A method and system for automatic updating of diagnostic rules

Abstract

The application discloses a kind of diagnostic rule automatic updating method and system, including data acquisition module, statistical module, influence factor calculation module and update module, data acquisition module is used to extract the diagnostic rule of PHM system, diagnostic process output, diagnostic result output and field maintenance feedback information;Statistical module is used to according to diagnostic result output and field maintenance feedback information statistics diagnostic rule fusion decision layer's accuracy, false alarm rate and false negative rate;Influence factor calculation module is used to calculate the influence factor of the submodel of fusion decision layer to comprehensive diagnostic result.The data analysis efficiency of the application is high, can significantly shorten data analysis period, and can automatically update the diagnostic rule of PHM system, so that PHM system realizes automatic closed loop.

Description

Technical Field

[0001] This invention relates to a rule updating method, and more particularly to a method and system for automatically updating diagnostic rules. Background Technology

[0002] Currently, the PHM system for railcar doors relies mainly on offline data analysis for information processing after early warning pushes, such as credit investigation, operational data analysis, model analysis, and model optimization. However, offline data analysis is inefficient and prone to errors, resulting in high tracking and maintenance costs for the PHM system and long model optimization cycles, which is not conducive to improving the diagnostic effectiveness of the PHM system. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a method and system for automatically updating diagnostic rules with high data analysis efficiency and short analysis cycle.

[0004] Technical solution: The present invention provides an automatic update method for diagnostic rules, comprising the following steps:

[0005] Extract data, including the diagnostic rules, diagnostic process output, diagnostic result output, and on-site maintenance feedback information of the PHM system. The diagnostic rules include the fusion decision layer and the initial diagnostic layer.

[0006] Based on the diagnostic results output and on-site maintenance feedback, the accuracy p of the fusion diagnostic results of the fusion decision layer is statistically analyzed. c False alarm rate p e , false negative rate p l ;

[0007] Analyze the factors influencing the comprehensive diagnostic results of the sub-models of the fusion decision layer;

[0008] Set the conditions for automatic update triggering and iteration termination of diagnostic rules;

[0009] When the automatic update trigger condition is met, the weights of the sub-models in the fusion decision layer are updated in combination with the influence factors;

[0010] When the automatic update trigger condition is met, update the machine learning algorithm class model of the initial diagnostic layer and update the parameters of the threshold judgment class model of the initial diagnostic layer;

[0011] The accuracy p of the updated fusion decision layer and initial diagnostic layer is statistically analyzed. c False alarm rate p e , false negative rate p l The algorithm checks whether the updated value meets the termination condition of the iteration. If it does, the update ends; otherwise, it proceeds to the next update until the termination condition is met.

[0012] Furthermore, the accuracy p of the fusion diagnostic results of the statistical fusion decision layer...c False alarm rate p e , false negative rate p l The calculation method is as follows: the statistical period is set from the (u-1)th update to before the uth update.

[0013]

[0014]

[0015]

[0016] Where TotalNum(u-1) refers to the sum of the number of early warning pushes from the PHM system and the number of expected omissions based on on-site maintenance feedback, N c (u-1) represents the number of accurate early warnings reported from on-site maintenance, N e (u-1) represents the number of on-site maintenance feedback warning errors, N l (u-1) represents the number of missed early warnings from on-site maintenance feedback;

[0017] The automatic trigger condition for data statistics is when TotalNum(u-1) reaches the set threshold TotalNum(th).

[0018] Furthermore, the calculation method for the influence factor of the sub-model of the fusion decision layer on the comprehensive diagnostic results is as follows:

[0019] For the k-th accurately predicted data point, the prediction confidence scores of the initial diagnostic layer and the fusion decision layer are obtained through the diagnostic process output. These scores are then sorted from high to low. Assuming the initial diagnostic layer has M sub-models, and each sub-model corresponds to N candidate sub-health types, the prediction confidence matrix P(k) of the initial diagnostic layer's sub-models is obtained.

[0020]

[0021] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, p ij (k)∈[0,1];

[0022] The confidence matrix R(k) of the fusion diagnosis results of the fusion decision layer is obtained.

[0023] R(k)=[q1(k) q2(k) Λ q j (k) Λ q N (k)]

[0024] Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1];

[0025] The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k).

[0026]

[0027] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is either 0 or 1;

[0028] Calculate the consistency coefficient matrix A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results.

[0029] A(k)=[α1(k) α2(k) ... α i (k) ... α M (k)] T

[0030] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N,

[0031] Calculate the confidence distance matrix D(k) between the sub-models of the initial diagnostic layer and each prediction type of the fused diagnostic results.

[0032] D(k)=[d1(k) d2(k) Λ d i (k)Λ d M (k)] T

[0033] Where 1≤k≤N c (u-1), 1≤i≤M,

[0034] Calculate the rectangle F(k) representing the influence factor of the sub-model prediction results of the initial diagnostic layer on the fused diagnostic results.

[0035] F(k)=[β1(k) ... β i (k) ... β M (k)] T

[0036] Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. Where g(d) i(k) is the confidence distance d i (k) is a function of the independent variable, 0 < g(d) i (k))≤1, use an exponential function with a base greater than 1 for calculation;

[0037] Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results.

[0038] F(u-1)=[γ1(u-1) ... γ i (u-1) ... γ M (u-1)] T

[0039] Where 1≤i≤M,

[0040] Furthermore, the automatic update trigger condition refers to: setting an initial accuracy p. c (0), false alarm rate p e (0), false negative rate p l (0), set the thresholds for accurate early warning, false alarms, and missed alarms as p respectively. c (th), p e (th), p l (th);

[0041] Set automatic update trigger conditions: p c (u-1)<p c (th), p e (u-1)>p e (th), p l (u-1)>p l (th) l p c (u-1)<p c (u-2), p e (u-1)>p e (u-2), p l (u-1)>p l (u-2);

[0042] When u=1, if any one of the first three conditions is met, the fusion decision layer and the initial diagnosis layer will be updated.

[0043] When u≥2, and any one of the six conditions is met, the fusion decision layer and the initial diagnosis layer are updated.

[0044] Furthermore, the condition for terminating the iteration refers to: assuming that after the v-th iteration before the u-th update calculation deadline, the accuracy of the fusion diagnosis result of the fusion decision layer is p. c (v) The false alarm rate is pe (v) The false negative rate is p l (v);

[0045] Set the iteration termination condition: p c (v)≥p c (th), p e (v)≤p e (th), p l (v)≤p l (th);

[0046] The iteration calculation is terminated when any one of the three conditions is met.

[0047] Furthermore, the formula for calculating the weights of the sub-models of the fusion decision layer by combining influence factors is as follows: Let W be the weight matrix of the sub-models of the fusion decision layer after the U-th update. U ,

[0048]

[0049] Where M refers to the number of sub-models in the fusion decision layer, w i (u) refers to the weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M;

[0050] Let the initial prediction accuracy of each sub-model be:

[0051] P c (0)=[p1(0) p2(0) Λ p i (0) Λ p M (0)];

[0052] Based on the initial prediction accuracy of each sub-model, set the initial weights for each sub-model:

[0053] W(0)=[w1(0) w2(0) Λ w i (0) Λ w M (0)]

[0054] Wherein, the weight of the i-th sub-model

[0055] The formula for calculating the weight value after the u-th update is as follows:

[0056]

[0057] Furthermore, the update of the machine learning algorithm model refers to adding positive true examples to the sample library, setting the model training parameters optimization space, retraining the model using new samples, and stopping training when the model's prediction accuracy increases or reaches a preset value, and updating the model and its configuration parameters.

[0058] Furthermore, the parameters of the updated threshold judgment model refer to: assuming that before the u-th update calculation begins, the confidence push threshold parameter for the j-th candidate sub-health type of the i-th sub-model is t. ij (u-1), the probability distribution of the confidence level of the positive feedback error result is [a ij ,b ij ], then the threshold parameter adjustment step size calculated in the v-th iteration before the u-th update deadline is:

[0059]

[0060] Where θ i (v) is the influence factor adjustment parameter, and its value range is [0,1].

[0061] If the false negative rate is greater than the accuracy rate or the false negative rate increases, then lower the push threshold, set the upper and lower confidence limits for false negative sub-health type j corresponding to a and b, and adjust the threshold parameter t. ij (v+1)=t ij (u-1)- Δ t ij (v);

[0062] If the false alarm rate is greater than the accuracy rate or the false alarm rate increases, the push threshold is increased, and the upper and lower confidence limits of the false alarm results are set for a and b respectively. The adjusted threshold parameter t is then adjusted. ij (v+1)=t ij (u-1)+ Δ t ij (v).

[0063] Based on the same inventive concept, the present invention also provides an automatic diagnostic rule update system, comprising,

[0064] The data acquisition module is used to extract diagnostic rules, diagnostic process outputs, diagnostic result outputs, and on-site maintenance feedback information from the PHM system.

[0065] The statistics module is used to calculate the accuracy, false alarm rate, and false negative rate of the integrated decision-making layer of diagnostic rules based on the diagnostic results output and on-site maintenance feedback information.

[0066] The impact factor calculation module is used to calculate the impact factors of the sub-models of the fusion decision layer on the comprehensive diagnostic results.

[0067] The update module is used to set the conditions for automatic update triggering and iteration termination of diagnostic rules based on the early warning accuracy, false alarm rate and false negative rate. When the automatic update triggering conditions are met, the weights of the sub-models of the fusion decision layer are updated in combination with the influencing factors, and the parameters of the machine learning algorithm model and threshold judgment model of the initial diagnostic layer are updated until the iteration termination conditions are met.

[0068] Furthermore, the influencing factor is calculated as follows: For the k-th accurately predicted data point, the prediction confidence of the initial diagnostic layer and the fusion decision layer is obtained through the diagnostic process output. The prediction confidence of the initial diagnostic layer and the fusion decision layer is sorted from high to low. Assuming that the initial diagnostic layer has M sub-models and each sub-model corresponds to N candidate sub-health types, the prediction confidence matrix P(k) of the sub-models of the initial diagnostic layer is obtained.

[0069]

[0070] Where 1≤i≤M, 1≤j≤N, p ij ∈[0,1];

[0071] The confidence level R(k) of the fusion diagnosis results of the fusion decision layer is obtained.

[0072] R(k)=[q1(k) q2(k) Λ q j (k) Λ q N (k)]

[0073] Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1];

[0074] The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k).

[0075]

[0076] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is 0 or 1;

[0077] Calculate the consistency coefficient A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results.

[0078] A(k)=[α1(k) α2(k) ... α i (k) ... α M (k)] T

[0079] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N,

[0080] Calculate the confidence distance D(k) between the sub-model of the initial diagnostic layer and each prediction type of the fusion diagnostic result.

[0081] D(k)=[d1(k) d2(k) Λ d i (k)Λ d M (k)] T

[0082] Where 1≤k≤N c (u-1), 1≤i≤M,

[0083] Calculate the influence factor F(k) of the sub-model prediction results of the initial diagnostic layer on the fusion diagnostic results.

[0084] F(k)=[β1(k) ... β i (k) ... β M (k)] T

[0085] Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. Where g(d) i (k) is the confidence distance d i (k) is a function of the independent variable, 0 < g(d) i (k))≤1, use an exponential function with a base greater than 1 for calculation.

[0086] Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results.

[0087] F(u-1)=[γ1(u-1) ... γ i (u-1) ... γ M (u-1)] T

[0088] Where 1≤i≤M,

[0089] The formula for calculating the weights of the sub-models of the fusion decision layer by combining influence factors is as follows: Let W be the weight matrix of the sub-models of the fusion decision layer after the U-th update. U ,

[0090]

[0091] Where M refers to the number of sub-models in the fusion decision layer, w i (u) refers to the weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M;

[0092] Let the initial prediction accuracy of each sub-model be:

[0093] P c (0)=[p1(0) p2(0) Λ p i (0) Λ p M (0)];

[0094] Based on the initial prediction accuracy of each sub-model, set the initial weights for each sub-model:

[0095] W(0)=[w1(0) w2(0) Λ w i (0) Λ w M (0)]

[0096] Wherein, the weight of the i-th sub-model

[0097] The formula for calculating the weight value after the u-th update is as follows:

[0098]

[0099] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention can automatically adjust the diagnostic rules of the PHM system; the present invention has high data analysis efficiency and can significantly shorten the data analysis cycle. It adopts online analysis to automatically analyze the data on the impact of on-site maintenance feedback information and diagnostic rules on early warning results, thereby reducing the subsequent PHM system tracking and maintenance costs and enabling the PHM system to achieve automatic closed-loop processing. Attached Figure Description

[0100] Figure 1 This is a flowchart of the method of the present invention.

[0101] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0102] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0103] Example 1

[0104] like Figure 1 As shown, the automatic update method for diagnostic rules according to the present invention includes the following steps:

[0105] (1) Data extraction includes the diagnostic rules, diagnostic process output, diagnostic result output, and on-site maintenance feedback information of the PHM system. The diagnostic rules include a fusion decision layer and an initial diagnostic layer. The initial diagnostic layer contains multiple sub-models (referred to as "sub-health diagnostic sub-models"), and each sub-model outputs initial results (referred to as "sub-health diagnostic initial results"). The initial results of the initial diagnostic layer are used as input data to the fusion decision layer. The fusion decision layer makes fusion decisions based on the initial results of each sub-model and the fusion strategy, and outputs a comprehensive diagnostic result (i.e., diagnostic result output). The PHM system diagnostic result output is used to push early warnings. Staff perform maintenance based on the early warning pushes. The accuracy of the early warning pushes can be judged by the on-site maintenance feedback information. That is, the on-site maintenance feedback information can be used as analytical data to provide a basis for subsequent optimization of the PHM system rule updates.

[0106] (2) Based on the diagnostic results output and on-site maintenance feedback information, the accuracy p of the fusion diagnostic results of the fusion decision layer is statistically analyzed. c False alarm rate p e , false negative rate p l .

[0107] The accuracy p of the fusion diagnostic results of the statistical fusion decision layer c False alarm rate p e , false negative rate p l The calculation method is as follows: the statistical period is set from the (u-1)th update to before the uth update.

[0108]

[0109]

[0110]

[0111] Where TotalNum(u-1) refers to the sum of the number of early warning pushes from the PHM system and the number of expected omissions based on on-site maintenance feedback, N c (u-1) represents the number of accurate early warnings reported from on-site maintenance, N e (u-1) represents the number of on-site maintenance feedback warning errors, N l (u-1) represents the number of missed early warnings from on-site maintenance feedback. The accuracy, false alarm rate, and false negative rate of the integrated diagnostic results are the same as the accuracy, false alarm rate, and false negative rate of the PHM system's diagnostic output results.

[0112] The automatic trigger condition for data statistics is when TotalNum(u-1) reaches the set threshold TotalNum(th).

[0113] (3) Analyze the factors that influence the sub-models of the fusion decision layer on the comprehensive diagnostic results.

[0114] The specific calculation method is as follows:

[0115] For the k-th accurately predicted data point, the prediction confidence scores of the initial diagnostic layer and the fusion decision layer are obtained through the diagnostic process output. These scores are then sorted from high to low. Assuming the initial diagnostic layer has M sub-models, and each sub-model corresponds to N candidate sub-health types, the prediction confidence matrix P(k) of the initial diagnostic layer's sub-models is obtained.

[0116]

[0117] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, p ij (k)∈[0,1];

[0118] The confidence matrix R(k) of the fusion diagnosis results of the fusion decision layer is obtained.

[0119] R(k)=[q1(k) q2(k) Λ q j (k) Λ q N (k)]

[0120] Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1];

[0121] The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k).

[0122]

[0123] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is either 0 or 1;

[0124] Calculate the consistency coefficient matrix A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results.

[0125] A(k)=[α1(k) α2(k) ... α i (k) ... α M (k)] T

[0126] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N,

[0127] Calculate the confidence distance matrix D(k) between the sub-models of the initial diagnostic layer and each prediction type of the fused diagnostic results.

[0128] D(k)=[d1(k) d2(k) Λ d i (k)Λ d M (k)] T

[0129] Where 1≤k≤N c (u-1), 1≤i≤M,

[0130] Calculate the rectangle F(k) representing the influence factor of the sub-model prediction results of the initial diagnostic layer on the fused diagnostic results.

[0131] F(k)=[β1(k) ... β i (k) ... β M (k)] T

[0132] Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. Where g(d) i (k) is the confidence distance d i (k) is a function of the independent variable, 0 < g(d) i (k))≤1, use an exponential function with a base greater than 1 for calculation;

[0133] Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results.

[0134] F(u-1)=[γ1(u-1) ... γ i (u-1) ... γ M (u-1)] T

[0135] Where 1≤i≤M,

[0136] (4) Set the conditions for automatic update triggering and iteration termination of diagnostic rules.

[0137] Automatic update trigger condition: Set the initial accuracy p c (0), false alarm rate p e (0), false negative rate p l (0), set the thresholds for accurate early warning, false alarms, and missed alarms as p respectively. c(th), p e (th), p l (th);

[0138] Set automatic update trigger conditions: p c (u-1)<p c (th), p e (u-1)>p e (th), p l (u-1)>p l (th), p c (u-1)<p c (u-2), p e (u-1)>p e (u-2), p l (u-1)>p l (u-2);

[0139] When u=1, and any one of the first three conditions is met, the update of the fusion decision layer and the initial diagnosis layer is triggered; that is, during the first update, the initial accuracy p c (0) Below the accuracy threshold th pc Initial false alarm rate p e (0) Above the false alarm rate threshold th pe Initial false negative rate p e (0) Above the false negative rate threshold th pl If any one of the three conditions is met, then the automatic update trigger condition is met;

[0140] When u≥2, the fusion decision layer and the initial diagnosis layer are updated when any one of the six conditions is met; that is, when it is not the first update, the automatic update condition is met when any one of the following six conditions is met: accuracy is lower than the threshold, false positive rate is higher than the threshold, false negative rate is higher than the threshold, accuracy is decreasing, false positive rate is increasing, or false negative rate is increasing.

[0141] The condition for termination of iteration:

[0142] Let p be the accuracy of the fusion diagnosis results of the fusion decision layer after the v-th iteration before the u-th update calculation deadline. c (v) The false alarm rate is p e (v) The false negative rate is p l (v);

[0143] Set the iteration termination condition: p c (v)≥p c (th), p e (v)≤p e (th), p l (v)≤p l (th);

[0144] The iteration calculation is terminated when any one of the three conditions is met; that is, after the v-th update, the accuracy reaches or exceeds the threshold, the false alarm rate is lower than or equal to the threshold, and the false negative rate is lower than or equal to the threshold. If any one of the three conditions is met, the iteration termination condition is met.

[0145] (5) When the automatic update triggering condition is met, the weights of the sub-models of the fusion decision layer are updated in combination with the influencing factors.

[0146] The formula for calculating the weights of the sub-models in the fusion decision layer is as follows: Let W be the weight matrix of the sub-models in the fusion decision layer after the U-th update. U ,

[0147]

[0148] Where M refers to the number of sub-models in the fusion decision layer, w i (u) refers to the weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M;

[0149] Let the initial prediction accuracy of each sub-model be:

[0150] P c (0)=[p1(0) p2(0) Λ p i (0) Λ p M (0)];

[0151] Based on the initial prediction accuracy of each sub-model, set the initial weights for each sub-model:

[0152] W(0)=[w1(0) w2(0) Λ w i (0) Λ w M (0)]

[0153] Wherein, the weight of the i-th sub-model

[0154] The formula for calculating the weight value after the u-th update is as follows:

[0155]

[0156] If a sub-model has a significant positive impact on the fusion diagnostic results (i.e., early warning results), then the weight of the sub-model should be increased to strengthen its impact; otherwise, the weight of the sub-model should be decreased to weaken its impact.

[0157] (6) When the automatic update trigger condition is met, update the machine learning algorithm class model of the initial diagnostic layer and update the parameters of the threshold judgment class model of the initial diagnostic layer.

[0158] Updating a machine learning algorithm model involves adding positive real examples to the sample library, setting the model's training parameters to optimize the space, retraining the model using the new samples, and stopping training when the model's prediction accuracy increases or reaches a preset value, and then updating the model and its configuration parameters.

[0159] The parameter update for the threshold judgment model refers to: assuming that before the u-th update calculation begins, the confidence threshold parameter for the j-th candidate sub-health type of the i-th sub-model is t. ij (u-1), the probability distribution of the confidence level of the positive feedback error result is [a ij ,b ij ], then the threshold parameter adjustment step size calculated in the v-th iteration before the u-th update deadline is:

[0160]

[0161] Where θ i (v) is the influence factor adjustment parameter, and its value range is [0,1].

[0162] If the false negative rate is greater than the accuracy rate or the false negative rate increases, then lower the push threshold, set the upper and lower confidence limits for false negative sub-health type j corresponding to a and b, and adjust the threshold parameter t. ij (v+1)=t ij (u-1)- Δ t ij (v);

[0163] If the false alarm rate is greater than the accuracy rate or the false alarm rate increases, the push threshold is increased, and the upper and lower confidence limits of the false alarm results are set for a and b respectively. The adjusted threshold parameter t is then adjusted. ij (v+1)=t ij (u-1)+ Δ t ij (v).

[0164] (7) Calculate the accuracy p of the updated fusion decision layer and the initial diagnostic layer. c False alarm rate p e , false negative rate p l The algorithm checks whether the updated value meets the termination condition of the iteration. If it does, the update ends; otherwise, it proceeds to the next update until the termination condition is met.

[0165] Example 2

[0166] like Figure 2 As shown, the present invention provides an automatic diagnostic rule update system, comprising:

[0167] The data acquisition module is used to extract diagnostic rules, diagnostic process outputs, diagnostic result outputs, and on-site maintenance feedback information from the PHM system.

[0168] The statistics module is used to calculate the accuracy, false alarm rate, and false negative rate of the integrated decision-making layer of diagnostic rules based on the diagnostic results output and on-site maintenance feedback information.

[0169] The impact factor calculation module is used to calculate the impact factor of the sub-diagnostic models of the fusion decision layer on the comprehensive diagnostic results.

[0170] The update module is used to set the conditions for automatic update triggering and iteration termination of diagnostic rules based on the early warning accuracy, false alarm rate and false negative rate. When the automatic update triggering conditions are met, the weights of the sub-diagnostic models of the fusion decision layer are updated in combination with the influencing factors, and the parameters of the machine learning algorithm model and threshold judgment model of the initial diagnostic layer are updated until the iteration termination conditions are met.

[0171] The calculation method for the influencing factor is as follows: For the k-th accurately predicted data, the prediction confidence of the initial diagnostic layer and the fusion decision layer is obtained through the diagnostic process output. The prediction confidence of the initial diagnostic layer and the fusion decision layer are sorted from high to low. Assuming that the initial diagnostic layer has M sub-models and N candidate sub-health types corresponding to the sub-models, the prediction confidence matrix P(k) of the initial diagnostic layer is obtained.

[0172]

[0173] Where 1≤i≤M, 1≤j≤N, p ij ∈[0,1];

[0174] The confidence level R(k) of the fusion diagnosis results of the fusion decision layer is obtained.

[0175] R(k)=[q1(k) q2(k) Λ q j (k) Λ q N (k)]

[0176] Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1];

[0177] The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k).

[0178]

[0179] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is 0 or 1;

[0180] Calculate the consistency coefficient A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results.

[0181] A(k)=[α1(k) α2(k) ... α i (k) ... α M (k)] T

[0182] Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N,

[0183] Calculate the confidence distance D(k) between the sub-model of the initial diagnostic layer and each prediction type of the fusion diagnostic result.

[0184] D(k)=[d1(k) d2(k) Λ d i (k)Λ d M (k)] T

[0185] Where 1≤k≤N c (u-1), 1≤i≤M,

[0186] Calculate the influence factor F(k) of the sub-model prediction results of the initial diagnostic layer on the fusion diagnostic results.

[0187] F(k)=[β1(k) ... β i (k) ... β M (k)] T

[0188] Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. Where g(d) i (k) is the confidence distance d i (k) is a function of the independent variable, 0 < g(d) i (k))≤1, use an exponential function with a base greater than 1 for calculation.

[0189] Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results.

[0190] F(u-1)=[γ1(u-1) ... γ i (u-1) ... γ M (u-1)] T

[0191] Where 1≤i≤M,

[0192] The formula for calculating the weights of the sub-models of the fusion decision layer by combining influence factors is as follows: Let W be the weight matrix of the sub-models of the fusion decision layer after the U-th update. U ,

[0193]

[0194] Where M refers to the number of sub-models in the fusion decision layer, w i (u) refers to the weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M;

[0195] Let the initial prediction accuracy of each sub-diagnostic model be:

[0196] P c (0)=[p1(0) p2(0) Λ p i (0) Λ p M (0)];

[0197] Based on the initial prediction accuracy of each sub-diagnostic model, set the initial weights for each sub-diagnostic model:

[0198] W(0)=[w1(0) w2(0) Λ w i (0) Λ w M (0)]

[0199] Wherein, the weight of the i-th sub-diagnostic model The formula for calculating the weight value after the u-th update is as follows:

[0200]

[0201] Example 3

[0202] This is one of the practical applications of the present invention.

[0203] (1) Extracting data

[0204] The statistical period is defined as from the (u-1)th update to the uth update (hereinafter the same). Before the uth update begins, the PHM system diagnostic rules, diagnostic process output, diagnostic result output, and on-site maintenance feedback information are read. The diagnostic rules include the sub-model configuration parameters of the initial diagnostic layer and the fusion diagnostic rules of the fusion decision layer; the diagnostic process output includes the sub-model prediction confidence matrix of the initial diagnostic layer and the fusion diagnostic result confidence distribution matrix of the fusion decision layer; the diagnostic result output is the sub-health warning type and confidence level after fusion decision; the on-site maintenance feedback information includes warning feedback marker information, etc.

[0205] The sub-model configuration parameters of the initial diagnostic layer and the fusion diagnostic rules of the fusion decision layer are shown in Table 1 below:

[0206] Table 1 Submodel Configuration Parameters and Fusion Diagnostic Rules

[0207]

[0208] The diagnostic process outputs are shown in Tables 2 and 3 below:

[0209] Table 2 Sub-model 1 Sub-health Diagnosis Results

[0210]

[0211] Note 1: This value indicates that the confidence level of the actual sample "sub7" being identified as "sub1" among all candidate sub-health types is 0.45, and the same applies below.

[0212] Note 2: This value indicates that the diagnostic output results for the actual sample "sub7" are sorted from largest to smallest according to the confidence level of all candidate points. The confidence level value (0.55) corresponding to the candidate point "sub7" is ranked first, and the same applies below.

[0213] Note 3: This indicates that the on-site maintenance feedback information is consistent with the PHM system diagnostic process / result output, and the same applies below.

[0214] Table 3 Sub-model 2 Sub-health Diagnosis Results

[0215]

[0216] The fusion diagnosis results are shown in Table 4 below:

[0217] Table 4. Fusion Diagnostic Results

[0218]

[0219]

[0220] (2) Based on the feedback information from on-site maintenance, the diagnostic information is divided into three major datasets, and the prediction accuracy, false alarm rate, and false alarm rate of the integrated decision-making layer and the initial diagnostic layer are statistically analyzed.

[0221] Let the initial accuracy of the model be p obtained during initial model training. c (0), false alarm rate p e (0), false negative rate p l (0), the thresholds for accurate warning, false alarm, and missed warning are p respectively. c (th), p e (th), p l (th).

[0222] For the fusion decision layer, the positive prediction accuracy p of the statistical fusion diagnostic results is... c (u-1), false alarm rate p e (u-1), false negative rate p l (u-1), as shown in Table 5 below:

[0223] Table 5. Prediction accuracy, false alarm rate, and false negative rate of the fusion decision layer

[0224]

[0225] (3) Analyze the factors influencing the comprehensive diagnostic results of the sub-diagnostic models of the integrated decision-making layer.

[0226] For the first accurate diagnostic result record (taking the data with serial number 1 in Tables 2, 3, and 4 as an example), the prediction results of the initial diagnostic layer and the fusion decision layer are sorted from high to low prediction confidence. Then, for the two sub-models and eight candidate sub-health types, the prediction confidence matrix of each sub-model is as follows:

[0227]

[0228] The confidence matrix of the fusion diagnosis results of the fusion decision layer is obtained as follows:

[0229] R(1)=[0.15 0.03 0.00 0.00 0.00 0.29 0.53 0.01]

[0230] Calculate the consistency between the sub-model prediction results and the predicted diagnostic results types in the initial diagnostic layer. For the first diagnostic result record, use R as a sliding window and match the confidence ranking numbers of candidate sub-health types from top to bottom for P. If they match, record it as 1; otherwise, record it as 0. The consistency matrix for the first diagnostic result record is:

[0231]

[0232] Article 1: Consistency coefficient of prediction type between each sub-model and the fusion diagnosis result recorded in the accurate early warning diagnosis result record:

[0233]

[0234] The confidence distance between each sub-model and each prediction type of the fusion diagnostic result recorded in the first diagnostic result record:

[0235] D(1) = [0.1924 0.0357]

[0236] Calculate the influence factors of the sub-model prediction results of the initial diagnostic layer on the fusion diagnostic results. Utilize prediction type consistency and confidence distance to calculate the influence factors of the sub-model prediction results on the fusion diagnostic results.

[0237]

[0238] Following the steps above, the impact factors of other accurately predicted data are calculated. The comprehensive impact factor of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fused diagnostic results is then:

[0239]

[0240] (4) Formulate rule adjustment strategies

[0241] (41) Set diagnostic rules to automatically update trigger conditions and iteration termination conditions.

[0242] If TotalNum(u-1) reaches the set threshold TotalNum(th), the system triggers an automatic rule update judgment.

[0243]

[0244] (42) Update the fusion diagnosis rules for the fusion decision layer. For the fusion decision layer, update the weights of the sub-models. The larger the influence coefficient, the higher the weight; otherwise, reduce the weight.

[0245] If the initial accuracies of the sub-models are 0.95 and 0.91 respectively, then the initial weights W(0) = [0.5108 0.4892].

[0246] If the accuracy of the sub-model before the (u-1)th update is 0.85 and 0.89 respectively, then the weight W(u-1) before the (u-1)th update is [0.4885 0.5115].

[0247] Combining the influence coefficient before the u-th update

[0248] Update the weight value W(u) = [0.2928 0.7072].

[0249] (43) Update the machine learning algorithm class model of the initial diagnostic layer. Add the samples marked as "accurate" to the sample library, retrain the sub-model, and update the corresponding model parameters.

[0250] (44) Update the threshold judgment model parameters of the initial diagnostic layer. For the threshold judgment diagnostic model, statistically analyze the probability distribution of each sub-model predicting the wrong sub-health type. Let t be the confidence push threshold parameter for the 7th sub-health type (sub7) of the 2nd sub-model after the (u-1)th rule update. 27 Given (u-1) = 0.5, and the probability distribution of the confidence level of the positive feedback error result is [0.51, 0.61], then the step size for adjusting the threshold parameter in the v-th iteration before the u-th update deadline is:

[0251]

[0252] If the false alarm rate increases, lower the threshold:

[0253] t 27 (v)=t 27 (u-1)- Δ t 27 (v)

[0254] After the v-th iteration, calculate the prediction accuracy, false alarm rate, and false negative rate of the fusion layer, compare them with the iteration termination condition, and proceed to the next update if the condition is not met, otherwise terminate the iteration and output the latest model parameters.

Claims

1. A method for automatically updating diagnostic rules, characterized in that: Includes the following steps: Extract data, including the diagnostic rules, diagnostic process output, diagnostic result output, and on-site maintenance feedback information of the PHM system. The diagnostic rules include the fusion decision layer and the initial diagnostic layer. Based on the diagnostic results output and on-site maintenance feedback, the accuracy p of the fusion diagnostic results of the fusion decision layer is statistically analyzed. c False alarm rate p e , false negative rate p l ; Analyze the factors influencing the comprehensive diagnostic results of the sub-models of the fusion decision layer; Set the conditions for automatic update triggering and iteration termination of diagnostic rules; When the automatic update trigger condition is met, the weights of the sub-models in the fusion decision layer are updated in combination with the influence factors; When the automatic update trigger condition is met, update the machine learning algorithm class model of the initial diagnostic layer and update the parameters of the threshold judgment class model of the initial diagnostic layer; The accuracy p of the updated fusion decision layer c False alarm rate p e , false negative rate p l The algorithm checks whether the updated value meets the termination condition of the iteration. If it does, the update ends; otherwise, it proceeds to the next update until the termination condition is met.

2. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The accuracy p of the fusion diagnostic results of the statistical fusion decision layer c False alarm rate p e , false negative rate p l The calculation method is as follows: the statistical period is set from the (u-1)th update to the uth update. ， ， ， Where TotalNum(u-1) refers to the sum of the number of early warning pushes from the PHM system and the number of missed early warnings from on-site maintenance feedback, N c (u-1) represents the number of accurate early warnings reported from on-site maintenance, N e (u-1) represents the number of on-site maintenance feedback warning errors, N l (u-1) represents the number of missed early warnings from on-site maintenance feedback; The automatic trigger condition for data statistics is when TotalNum(u-1) reaches the set threshold TotalNum(th).

3. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The calculation method for the influence factor of the sub-model of the fusion decision layer on the comprehensive diagnostic results is as follows: For the k-th accurately predicted data point, the prediction confidence scores of the initial diagnostic layer and the fusion decision layer are obtained through the diagnostic process output. These scores are then sorted from high to low. Assuming the initial diagnostic layer has M sub-models, and each sub-model corresponds to N candidate sub-health types, the prediction confidence matrix P(k) of the initial diagnostic layer's sub-models is obtained. ， Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, p ij (k)∈[0,1]; The confidence matrix R(k) of the fusion diagnosis results of the fusion decision layer is obtained. ， Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1]; The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k). ， Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is either 0 or 1; Calculate the consistency coefficient matrix A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results. ， Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, ； Calculate the confidence distance matrix D(k) between the sub-models of the initial diagnostic layer and each prediction type of the fused diagnostic results. ， Where 1≤k≤N c (u-1), 1≤i≤M, ； Calculate the rectangle F(k) representing the influence factor of the sub-model prediction results of the initial diagnostic layer on the fused diagnostic results. ， Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. ,in Based on confidence distance d i (k) is a function of the independent variable. It uses an exponential function with a base greater than 1 for calculation; Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results. ， Where 1≤i≤M, 。 4. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The automatic update trigger condition refers to: Let p be the initial accuracy of the fusion diagnostic results of the fusion decision layer. c (0) False alarm rate p e (0) Underreporting rate p l (0), let the thresholds for early warning accuracy, false alarm rate, and false negative rate be p, respectively. c (th), p e (th), p l (th); Set automatic update trigger conditions: p c (u-1) < p c (th), p e (u-1) > p e (th), p l (u-1) > p l (th), p c (u-1) < p c (u-2), p e (u-1) > p e (u-2), p l (u-1) > p l (u-2); When u=1, if any of the first three conditions are met, the fusion decision layer and the initial diagnosis layer will be updated. When u≥2, and any one of the six conditions is met, the fusion decision layer and the initial diagnosis layer are updated.

5. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The condition for terminating the iteration refers to: Let p be the accuracy of the fusion diagnosis results of the fusion decision layer after the v-th iteration before the u-th update calculation deadline. c (v) The false alarm rate is p e (v) The false negative rate is p l (v); Set the iteration termination condition: p c (v)≥p c (th), p e (v)≤p e (th), p l (v)≤p l (th); The iteration calculation is terminated when any one of the three conditions is met.

6. The method for automatically updating diagnostic rules according to claim 3, characterized in that: The formula for calculating the weights of the sub-models of the fusion decision layer by combining influence factors is as follows: Let W be the weight matrix of the sub-models of the fusion decision layer after the U-th update. U , ， Where M refers to the number of sub-models in the fusion decision layer. The weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M; Let the initial prediction accuracy of each sub-model be: ； Based on the initial prediction accuracy of each sub-model, set the initial weights for each sub-model: ， Wherein, the weight of the i-th sub-model ； The formula for calculating the weight value after the u-th update is as follows: 。 7. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The machine learning algorithm model for updating the initial diagnostic layer refers to adding positive true examples to the sample library, setting the model's training parameters to optimize the space, retraining the model using the new samples, and stopping training when the model's prediction accuracy increases or reaches a preset value, and then updating the model and its configuration parameters.

8. The method for automatically updating diagnostic rules according to claim 1, characterized in that: The parameters of the threshold judgment model for updating the initial diagnostic layer refer to: Let the confidence push threshold parameter for the j-th candidate sub-health type of the i-th sub-model be set before the u-th update calculation. The confidence probability distribution of the positive feedback error result is [a ij ,b ij ], then the threshold parameter adjustment step size calculated in the v-th iteration before the u-th update deadline is: ， in This is the adjustment parameter for the influence factor, and its value range is [0,1]. If the false negative rate is greater than the accuracy rate or the false negative rate increases, then lower the push threshold and set the upper and lower confidence limits for false negative sub-health type j corresponding to a and b. Adjust the threshold parameters accordingly. ; If the false alarm rate exceeds the accuracy rate or the false alarm rate increases, the push threshold will be increased, and the upper and lower confidence limits for false alarm results (a and b) will be set accordingly. .

9. A diagnostic rule automatic update system, characterized in that: include, The data acquisition module is used to extract diagnostic rules, diagnostic process outputs, diagnostic result outputs, and on-site maintenance feedback information from the PHM system. The statistics module is used to calculate the accuracy, false alarm rate, and false negative rate of the integrated decision-making layer of diagnostic rules based on the diagnostic results output and on-site maintenance feedback information. The impact factor calculation module is used to calculate the impact factors of the sub-models of the fusion decision layer on the comprehensive diagnostic results. The update module is used to set the conditions for automatic update triggering and iteration termination of diagnostic rules based on the early warning accuracy, false alarm rate and false negative rate. When the automatic update triggering conditions are met, the weights of the sub-models of the fusion decision layer are updated in combination with the influencing factors, and the parameters of the machine learning algorithm model and threshold judgment model of the initial diagnostic layer are updated until the iteration termination conditions are met.

10. The diagnostic rule automatic update system according to claim 9, characterized in that: The influencing factor is calculated as follows: For the k-th accurately predicted data point, the prediction confidence of the initial diagnostic layer and the fusion decision layer is obtained through the diagnostic process output. The prediction confidence of the initial diagnostic layer and the fusion decision layer is sorted from high to low. Assuming that the initial diagnostic layer has M sub-models and each sub-model corresponds to N candidate sub-health types, the prediction confidence matrix P(k) of the sub-models of the initial diagnostic layer is obtained. ， Where 1≤i≤M, 1≤j≤N, p ij ∈[0,1]; The confidence level R(k) of the fusion diagnosis results of the fusion decision layer is obtained. ， Where 1≤k≤N c (u-1), 1≤j≤N, q j (k)∈[0,1]; The consistency between the sub-model prediction results and the fusion diagnosis results of the initial diagnostic layer is calculated. Using R(k) as a sliding window, the confidence ranking numbers of candidate sub-health types are matched from top to bottom for P(k) to obtain the consistency matrix H(k). ， Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, h ij (k) is 0 or 1; Calculate the consistency coefficient A(k) between the sub-model of the initial diagnostic layer and the predicted type of the fused diagnostic results. ， Where 1≤k≤N c (u-1), 1≤i≤M, 1≤j≤N, ； Calculate the confidence distance D(k) between the sub-model of the initial diagnostic layer and each prediction type of the fusion diagnostic result. ， Where 1≤k≤N c (u-1), 1≤i≤M, ； Calculate the influence factor F(k) of the sub-model prediction results of the initial diagnostic layer on the fusion diagnostic results. ， Where 1≤k≤N c (u-1), 1≤i≤M, β i (k) represents the influence factor of the prediction result of the i-th sub-model corresponding to the k-th accurate prediction data on the fusion diagnosis result. ,in Based on confidence distance d i (k) is a function of the independent variable. It uses an exponential function with a base greater than 1 for calculation; Calculate the influence factor F(u-1) of the sub-model prediction results of the initial diagnostic layer before the u-th update on the fusion diagnostic results. ， Where 1≤i≤M, , The formula for calculating the weights of the sub-models of the fusion decision layer by combining influence factors is as follows: Let W be the weight matrix of the sub-models of the fusion decision layer after the U-th update. U , ， Where M refers to the number of sub-models in the fusion decision layer. The weight value of the i-th sub-model after the u-th update, 1≤u≤U, 1≤i≤M; Let the initial prediction accuracy of each sub-model be: ； Based on the initial prediction accuracy of each sub-model, set the initial weights for each sub-model: ， Wherein, the weight of the i-th sub-model ； The formula for calculating the weight value after the u-th update is as follows: 。

Images