A high-speed train suspension fault diagnosis method based on T-S fuzzy data-driven ToMFIR

Abstract

The application discloses a high-speed train suspension fault diagnosis method based on T-S fuzzy data-driven ToMFIR, which comprises the following steps: T-S fuzzy modeling is carried out on a train suspension system and a regular system, then T-S fuzzy models of the obtained real system and the regular system are subjected to data modeling to obtain data models of the real system and the regular system; the output of the real system and the regular system when the train suspension system is running is acquired by using an instrument, and input-output data matrices of the real system and the regular system are constructed; a fault information full-metric residual based on data driving is designed by using identification technology; an evaluation function is constructed, and an alarm is given when a detection index is greater than a threshold value; a suspension system sensor fault isolation algorithm is designed to isolate sensor faults; and a suspension system actuator fault isolation algorithm is designed to isolate actuator faults. The method can effectively detect and isolate faults when the suspension system fails, even when the failure is a slowly changing or intermittent small failure.

Description

Technical Field

[0001] This invention belongs to the field of high-speed train suspension system fault diagnosis technology, and relates to a high-speed train suspension fault diagnosis method based on TS fuzzy data driven ToMFIR. Background Technology

[0002] The suspension system of high-speed trains supports the car body and bogies, while isolating wheel-rail forces caused by track irregularities, ensuring the stability and safety of the train during high-speed operation. Therefore, it requires high reliability. Since the introduction of CRH (China Railway High-speed) EMU trains with speeds exceeding 200 km / h on major trunk lines in 2008, my country has built the world's largest and fastest high-speed railway network after more than a decade of development. Research on the detection of minute faults in the suspension system is of great significance for improving the safety of my country's high-speed rail operations.

[0003] High-speed train suspension systems are divided into active suspension and semi-active suspension, employing a closed-loop control structure. The suspension system uses a two-stage suspension system: the primary suspension is located between the axle box and the bogie frame, while the secondary suspension is located between the bogie frame and the car body. It comprises numerous components, including coil springs, lateral / vertical dampers, air springs, active actuators, and sensors. Among these, the active actuator, as a crucial actuator component, is vital for the safe operation and ride comfort of high-speed trains. The active actuator calculates the active control force required by the active suspension system based on vehicle output signals measured by sensors; therefore, the sensors are also of great importance.

[0004] Suspension system failures can be categorized into actuator failures, sensor failures, and mechanical component failures. As trains operate on the track for longer periods, some components in the suspension system, such as coil springs, shock absorbers, air springs, active / semi-active actuators, and sensors, will experience a certain degree of performance degradation. This can lead to minor faults such as micro-cracks in coil springs, slight oil / fluid leaks in shock absorbers, slight air leaks in air springs, small-amplitude loss of actuator actuation efficiency, and sensor drift, posing potential dangers to train operation safety.

[0005] Existing fault diagnosis methods for high-speed train suspension systems primarily target permanent faults, i.e., faults that do not disappear automatically after they occur. Since intermittent faults occur at unknown times, have random amplitudes, and disappear spontaneously after occurrence, traditional detection methods for permanent faults are difficult to directly apply to the diagnosis of intermittent faults. Furthermore, most fault diagnosis methods for high-speed train suspension systems neglect the nonlinearities within the suspension system, which affects the accuracy of fault diagnosis. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-speed train suspension fault diagnosis method based on TS fuzzy data driven ToMFIR. This method targets the nonlinear suspension system of high-speed trains with a closed-loop control structure, using active actuators and sensors as fault diagnosis objects. It can detect faults in real time and accurately isolate them when faults occur in the suspension system, even early, minor faults with slowly changing or intermittent characteristics.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0008] A high-speed train suspension fault diagnosis method based on TS fuzzy data-driven ToMFIR, using active actuators and sensors as fault diagnosis objects, includes the following steps:

[0009] Step 1. Perform TS fuzzy modeling on the high-speed train suspension system and its regularized system, and then perform data modeling on the obtained real system TS fuzzy model and regularized system TS fuzzy model to obtain the input-output data model of the real system and the input-output data model of the regularized system.

[0010] Step 2. Use displacement sensors and gyroscopes to acquire the vertical displacement and angular velocity signals of the center of mass of the train car and the vertical displacement signal of the center of mass of the bogie frame during high-speed train operation. The signals from the sensors and gyroscopes are the outputs of the real system. The obtained input and output signals of the real system are used to construct the input and output data matrix of the real system, and the obtained input and output signals of the regular system are used to construct the input and output data matrix of the regular system.

[0011] Step 3. Using the input-output data matrix of the real system and the input-output data matrix of the regular system obtained in Step 2, perform matrix identification and design a data-driven fault information full-measure residual.

[0012] Step 4. Construct an evaluation function J using the data-driven fault information full measurement residual designed in Step 3, and combine it with the fault detection alarm threshold to trigger an alarm when the detection index reaches the alarm threshold.

[0013] Step 5. Using the data-driven fault information full measurement residual designed in Step 3, design a suspension system sensor fault isolation algorithm to isolate sensor faults;

[0014] Step 6. Using the data-driven fault information full measurement residual designed in Step 3, design a suspension system actuator fault isolation algorithm to isolate actuator faults;

[0015] The specific process of step 1 includes:

[0016] Step 1.1. The discrete nonlinear real system G and the regular system G0 of the high-speed train suspension system can be expressed as follows:

[0017]

[0018]

[0019] Among them, A, B, C, E d E f ,F f Let f(k) be the coefficient matrix corresponding to the space state equation; x(k), g(·), u(k), d(k), y(k) are the state variables, nonlinear functions, control input variables, track disturbance excitations, and output variables of the real system, respectively; f(k) represents all possible faults. ξ(k) represents the process noise and measurement noise, respectively; x0(k), y0(k), and u0(k) represent the state variable, output variable, and control input variable of the canonical system, respectively.

[0020] Step 1.2. From the discrete nonlinear real system G and the regular system G0 of the suspension system in Step 1.1, we have:

[0021] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0022]

[0023]

[0024] Where θ(k)=[θ1(k) … θ p [(k)] represents the measurable premise variable; N represents the i-th fuzzy inference rule; j,i (j = 1, 2, ..., p) represents a fuzzy set; ν represents the number of inference rules;

[0025] The output of a fuzzy system can be inferred using standard fuzzy inference methods:

[0026]

[0027]

[0028] Where μ i (θ(k)) represents the fuzzy membership function, satisfying μ i(θ(k))≥0 and

[0029] Step 1.3. From the TS fuzzy system (3) of the actual suspension system in Step 1.2, we have:

[0030] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0031]

[0032] Where y l (k),u l (k),d l (k),f l (k), ξ l (k) is the stack matrix, Γ i,l H i,u,l H i,d,l H i,f,l , The corresponding coefficient matrix has the following specific form:

[0033]

[0034]

[0035]

[0036]

[0037] The expression (7) obtained in step 1.4.1.3 contains the state variable x(k). To eliminate the state variable x(k), we can obtain the TS fuzzy system (3) of the real system of the suspension system in step 1.2:

[0038]

[0039] in,

[0040]

[0041] Step 1.5. Substitute the expression (10) obtained in Step 1.4 into the expression (7) obtained in Step 1.3, and we have

[0042] Plant rule IF θ1(k) is N 1,iand θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0043]

[0044] in

[0045] H i,u,p,l =[Γ i,l Q i,u H i,u,l ],

[0046] H i,d,p,l =[Γ i,l Q i,d H i,d,l ],

[0047] H i,f,p,l =[Γ i,l Q i,f H i,f,l ],

[0048]

[0049] Step 1.6. Similarly, from the TS fuzzy system (4) of the suspension system regularization system in Step 1.2, we have:

[0050] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0051]

[0052] in,

[0053] Specifically, the process of designing the data-driven fault information full measurement residual in step 3 includes:

[0054] Step 3.1. Define the output residual r y (k) is used to characterize the output difference between the real system and the canonical system.

[0055]

[0056] Step 3.2. Define the controller residual ru (k) is used to characterize the output difference of the controller in the real system and the canonical system.

[0057]

[0058] Step 3.3. Define the fault information full measurement residual under the closed-loop control structure.

[0059]

[0060] in Characterizes the system output of a regularized system driven by a real-time input signal;

[0061] Step 3.4. Considering the time interval N, from the expression (11) obtained in step 1.5, we can get:

[0062] Y k,l =H i,u,p,l U k,p,l +H i,d,p,l D k,p,l +H i,f,p,l F k,p,l +H i,e,p,l E k,p,l (17)

[0063] Step 3.5. Perform LQ decomposition on the process data:

[0064]

[0065] achievable

[0066] Step 3.6. Based on the matrix H identified in Step 3.5 i,u,p,l Fuzzy data-driven ToMFIR can be written as:

[0067] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0068]

[0069] in Then, the global data-driven ToFMIR can be written as:

[0070]

[0071] in

[0072] Specifically, in step 4, the detection of minor faults in the suspension system is performed using the data-driven fault information full measurement residual. The process includes:

[0073] Step 4.1. Residual signal ToMFIR(k)∈R m It can be written as ToMFIR(k)=[τ1,…,τ m ] T Assuming ToMFIR(k) follows or approximately follows a Gaussian distribution, we have in

[0074] Step 4.2. Introduce fault-free residual signals ToMFIR rf (k) is the ToMFIR(k) at the fault-free time. Similarly, according to step 4.1, we have... in

[0075] Step 4.3. Based on the Jensen-Shannon divergence, the following evaluation function can be defined:

[0076]

[0077] Step 4.4. Utilizing the designed data-driven fault information full measurement residual and evaluation function, the fault detection mechanism for minor faults in the high-speed train suspension system is as follows:

[0078]

[0079] Specifically, in step 5, the residual of the full measurement of fault information based on data-driven analysis is used to isolate minor faults in the suspension system sensors. The process includes:

[0080] Step 5.1. Output residual r y,l (k)=[r y (k),…,r y (k+l f )] T It follows or approximately follows a Gaussian distribution, and has in

[0081] Step 5.2. When a fault is detected, collect fault operation data, and then establish a dynamic fault model for the output residuals as follows:

[0082]

[0083] in Indicates online data, express The fault-free part after reconstruction Represents the identity matrix and Ξ s The Kronecker product, where Let g represent the fault distribution matrix of the output residuals, where n represents the number of fault variables, and g i ∈{1,2,…,k y} represents the position of the i-th fault variable. identity matrix The i-th column;

[0084] The probability density function of the sample vector is:

[0085]

[0086] in

[0087] To estimate the size of the fault, the generalized least squares approach is established as follows:

[0088]

[0089] f s,l The maximum likelihood estimate of (k) is

[0090]

[0091] The reconstructed output residual can be written as

[0092]

[0093] The refactored global data-driven ToFMIR can be written as

[0094]

[0095] The reconstructed evaluation function is

[0096] J(ToMFIR * (k))=JS(ToMFIR * (k)‖ToMFIR rf (k)) (29)

[0097] Step 5.3. Initialization: Let n=0 indicates Ξ s The number of distribution vectors in the middle;

[0098] Step 5.4. for i = 1:k y -n, construct in It is the identity matrix The i-th column;

[0099] Step 5.5. Based on expressions (26)-(29) in Step 5.2, calculate the expression with... Relevant Reconstruction Statistics

[0100] Step 5.6. Insert into Ξ s In the middle, let n = n + 1;

[0101] Step 5.7. Calculate and Ξ s Relevant Reconstruction Statistics like Return to step 5.4;

[0102] Step 5.8. Based on the results obtained in Step 5.7 Isolate fault variables

[0103] Specifically, in step 6, the residual of the full measurement of fault information based on data-driven faults is used to isolate minor faults in the suspension system actuators. The process includes:

[0104] Step 6.1. Similar to the output residual, the controller residual follows a Gaussian distribution. in

[0105] Step 6.2. When a fault is detected, collect fault operation data, and then establish a dynamic fault model for the controller residuals as follows:

[0106]

[0107] in Indicates online data, express The fault-free part after reconstruction Represents the identity matrix and Ξ a The Kronecker product; where The fault distribution matrix represents the controller residuals, where n represents the number of fault variables, and g... i ∈{1,2,…,k y} represents the position of the i-th fault variable. identity matrix The i-th column;

[0108] The probability density function of the sample vector is:

[0109]

[0110] in

[0111] To estimate the size of the fault, the generalized least squares approach is established as follows:

[0112]

[0113] f a,p,l The maximum likelihood estimate of (k) is

[0114]

[0115] The reconstructed controller residual can be written as

[0116]

[0117] When the actuator fails, the output will also contain fault information; the reconstructed controller residuals can be used. The output residual corresponding to the reconstructed controller residual is represented by the matrix determined in step 3.5; according to expression (11) in step 1.5 and expression (13) in step 1.6, we have:

[0118] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0119]

[0120] The reconstructed fuzzy data-driven ToFMIR can be written as

[0121] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0122]

[0123] The refactored global data-driven ToFMIR can be written as

[0124]

[0125] The reconstructed evaluation function is

[0126] J(ToMFIR * (k))=JS(ToMFIR *(k)‖ToMFIR rf (k)) (38)

[0127] Step 6.3. Initialization: Let n=0 indicates Ξ a The number of distribution vectors in the middle;

[0128] Step 6.4. for i = 1:k u -n, construct in It is the identity matrix The i-th column;

[0129] Step 6.5. Calculate the expression (33)-(38) from Step 6.2. Relevant Reconstruction Statistics

[0130] Step 6.6. Insert into Ξ a In the middle, let n = n + 1;

[0131] Step 6.7. Calculate Ξ a Relevant Reconstruction Statistics like Return to step 6.4;

[0132] Step 6.8. Based on the results obtained in Step 6.7 Isolate fault controller variables

[0133] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0134] 1. The method of the present invention targets the suspension system of the closed-loop control structure of high-speed trains. It uses active actuators and sensors as the objects of fault diagnosis. When a fault occurs in the suspension system, even a minor early fault such as gradual change or intermittent fault, it can detect and accurately isolate the fault in real time.

[0135] 2. The fault diagnosis method of the present invention only relies on the input and output data of the actual system and the regular system of the train suspension system, and does not require knowledge of the precise system model. It is a data-driven fault diagnosis method.

[0136] 3. The fault diagnosis method of the present invention can collect more comprehensive fault information, making it more sensitive to minor faults in actuators and sensors of high-speed railway suspension systems, and can effectively and timely detect early minor faults such as gradual changes and intermittent faults.

[0137] 4. The fault diagnosis method of the present invention can effectively isolate minor faults in actuators and sensors of high-speed railway suspension systems.

[0138] 5. The fault diagnosis method of the present invention has high sensitivity to faults, even minor faults, in the high-speed train suspension system. It effectively solves the problems of diagnosing minor faults in the high-speed train suspension system under the closed-loop control structure and their engineering application. It is of great significance for the early warning and real-time monitoring of minor faults in the high-speed train suspension system. Attached Figure Description

[0139] Figure 1 This is a schematic diagram of the structure of a high-speed train suspension system fault diagnosis system according to an embodiment of the present invention.

[0140] Figure 2 This is a schematic diagram of the installation positions of sensors and gyroscopes in a high-speed train suspension system according to an embodiment of the present invention.

[0141] Figure 3 This is a flowchart of a fault diagnosis method according to an embodiment of the present invention.

[0142] Figure 4 This is a simulation curve of minor fault detection of a suspension sensor according to an embodiment of the present invention.

[0143] Figure 5 This is a simulation curve of a minor fault isolation of a suspension sensor according to an embodiment of the present invention.

[0144] Figure 6 This is a simulation curve of a suspension actuator for detecting minor faults according to an embodiment of the present invention.

[0145] Figure 7 This is a simulation curve of a suspension actuator with gradual micro-fault isolation according to an embodiment of the present invention.

[0146] Figure 8 This is a simulation curve of intermittent minor fault detection of a suspension actuator according to an embodiment of the present invention.

[0147] Figure 9 This is a simulation curve of intermittent minor fault isolation of a suspension actuator according to an embodiment of the present invention. Detailed Implementation

[0148] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0149] Figure 1 This is a schematic diagram of a high-speed train suspension system fault diagnosis system according to an embodiment of the present invention. Figure 1As shown, the fault diagnosis system includes: sensors for acquiring the vertical displacement and pitch angle of the train's center of gravity and the vertical displacement of the bogie frame's center of gravity; a data acquisition system for acquiring, processing, and sending the sensor data to the fault diagnosis host; and the fault diagnosis host for determining whether a fault has occurred in the high-speed train's suspension system based on the acquired information and estimating the fault amplitude. The sensor and gyroscope installation locations in each carriage are shown in the diagram. Figure 2 As shown.

[0150] like Figure 3 As shown, an embodiment of the present invention provides a high-speed train suspension fault diagnosis method based on TS fuzzy data driven ToMFIR, which uses active actuators and sensors as fault diagnosis objects, and includes the following steps:

[0151] Step 1. Perform TS fuzzy modeling on the high-speed train suspension system and its regularized system, and then perform data modeling on the obtained real system TS fuzzy model and regularized system TS fuzzy model to obtain the input-output data model of the real system and the input-output data model of the regularized system.

[0152] Step 2. Use displacement sensors and gyroscopes to acquire the vertical displacement and angular velocity signals of the center of mass of the train car and the vertical displacement signal of the center of mass of the bogie frame during high-speed train operation. The signals from the sensors and gyroscopes are the outputs of the real system. The obtained input and output signals of the real system are used to construct the input and output data matrix of the real system, and the obtained input and output signals of the regular system are used to construct the input and output data matrix of the regular system.

[0153] Step 3. Using the input-output data matrix of the real system and the input-output data matrix of the regular system obtained in Step 2, perform matrix identification and design a data-driven fault information full-measure residual.

[0154] Step 4. Construct an evaluation function J using the data-driven fault information full measurement residual designed in Step 3, and combine it with the fault detection alarm threshold to trigger an alarm when the detection index reaches the alarm threshold.

[0155] Step 5. Using the data-driven fault information full measurement residual designed in Step 3, design a suspension system sensor fault isolation algorithm to isolate sensor faults;

[0156] Step 6. Using the data-driven fault information full measurement residual designed in Step 3, design a suspension system actuator fault isolation algorithm to isolate actuator faults.

[0157] In step 1, the specific process of performing TS fuzzy modeling on the high-speed train suspension system and its regularized system, and then performing data modeling on the fuzzy modeling results of the real system TS fuzzy model and the regularized system TS fuzzy model, includes:

[0158] Step 1.1. According to Figure 2 The vertical suspension system shown can be represented by its discrete nonlinear real system G and regular system G0 as follows:

[0159]

[0160]

[0161] Among them, A, B, C, E d E f ,F f Let f(k) be the coefficient matrix corresponding to the space state equation; x(k), g(·), u(k), d(k), y(k) are the state variables, nonlinear functions, control input variables, track disturbance excitations, and output variables of the real system, respectively; f(k) represents all possible faults. ξ(k) represents the process noise and measurement noise, respectively. x0(k), y0(k), and u0(k) represent the state variable, output variable, and control input variable of the canonical system, respectively.

[0162] Step 1.2. From the discrete nonlinear real system G and the regular system G0 of the suspension system in Step 1.1, we have:

[0163] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0164]

[0165]

[0166] Where θ(k)=[θ1(k) … θ p [(k)] represents the measurable premise variable; N represents the i-th fuzzy inference rule; j,i (j=1,2,…,p) represents a fuzzy set. ν represents the number of inference rules.

[0167] Using standard fuzzy inference methods, the output of the fuzzy system can be inferred as follows:

[0168]

[0169]

[0170] Where μi (θ(k)) represents the fuzzy membership function, satisfying μ i (θ(k))≥0 and

[0171] Step 1.3. From the TS fuzzy system (3) of the actual suspension system in Step 1.2, we have:

[0172] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0173]

[0174] Where y l (k),u l (k),d l (k),f l (k), ξ l (k) is the stack matrix, Γ i,l H i,u,l H i,d,l H i,f,l , The corresponding coefficient matrix has the following specific form:

[0175]

[0176]

[0177]

[0178]

[0179] The expression (7) obtained in step 1.4.1.3 contains the state variable x(k). To eliminate the state variable x(k), we can obtain the TS fuzzy system (3) of the real system of the suspension system in step 1.2.

[0180]

[0181] in

[0182]

[0183] Step 1.5. Substitute the expression (10) obtained in Step 1.4 into the expression (7) obtained in Step 1.3, and we have

[0184] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0185]

[0186] in

[0187] H i,u,p,l =[Γ i,l Q i,u H i,u,l ],

[0188] H i,d,p,l =[Γ i,l Q i,d H i,d,l ],

[0189] H i,f,p,l =[Γ i,l Q i,f H i,f,l ],

[0190]

[0191] Step 1.6. Similarly, from the TS fuzzy system (4) of the suspension system regularization system in Step 1.2, we have:

[0192] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0193]

[0194] in,

[0195] In step 3, the specific process of designing the data-driven fault information full measurement residual includes:

[0196] Step 3.1. Define the output residual r y (k) is used to characterize the output difference between the real system and the canonical system.

[0197]

[0198] Step 3.2. Define the controller residual r u (k) is used to characterize the output difference of the controller in the real system and the canonical system.

[0199]

[0200] Step 3.3. Define the fault information full measurement residual under the closed-loop control structure.

[0201]

[0202] in Characterizes the output of a regular system driven by a real-time input signal.

[0203] Step 3.4. Considering the time interval N, from the expression (11) obtained in step 1.5, we can get:

[0204] Y k,l =H i,u,p,l U k,p,l +H i,d,p,l D k,p,l +H i,f,p,l F k,p,l +H i,e,p,l E k,p,l (17)

[0205] Step 3.5. Perform LQ decomposition on the process data:

[0206]

[0207] achievable

[0208] Step 3.6. Based on the matrix H identified in Step 3.5 i,u,p,l Fuzzy data-driven ToMFIR can be written as:

[0209] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0210]

[0211] in Then, the global data-driven ToFMIR can be written as:

[0212]

[0213] in

[0214] In step 4, the residual of the full measurement of fault information based on data-driven analysis is used to detect minor faults in the suspension system. The specific steps include:

[0215] Step 4.1. Residual signal ToMFIR(k)∈R m It can be written as ToMFIR(k)=[τ1,…,τ m ] T Assuming ToMFIR(k) follows or approximately follows a Gaussian distribution, we have in

[0216] Step 4.2. Introduce fault-free residual signals ToMFIR rf (k) is the ToMFIR(k) at the fault-free time. Similarly, according to step 4.1, we have... in

[0217] Step 4.3. Based on the Jensen-Shannon divergence, the following evaluation function can be defined:

[0218]

[0219] Step 4.4. Utilizing the designed data-driven fault information full measurement residual and evaluation function, the fault detection mechanism for minor faults in the high-speed train suspension system is as follows:

[0220]

[0221] In step 5, the residual of the full measurement based on data-driven fault information is used to isolate minor faults in the suspension system sensors. The specific steps include:

[0222] Step 5.1. Output residual r y,l (k)=[r y (k),…,r y (k+l f )] T It follows or approximately follows a Gaussian distribution, and has in

[0223] Step 5.2. When a fault is detected, collect fault operation data, and then establish a dynamic fault model for the output residuals as follows:

[0224]

[0225] in Indicates online data, express The fault-free part after reconstruction Represents the identity matrix and Ξ s The Kronecker product, where Let g represent the fault distribution matrix of the output residuals, where n represents the number of fault variables, and g i ∈{1,2,…,k y} represents the position of the i-th fault variable. identity matrix The i-th column.

[0226] The probability density function of the sample vector is:

[0227]

[0228] in

[0229] To estimate the size of the fault, the generalized least squares approach is established as follows:

[0230]

[0231] f s,l The maximum likelihood estimate of (k) is

[0232]

[0233] The reconstructed output residual can be written as

[0234]

[0235] The refactored global data-driven ToFMIR can be written as

[0236]

[0237] The reconstructed evaluation function is

[0238] J(ToMFIR * (k))=JS(ToMFIR * (k)‖ToMFIR rf (k)) (29)

[0239] Step 5.3. Initialization: Let n=0 indicates Ξ s The number of distribution vectors in the middle.

[0240] Step 5.4. for i = 1:k y -n, construct in It is the identity matrix The i-th column.

[0241] Step 5.5. Based on expressions (26)-(29) in Step 5.2, calculate the expression with... Relevant Reconstruction Statistics

[0242] Step 5.6. Insert into Ξ s In the middle, let n = n + 1;

[0243] Step 5.7. Calculate and Ξ s Relevant Reconstruction Statistics like Return to step 5.4.

[0244] Step 5.8. Based on the results obtained in Step 5.7 Isolate fault variables

[0245] In step 6, the residuals of the full measurement based on data-driven fault information are used to isolate minor faults in the suspension system actuators. The specific steps include:

[0246] Step 6.1. Similar to the output residuals, the controller residuals follow a Gaussian distribution. in

[0247] Step 6.2. When a fault is detected, collect fault operation data, and then establish a dynamic fault model for the controller residuals as follows:

[0248]

[0249] in Indicates online data, express The fault-free part after reconstruction Represents the identity matrix and Ξ a The Kronecker product, where Let g represent the fault distribution matrix of the controller residuals, where n represents the number of fault variables, and g i ∈{1,2,…,k y} represents the position of the i-th fault variable. identity matrix The i-th column.

[0250] The probability density function of the sample vector is:

[0251]

[0252] in

[0253] To estimate the size of the fault, the generalized least squares approach is established as follows:

[0254]

[0255] f a,p,l The maximum likelihood estimate of (k) is

[0256]

[0257] The reconstructed controller residual can be written as

[0258]

[0259] When an actuator fails, the output will also contain fault information. The reconfigured controller residuals can be used. The matrix determined in step 3.5 is used to represent the output residual corresponding to the reconstructed controller residual. Based on expression (11) in step 1.5 and expression (13) in step 1.6, we have:

[0260] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0261]

[0262] The reconstructed fuzzy data-driven ToFMIR can be written as

[0263] Plant rule IF θ1(k) is N 1,i and θ2(k) is N 2,i and … and θ p (k) is N p,i THEN

[0264]

[0265] The refactored global data-driven ToFMIR can be written as

[0266]

[0267] The reconstructed evaluation function is

[0268] J(ToMFIR * (k))=JS(ToMFIR * (k)‖ToMFIR rf (k)) (38)

[0269] Step 6.3. Initialization: Let n=0 indicates Ξ a The number of distribution vectors in the middle.

[0270] Step 6.4. for i = 1:k u -n, construct in It is the identity matrix The i-th column.

[0271] Step 6.5. Calculate the expression (33)-(38) from Step 6.2. Relevant Reconstruction Statistics

[0272] Step 6.6. Insert into Ξ a In the middle, let n = n + 1;

[0273] Step 6.7. Calculate Ξ a Relevant Reconstruction Statistics like Return to step 6.4.

[0274] Step 6.8. Based on the results obtained in Step 6.7 Isolate fault controller variables

[0275] The method of the present invention will be verified by simulation below.

[0276] Step 1. Set the characteristic information of sensor drift fault, including: the carriage sensor has a slow drift micro fault, the drift amount is θ(t)=0.01×(0.25+0.02sin(t)+0.01sin(0.2t)), the fault start time is the 60th second, the fault end time is the simulation end time, the fault is injected into the software through the fault injection module to establish the fault model;

[0277] Step 2. Set the characteristic information of the actuator gradual failure, including: 10% gradual failure of the front bogie actuator, the failure start time is the 60th second, and the failure end time is the simulation end time. The fault is injected into the software through the fault injection module to establish a fault model.

[0278] Step 3. Set the characteristic information of the actuator intermittent failure, including: 10% failure of the front bogie actuator, the failure time is from the 20th to the 30th second and from the 60th to the 80th second, the failure is injected into the software through the fault injection module to establish a fault model;

[0279] Step 4. Use SIMPACK and Matlab / Simulink for co-simulation. Import the track-train coupling model built in SIMPACK into Matlab / Simulink, and build a simulation model of the train's vertical suspension system control in Simulink. Set the vehicle's running speed to 250 km / h and the simulation duration to 100 seconds.

[0280] like Figure 4 As shown, when a minor fault of gradual drift occurs in the sensor of the high-speed train suspension system, the method proposed in this invention can effectively detect the occurrence of such a minor fault.

[0281] like Figure 5 As shown, when a minor fault of gradual drift occurs in the sensor of the high-speed train suspension system, the method proposed in this invention can effectively isolate the minor fault of the sensor of the train suspension system.

[0282] like Figure 6 As shown, when a slow-change minor fault occurs in the actuator of the high-speed train suspension system, the method proposed in this invention can effectively detect the occurrence of the slow-change minor fault in the actuator of the train suspension system.

[0283] like Figure 7 As shown, when a slow-changing minor fault occurs in the actuator of the high-speed train suspension system, the method proposed in this invention can effectively isolate the slow-changing minor fault in the actuator of the train suspension system.

[0284] like Figure 8 As shown, when intermittent minor faults occur in the actuators of the high-speed train suspension system, the method proposed in this invention can effectively detect the occurrence of intermittent minor faults in the actuators of the train suspension system.

[0285] like Figure 9 As shown, when intermittent minor faults occur in the actuators of the high-speed train suspension system, the method proposed in this invention can effectively isolate these intermittent minor faults.

[0286] The fault detection method of this invention has high sensitivity to minor faults in the high-speed train suspension system. It can effectively detect and isolate minor faults such as gradual changes and intermittent faults in the high-speed train suspension system. It effectively solves the problems of detecting minor faults under the closed-loop control structure and their engineering application. This is of great significance for the early warning and real-time monitoring of minor faults in the high-speed train suspension system.

Claims

1. A high-speed train suspension fault diagnosis method based on T-S fuzzy data-driven ToMFIR, characterized in that, The active actuator and sensor are taken as the fault diagnosis objects, and the method comprises the following steps: Step 1. T-S fuzzy modeling is conducted on the high-speed train suspension system and its regular system, and then data modeling is conducted on the obtained T-S fuzzy model of the real system and the T-S fuzzy model of the regular system to obtain the input-output data model of the real system and the input-output data model of the regular system, and the specific process comprises: Step 1.

1. The discrete nonlinear real system G and the regular system G0 of the high-speed train suspension system can be represented as: where A, B, C, E d , F f f are the coefficient matrices of the spatial state equation; x(k), g(·), u(k), d(k), y(k) are the state variable, nonlinear function, control input variable, track disturbance excitation, and output variable of the real system, respectively; f(k) represents all possible faults; ξ(k) are the process noise and measurement noise, respectively; x0(k), y0(k), u0(k) are the state variable, output variable, and control input variable of the regular system, respectively.​ Step 1.

2. According to the discrete nonlinear real system G and the regular system G0 of the suspension system in step 1.1, there are: where θ(k) = [θ1(k)... θN(k)] represents the measurable premise variables; p (k) represents the measurable premise variables; represents the ith fuzzy inference rule; N j,i represents the fuzzy set, where j = 1, 2,..., v, and v represents the number of inference rules. Through the standard fuzzy reasoning method, the output of the fuzzy system can be inferred as: where μ i (θ(k)) denotes the fuzzy membership function, satisfying μ i (θ(k)) ≥ 0 and Step 1.

3. According to the T-S fuzzy system (3) of the real system of the suspension system in step 1.2, there are: where y l (k),u l (k),d l (k),f l (k), ξ l (k) is a stack matrix, Γ i,l ,H i,u,l ,H i,d,l ,H i,f,l , are the respective coefficient matrices, which have the following specific forms: Step 1.

4. In order to eliminate the state variable x(k), according to the T-S fuzzy system (3) of the real system of the suspension system in step 1.2, there are: Wherein, Step 1.

5. The expression (10) obtained in step 1.4 is substituted into the expression (7) obtained in step 1.3, and there are Wherein Step 1.

6. Similarly, according to the T-S fuzzy system (4) of the regular system of the suspension system in step 1.2, there are: wherein, Step 2. The vertical displacement and angular velocity signals of the train car mass center and the vertical displacement signals of the bogie frame mass center are obtained by using the displacement sensor and the gyroscope when the high-speed train is running, and the signals of the sensor and the gyroscope are the output of the real system; the input-output signals of the real system are used to construct the input-output data matrix of the real system, and the input-output signals of the regular system are used to construct the input-output data matrix of the regular system; Step 3. The input-output data matrix of the real system and the input-output data matrix of the regular system obtained in step 2 are used for matrix identification, and a fault information full-metric residual based on data driving is designed, and the specific process comprises: Step 3.

1. Define output residual r y,l (k) for characterizing the output difference of real and regular systems Step 3.

2. Definition of controller residual r u,p,l (k) for characterizing the output difference of controllers in real and regular systems Step 3.

3. The fault information full-metric residual under the closed-loop control structure is defined wherein characterizing the system output of the canonical system under real-time input signal driving; Step 3.

4. Considering the time interval N, according to the expression (11) obtained in step 1.5, there are: Y k,l = H i,u,p,l U k,p,l + H i,d,p,l D k,p,l + H i,f,p,l F k,p,l + H i,e,p,l E k,p,l (17) Step 3.

5. The process data is decomposed by LQ: available Step 3.

6. Matrix H is identified based on the recognition of step 3.5 i,u,p,l The fuzzy data-driven ToMFIR can be written as: wherein Then, the global data-driven ToFMIR can be written as: wherein μ i is the fuzzy membership degree; Step 4. An evaluation function J is constructed by using the fault information full-metric residual based on data driving designed in step 3, and an alarm is given when the detection index reaches the alarm threshold, wherein the fault information full-metric residual based on data driving is used for detecting the small fault of the suspension system, and the specific steps comprise: Step 4.

1. The residual signal ToMFIR(k) e R m may be written as ToMFIR(k) = [τ1,..., τ m ] T ; assuming that ToMFIR(k) is Gaussian or approximately Gaussian, there is where Step 4.

2. Introducing the fault-free residual signal ToMFIR rf (k) is the ToMFIR(k) for the fault-free instant, according to Step 4.1, and by the same reasoning we have where Step 4.

3. Based on the Jensen-Shannon divergence, the following evaluation function can be defined: Step 4.

4. By using the fault information full-metric residual based on data driving and the evaluation function, the fault detection mechanism for the small fault of the high-speed train suspension system is Step 5. The fault isolation algorithm of the suspension system sensor is designed by using the fault information full-metric residual based on data driving designed in step 3 to isolate the sensor fault, and the specific steps comprise: Step 5.

1. Output residual r y,l (k) = [r y (k),…, r y (k+L f )] T Subject to or approximately subject to a Gaussian distribution, there is where Step 5.

2. When the fault is detected, collect the fault run data, and then build the dynamic fault model of output residual as: wherein represents the online data, represents the reconstructed non-faulty part, represents the identity matrix and Ξ s is the Kronecker product of represents the fault distribution matrix of the output residual, where n represents the number of fault variables, g i ∈ {1, 2, …, k y} represents the position of the i-th fault variable, is the i-th column of the identity matrix I The probability density function of sample vector is: wherein To estimate the size of the fault, the generalized least squares is built as follows f s,l The maximum likelihood estimate of (k) is The reconstructed output residual can be written as The reconstructed global data-driven ToFMIR can be written as The reconstructed evaluation function is J(ToMFIR * (k)) = JS(ToMFIR * (k)‖ToMFIR rf (k)) (29) Step 5.

3. Initialization: Let n = 0 represents the number of distribution vectors in Ξ s Ξ Step 5.

4. for i = 1 :k y - n, construct wherein is the identity matrix the i-th column of Step 5.

5. Calculate the reconstruction statistics associated with the expression (26) - (29) in Step 5.

2. ​ Step 5.

6. Insert into Ξ s n = n + 1; Step 5.

7. Compute the statistics related to Ξ s reconstruction statistics If Go to Step 5.4; Step 5.

8. Based on the results obtained in step 5.7 Isolating fault variables Step 6. Use the data-driven fault information full metric residual designed in step 3 to design the actuator fault isolation algorithm of the suspension system to isolate the actuator fault, the specific steps include: Step 6.

1. As with the output residuals, the controller residuals are subject to a Gaussian distribution where Step 6.

2. When the fault is detected, collect the fault run data, and then build the dynamic fault model of controller residual as: wherein represents online data, represents reconstructed non-faulty part, represents identity matrix and Ξ a is the Kronecker product of and ; wherein represents the fault distribution matrix of controller residual, n represents the number of fault variables, g i ∈ {1, 2, …, k u} represents the position of the i-th fault variable, is the i-th column of identity matrix . The probability density function of sample vector is: wherein To estimate the size of the fault, the generalized least squares is built as follows f a,p,l The maximum likelihood estimate of (k) is The reconstructed controller residual can be written as When the actuator fails, the output will also contain the failure information; the reconstructed controller residual can be used and the identified matrix and to represent the output residual corresponding to the reconstructed controller residual; according to expression (11) in step 1.5 and expression (13) in step 1.6, we have: The reconstructed fuzzy data-driven ToFMIR can be written as The reconstructed global data-driven ToFMIR can be written as The reconstructed evaluation function is J(ToMFIR * (k)) = JS(ToMFIR * (k) || ToMFIR rf (k)) (38) Step 6.

3. Initialization: Let n = 0 represent the number of distribution vectors in Ξ a . Step 6.

4. for i = 1 :k u - n, construct wherein is the i-th column of the identity matrix I Step 6.

5. Calculate the reconstruction statistics associated with the expression (33) - (38) in Step 6.

2. ​ Step 6.

6. Set n = n + 1. Insert into S a n = n + 1; Step 6.

7. Compute the reconfiguration statistics associated with Ξ a If Go to Step 6.4;​ Step 6.

8. Based on the results obtained in step 6.7 Isolating fault controller variables

Images