence safety and the scoring data of the importance of risk factors, characterized in that: include: An SVTraNN scoring conversion module stores a preset SVTraNN language variable conversion table, and is used to obtain the SVTraNN evaluation matrix corresponding to each of the experts according to the SVTraNN language variable conversion table and the scoring data; An expert weight calculation module, for calculating, for each risk factor, the expert weight of each expert for the risk factor according to the SVTraNNs score corresponding to the risk factor in all the SVTraNNs evaluation matrices; An expert evaluation summary module, used for calculating the SVTraNN decision matrix according to all the SVTraNNs evaluation matrices and the expert weights; A subjective weight calculation module, storing an SVTraNN language variable conversion table and an SVTraN-FUCOM planning model, for calculating the final subjective weight corresponding to each of the risk factors according to the SVTraNN language variable conversion table, the SVTraN-FUCOM planning model and all the risk factor importance score data; An objective weight calculation module, storing a difference construction target planning model, for calculating the objective weight corresponding to each of the risk factors according to the SVTraNN decision matrix and the difference construction target planning model; A comprehensive weight calculation module, used to calculate the comprehensive weight corresponding to each of the risk factors according to all the final subjective weights and the objective weights; A comprehensive score calculation module, used to calculate the comprehensive score of each of the artificial intelligence security potential risk modes according to the SVTraN-CODAS method, the SVTraNN decision matrix and all the comprehensive weights; A risk ranking module, used to obtain the risk ranking results of each of the potential risk modes of artificial intelligence security based on all the comprehensive scores; A risk level module is used to perform clustering and stratification according to the comprehensive score, the scoring data and the comprehensive weight to obtain the risk level of each of the potential risk modes of artificial intelligence security. Among them, the higher the risk ranking result is, the higher the risk of the corresponding artificial intelligence security potential risk mode is.

2. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 1, Features: in, The expert weight calculation module includes: An average similarity calculation unit, used for calculating the average similarity corresponding to each of the risk factors according to the SVTraNN evaluation matrix of the experts in pairs; A confidence level calculation unit, used for calculating the confidence level corresponding to each of the experts for each of the risk factors according to the average similarity between the experts; The expert weight calculation unit is used to calculate the expert weight corresponding to each of the experts for each of the risk factors according to the confidence levels of all the experts.

3. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 2 is characterized in that: in, The average similarity AS between expert k and expert l on risk factor S S (e k ,e l ) is calculated as: f=<(a f ,b f ,c f ,d f );T f ,I f ,F f >, g=<(a g ,b g ,c g ,d g );T g ,I g ,F g >, Where m is the total number of potential risk patterns for AI security, f and g are two SVTraNNs, is the SVTraNN corresponding to the i-th AI security potential risk pattern in the SVTraNN evaluation matrix corresponding to expert k, is the SVTraNN corresponding to the i-th AI security potential risk mode in the SVTraNN evaluation matrix corresponding to expert l, x fT 、x gT ,y fT ,y gT Corresponding to the true membership, x fI 、x gI ,y fI ,y gI Corresponding to the uncertain membership, x fF 、x gF ,y fF ,y gF Corresponding to the pseudo membership, The confidence level CL corresponding to expert k under risk factor S S (e k ) is calculated as: Where t is the total number of experts, The expert weight ws corresponding to expert k under risk factor S k The calculation expression is:

4. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 1 is characterized in that: in, The SVTraNN evaluation matrix H corresponding to expert k k The expression is: The first, second, and third columns correspond to the scores of risk factors S, O, and D, respectively. The expression of the SVTraNN decision matrix is: Where i = 1,…,m, m is the total number of potential risk modes of AI security, and are the scores of the i-th AI security potential risk model in the SVTraNN evaluation matrix corresponding to the k-th expert under risk factors S, O, and D, respectively, and ws k ,wo k and wd k are the expert weights corresponding to the kth expert under risk factors S, O, and D respectively.

5. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 1, Features: in, The subjective weight calculation module includes: An SVTraNN conversion unit, used to convert the risk factor importance score data according to the SVTraNN language variable conversion table to obtain SVTraNN data corresponding to each expert; A subjective weight solving unit, used to solve the SVTraN-FUCOM planning model according to the SVTraNN data corresponding to each of the experts through the SVTraN-FUCOM planning model, and obtain the subjective weight of each of the experts for each of the risk factors; The final subjective weight calculation unit is used to calculate the average value of the subjective weights corresponding to all the experts for each risk factor, and use the calculation result as the final subjective weight corresponding to the risk factor.

6. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 5 is characterized in that: in, When solving the subjective weight of expert e1, the corresponding expression of the SVTraN-FUCOM planning model is: x S.t In the formula is the SVTraNN data corresponding to the score of the i-th expert on the importance of the j-th risk factor, is the SVTraNN data corresponding to the subjective weight of the i-th expert on the j-th risk factor, is the subjective weight of the i-th expert on the j-th risk factor.

7. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 1, Features: in, The objective weight calculation module comprises: A score matrix generating unit, used for performing SVTraNN score calculation and normalization on the SVTraNN decision matrix to obtain a SVTraNN normalized score matrix; A parameter calculation unit, used for calculating the standard deviation corresponding to each of the risk factors and the Pearson correlation coefficient between any two risk factors according to the SVTraNN normalized score matrix; The objective weight solving unit is used to solve the difference construction target programming model according to the standard deviation and the Pearson correlation coefficient to obtain the objective weight corresponding to each of the risk factors.

8. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 7 is characterized in that: in, The calculation expression of the SVTraNN score calculation is: Where <(a f ,b f ,c f ,d f );T f ,I f ,F f > is the expression corresponding to the SVTraNN element f, ξ(f) is the result of SVTraNN score calculation for the element f, The expression of the difference construction target programming model is: minλ b +0.5λ c S.t ωo j >0,j=1,2,3. Where ωo j is the objective weight of the j-th risk factor, σ j is the standard deviation of all normalized scores of the jth risk factor, r jl is the Pearson correlation coefficient between the j-th risk factor and all normalized scores of the l-th risk factor.

9. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 1, Features: in, The comprehensive score calculation module includes: A normalization unit, used for normalizing the SVTraNN decision matrix by using a normalization method of RAFSI to obtain a SVTraNN normalized decision matrix; A weighting unit, used for obtaining a weighted normalized matrix according to the comprehensive weight and the SVTraNN normalized decision matrix; A negative ideal solution unit, used for obtaining a negative ideal solution of each of the risk factors according to the SVTraNN normalized decision matrix and the corresponding SVTraNN score; A distance calculation unit, used to calculate the Euclidean distance and Manhattan distance corresponding to each of the artificial intelligence safety potential risk modes according to the SVTraNN normalized decision matrix and the negative ideal solution; A correlation evaluation matrix construction unit, used for constructing a correlation evaluation matrix according to all the Euclidean distances and the Manhattan distances; A comprehensive score calculation unit is used to calculate the comprehensive score of each of the potential risk modes of artificial intelligence security according to the relevant evaluation matrix.

10. The improved FMEA device considering the relationship between uncertainty and failure mode according to claim 9 is characterized in that: in, The expression of the weighted normalization matrix is: V m×3 =[v ij ] m×3 ' Where V m×3 is the weighted normalization matrix, v ij is the weighted normalization matrix V m×3 The element in the i-th row and j-th column of j is the comprehensive weight corresponding to the j-th risk factor, is the element in the i-th row and j-th column of the SVTraNN normalized decision matrix, The Euclidean distance E corresponding to the i-th AI security potential risk model i The calculation expression is: In the formula and is the true membership degree corresponding to the i-th AI safety potential risk model under the j-th risk factor, and is the uncertainty membership corresponding to the i-th AI safety potential risk mode under the j-th risk factor, and is the pseudo-membership corresponding to the ith AI safety potential risk model under the jth risk factor, and is the true membership degree corresponding to the negative ideal solution of the j-th risk factor, and is the uncertain membership corresponding to the negative ideal solution of the j-th risk factor, and is the pseudo-membership degree corresponding to the negative ideal solution of the j-th risk factor, Manhattan distance T corresponding to the i-th AI safety potential risk pattern i The calculation expression is: The expression of the relevant evaluation matrix is: RA=[o ik ] m×m , or ik =(And i -AND k )+Ψ(E i -AND k )*(T i -T k ), Where τ is the preset value, The comprehensive score O corresponding to the i-th AI security potential risk model i The calculation expression is: Where m is the total number of potential risk modes for AI security.

Images