Method for predicting service life of steer-by-wire system

Abstract

The invention discloses a steering-by-wire system life prediction method, which comprises the following steps: S1, determining a degradation characteristic influencing the life of a steering-by-wire system, and collecting a state monitoring quantity related to the degradation characteristic; s2, analyzing the correlation between the state monitoring quantity and the residual life, calculating the correlation between each state monitoring quantity and the degradation characteristic quantity, and screening the state monitoring quantity capable of reflecting the aging trend of the steer-by-wire system; s3, performing noise reduction on the screened state monitoring quantity by using singular value decomposition; s4, decomposing the degradation characteristic quantity into an eigenmode component and a residual component by using CEEMDAN; s5, global degradation trend modeling is carried out by inputting the eigenmode component into the MHA-BiLSTM model; s6, inputting the residual component into a hidden Markov model to carry out local fluctuation modeling; and S7, fusing prediction outputs of the global degradation trend model and the local fluctuation model. And local fluctuation caused by uncertainty is predicted by using a hidden Markov model, and finally, prediction output is fused to realize life prediction of the steering-by-wire system.

Description

Technical Field

[0001] The present invention relates to the technical field of wire-controlled steering systems, and in particular to a method for predicting the life of wire-controlled steering systems. Background Art

[0002] As a steering technology originated from aircraft, the steer-by-wire technology has great advantages in steering performance, space occupation, vehicle comfort, etc. due to the elimination of the mechanical connection between the steering wheel and the steering wheel. However, vehicles using steer-by-wire systems have added a large number of electrical components, increased system complexity, and increased risk of vehicle loss of control, which may eventually lead to catastrophic consequences. Therefore, based on functional safety analysis, studying the life prediction of steer-by-wire systems has become an urgent problem to be solved. In the prior art, the methods for predicting the life of steer-by-wire systems are mainly divided into data-driven and model-driven. The model-driven method relies on the mechanism degradation model, but in reality, the degradation process of the components of the steer-by-wire system is very complex, and it is difficult to establish an accurate mechanism model. The data-driven method relies on the original data to establish the behavior model. Although it can accurately construct the mapping relationship from massive degradation data to life information, the data-driven method is very dependent on high-quality training data. Due to the uncertainty of the actual working conditions, the data degradation process has a large randomness, which affects the prediction performance. Summary of the invention

[0003] In view of the technical problem that the data degradation process has great randomness under actual working conditions, the present invention proposes a life prediction method for a steer-by-wire system. MHA-BiLSTM is used to predict the global degradation trend, and a hidden Markov model is used to predict the local fluctuation caused by uncertainty. Finally, the prediction output is integrated to realize the life prediction of the steer-by-wire system.

[0004] The technical solution adopted by the present invention is as follows: A method for predicting the life of a wire-controlled steering system comprises the following steps:

[0005] S1, determining a degradation feature that affects the life of the steer-by-wire system, and collecting a state monitoring quantity and a degradation feature quantity related to the degradation feature;

[0006] S2, analyzing the correlation between the state monitoring quantity and the remaining life, calculating the correlation between each state monitoring quantity and the degradation characteristic quantity, and selecting the state monitoring quantity that can reflect the aging trend of the wire control steering system;

[0007] S3, using singular value decomposition to reduce the noise of the filtered state monitoring quantity;

[0008] S4, using CEEMDAN to decompose the degradation feature into eigenmode components and residual components, where the eigenmode components reflect the global degradation trend and the residual components reflect the local fluctuations caused by uncertainty;

[0009] S5, global degradation trend modeling is performed by inputting the eigenmode components into the MHA-BiLSTM model to predict the global degradation trend;

[0010] S6, local fluctuation modeling by inputting the residual component into a hidden Markov model to predict local fluctuations;

[0011] S7, the prediction output of the global degradation trend model and the prediction output of the local fluctuation model are integrated to achieve the prediction of the steering system degradation process.

[0012] Optionally, the step S1 specifically includes the following steps:

[0013] The state monitoring quantity of the degradation characteristics of the steer-by-wire system is collected. Assuming that the steer-by-wire system has n state monitoring quantities and monitors t time moments in total, the state monitoring quantity x is expressed as:

[0014] x={x1,x2,...,x n}

[0015] Among them, x i =[x i1 ,x i2 ,...,x it ] T , i = 1, 2, ..., n, that is, the state monitoring quantity x can be expressed as:

[0016]

[0017] Determine the degradation characteristic quantity corresponding to the state monitoring quantity of the wire control steering system as the accuracy of the steering system, collect degradation data, and express it as: Y = [y1, y2, ..., y t ] T

[0018] Among them, yj represents the accuracy value of the wire-controlled steering system at time j, j=1, 2,..., t.

[0019] Optionally, step S2 specifically includes the following steps:

[0020] S21, calculate the correlation between each state monitoring quantity and the remaining life, and the calculation formula of the correlation coefficient R is:

[0021]

[0022] Among them, x i is the i-th state monitoring sequence, is the state monitoring sequence x i The mean of y and x i The corresponding degenerate characteristic quantity is is the mean of the time series y;

[0023] S22, screening state monitoring quantities with strong correlation: when the correlation coefficient is closer to 1, the correlation between the state monitoring quantity and the degradation characteristic quantity is stronger, based on which the critical range of the correlation coefficient r is determined to be 0.9 to 1, and the state monitoring quantity within this range is selected.

[0024] Optionally, step S3 specifically includes the following steps:

[0025] S31, based on the phase space reconstruction theory, the state monitoring quantity is constructed into an m×n Hankel matrix H, and the matrix is ​​subjected to singular value decomposition to obtain the mutually unrelated noise subspace and signal subspace, the formula is:

[0026]

[0027] Where U represents an M×N orthogonal matrix; D represents a diagonal element; D=diag(σ1,σ2,σ3......σ n ) represents an M×N diagonal matrix; V represents an N×N orthogonal matrix;

[0028] S32, find the singular values ​​of the Hankel matrix A, retain the singular values ​​of the diagonal matrix that reflect the effective signal, set the smaller singular values ​​that reflect the noise signal to zero, preliminarily remove the noise in the high-frequency signal, and then perform the inverse process of singular value decomposition to finally obtain the matrix Then the matrix It is the best approximation moment of H with rank r, and the calculation formula is:

[0029]

[0030] Where σi represents the singular value of the matrix A; r represents the rank of the matrix H, T represents the transpose; ui represents the i-th eigenvector of the matrix AAT; vi represents the i-th eigenvector of the matrix AT A;

[0031] The matrix The anti-diagonal elements in are added and averaged to obtain the denoised state monitoring quantity.

[0032] Optionally, the step S4 specifically includes the following steps:

[0033] S41, Gaussian white noise ε0E1(ω i (t)) is added to the signal to be decomposed x(t), and then empirical mode decomposition is performed to obtain the first eigenmode component IMF1 i (t), for IMF1 i (t) is averaged to obtain the first-order eigenmode component IMF1(t), whose formula is:

[0034]

[0035] Where N is the number of eigenmode components, ω i (t) is the white noise added for the ith time, ε i is the standard deviation of white noise;

[0036] S42, calculate the first-order residual component r1(t), the formula is:

[0037] r1(t)=x(t)-IMF1(t)

[0038] S43, continue to add Gaussian white noise to r1(t), and the new signal r1(t)+ε1E1(ω i (t)) to perform empirical mode decomposition and obtain the second-order eigenmode component:

[0039]

[0040] Among them, E1(r1(t)+ε1E1(w i )) represents an empirical mode decomposition;

[0041] S44, and so on, we get the kth order residual r k (t), the formula is:

[0042] r k (t) = r k-1 (t)-IMF k (t)

[0043] S45, calculate the k+1th order eigenmode component as:

[0044]

[0045] The signal is decomposed cyclically until the residual obtained is a monotonic function and cannot be further decomposed. The final residual component formula is:

[0046]

[0047] The obtained eigenmode components It reflects the global degradation trend, and the residual component r(t) reflects the local fluctuation caused by uncertainty.

[0048] Optionally, the step S5 specifically includes the following steps:

[0049] S51, build a multi-head attention module, specifically:

[0050] S511, assuming that the tensor input to the module is input, after the linear transformation layer linear, map three tensors of the same size as the input: Q, K, V. These three tensors are put into the scaled dot product attention calculation part to obtain the attention weights and the reconstructed attention matrix, and the calculation formula is:

[0051]

[0052] S512, divide the tensors Q, K, and V into h equal-sized fragments in the same direction (usually the last dimension of the tensor), and calculate the attention matrix by placing each fragment of the tensor in a different subspace. After the calculation is completed, the matrices of these subspaces are stitched together in the division direction to obtain the multi-head attention matrix, which is calculated as follows:

[0053] Multihead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0054] Among them, head i is the attention matrix calculated in the i subspace, Concat is the concatenation operation of these sub-attention matrices in the split direction, and W O is the parameter matrix for linear transformation;

[0055] S52, embed the multi-head attention module into the LSTM unit structure based on BiLSTM, and its calculation formula is as follows:

[0056] f(t)=MultiHead(σ(w f x(t)+w f h(t)+b f ))

[0057] i(t)=MultiHead(σ(w i x(t)+w i h(t-1)+b i ))

[0058] o(t)=σ(w o x(t)+w o h(t-1)+b o )

[0059]

[0060] h(t)=o(t)⊙tanh(c(t))

[0061] Among them, f(t), i(t), and o(t) are the outputs of the forget gate, input gate, and output gate respectively. Since the forget gate and input gate structure contain multi-head attention modules, the outputs f(t) and i(t) become tensors after attention reconstruction and participate in the calculation of the current state c(t) and output state h(t); x(t) is the input layer at time t, that is, the state monitoring quantity obtained after singular value decomposition in step S3; h(t-1) is the hidden layer at time t-1; h(t) is the hidden layer at time t; w f 、w i 、w o 、w c are the weight matrices of the forget gate, input gate, output gate, and hidden layer respectively; b f 、b i 、b o 、b c are the forget gate, input gate, output gate, and bias term of the hidden layer respectively; σ and tanh are both activation functions, σ is the sigmoid function, and tanh is the hyperbolic tangent function; in the BiLSTM structure, the attention module exists in both directions of the information flow to adjust the weight of the degraded information;

[0062] S53, when training the prediction model, the root mean square error function is selected, and the formula is as follows:

[0063]

[0064] in, Represents the predicted value of the eigenmode component in the i-th data to be predicted, y i represents the actual value of the intrinsic mode component in the i-th data to be predicted, n is the number of data, the network is initialized by randomly initializing weights and biases, and the gradient descent algorithm is used to update the parameters until the loss function converges to obtain the final prediction model, that is, the global degradation trend model;

[0065] S54, inputting the new state monitoring value into the global degradation trend model after preprocessing, and outputting the eigenmode component reflecting the global degradation trend.

[0066] Optionally, step S6 specifically includes the following steps:

[0067] S61, the data sequence r={r1, r2, ..., r n} is represented as a set of observed states, and the data sequence of the state monitoring quantity obtained in step S3 is x={x1,x2,…,x n} is represented as a set of hidden states, where n represents the number of state values;

[0068] S62, the model parameters are estimated using the Baum-Welch algorithm, the formula is as follows:

[0069] λ=(A,B,P)

[0070] Among them, λ is the local fluctuation prediction model, A={a ij} N×N is the state transfer matrix, B is the probability of output observation, and P is the probability of the initial state;

[0071] The training parameters P and A are solved by the Baum-Welch algorithm to make the expected P(r t |λ) is the maximum; calculate the output observation probability B, the formula is as follows:

[0072]

[0073] Among them, j represents the state, K represents the number of Gaussian distributions, φ jl represents the weight of the lth Gaussian component, μ jl and Σ jl are the mean and covariance values ​​respectively;

[0074] S63, using the Viterbi algorithm to find the optimal hidden state sequence x corresponding to the observation sequence t , specifically:

[0075] S631, the calculation formula is as follows:

[0076] δ1(i)=p i b i (r1),1≤i≤N

[0077] ψ1(i)=0,1≤i≤N

[0078] Where N is the number of states, p i and b i is the i-th element of the model parameters P and B, δ1(i) and ψ1(i) are intermediate variables;

[0079] S632, recursively calculate the maximum probability of all single paths and their nodes, and the calculation formula is as follows:

[0080]

[0081] Where N is the number of states, δ t (i) is the total number of δ at the last moment. t-1 (i) with element a in A ij and element b in B i (r t ) product, where 1≤j≤N, ψ t (i) is to make δt-1 (i)a ij The index of the maximum correspondence;

[0082] S633, terminate the iteration, the calculation formula is as follows:

[0083]

[0084] Where N is the number of states and P* is the number of all ψ t The maximum value of the elements in (i), x T * is to make ψ t (i) the index of the maximum correspondence;

[0085] S634, perform optimal path backtracking to obtain the final optimal hidden state sequence, the calculation formula is as follows:

[0086]

[0087] Among them, x t * is the optimal hidden state sequence;

[0088] S64, input the new state monitoring value into the local fluctuation prediction model after preprocessing, and output the local fluctuation.

[0089] Optionally, step S7 specifically includes the following steps:

[0090] S71, using weighted fusion method to process the prediction output of global degradation trend model and uncertainty model to achieve degradation process modeling;

[0091] S72, prediction time t k The remaining life L k , the calculation formula is as follows:

[0092] L k = inf{l k :X(l k +t k )>w}

[0093] Where X is the degradation data value at a certain moment, w is the failure threshold, l k Indicates the number of remaining life moments.

[0094] The beneficial effects of the present invention are as follows: in response to the uncertainty problem existing in the data-driven model, the present invention proposes a life prediction method for a wire-controlled steering system that combines a multi-head attention mechanism-bidirectional long short-term memory network MHA-BiLSTM (Multi-Head Attention-Bi-directional Long Short-Term Memory) and a hidden Markov model that considers quantified uncertainty. The degradation data is decomposed into global degradation trends and local fluctuations through the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). Accordingly, MHA-BiLSTM is used to predict the global degradation trend, and the hidden Markov model is used to predict the local fluctuations caused by uncertainty, and finally the prediction output is fused to realize the life prediction of the wire-controlled steering system. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 A principle flow chart of a life prediction method for a steer-by-wire system according to an embodiment of the present invention;

[0096] Figure 2 A multi-head attention calculation flow chart of the life prediction method of the wire control steering system proposed in an embodiment of the present invention;

[0097] Figure 3 A schematic diagram of a BiLSTM neural network embedded with multi-head attention for a life prediction method for a wire-controlled steering system proposed in an embodiment of the present invention;

[0098] Figure 4 This is a hidden Markov model prediction flow chart of the life prediction method of the steer-by-wire system proposed in an embodiment of the present invention.

[0099] The symbols in the drawings are: DETAILED DESCRIPTION

[0100] The present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0101] like Figure 1-4 As shown, this embodiment discloses a method for predicting the life of a wire-controlled steering system, comprising the following steps:

[0102] S1, determining the degradation characteristics that affect the life of the steer-by-wire system, and collecting the state monitoring quantity and degradation characteristic quantity related to the degradation characteristics, specifically:

[0103] The state monitoring quantity of the degradation characteristics of the steer-by-wire system is collected. Assuming that the steer-by-wire system has n state monitoring quantities and monitors t time moments in total, the state monitoring quantity x is expressed as:

[0104] x={x1,x2,...,x n}

[0105] Among them, x i =[x i1 ,x i2 ,...,x it ] T , i = 1, 2, ..., n, that is, the state monitoring quantity x can be expressed as:

[0106]

[0107] Determine the degradation characteristic quantity corresponding to the state monitoring quantity of the wire control steering system as the accuracy of the steering system, collect degradation data, and express it as: Y = [y1, y2, ..., y t ] T

[0108] Among them, yj represents the accuracy value of the wire control steering system at time j, j = 1, 2, ..., t. The n state monitoring quantities come from the steering wheel angle, the steering wheel road sense motor rotor angle and phase current, the steering execution angle, the execution motor rotor angle and phase current, the rack displacement, the steering gear housing vibration and noise signal, etc.

[0109] S2, analyze the correlation between the state monitoring quantity and the remaining life, calculate the correlation between each state monitoring quantity and the degradation characteristic quantity, and select the state monitoring quantity that can reflect the aging trend of the wire control steering system, specifically:

[0110] S21, calculate the correlation between each state monitoring quantity and the remaining life, and the calculation formula of the correlation coefficient R is:

[0111]

[0112] Among them, x i is the i-th state monitoring sequence, is the state monitoring sequence x i The mean of y and x i The corresponding degenerate characteristic quantity is is the mean of the time series y;

[0113] S22, screening state monitoring quantities with strong correlation: when the correlation coefficient is closer to 1, the correlation between the state monitoring quantity and the degradation characteristic quantity is stronger, based on which the critical range of the correlation coefficient r is determined to be 0.9 to 1, and the state monitoring quantity within this range is selected.

[0114] S3, using singular value decomposition to reduce the noise of the filtered state monitoring quantity, specifically:

[0115] S31, based on the phase space reconstruction theory, the state monitoring quantity is constructed into an m×n Hankel matrix H, and the matrix is ​​subjected to singular value decomposition to obtain the mutually unrelated noise subspace and signal subspace, the formula is:

[0116]

[0117] Where U represents an M×N orthogonal matrix; D represents a diagonal element; D=diag(σ1,σ2,σ3......σ n ) represents an M×N diagonal matrix; V represents an N×N orthogonal matrix;

[0118] S32, find the singular values ​​of the Hankel matrix A, retain the singular values ​​of the diagonal matrix that reflect the effective signal, set the smaller singular values ​​that reflect the noise signal to zero, preliminarily remove the noise in the high-frequency signal, and then perform the inverse process of singular value decomposition to finally obtain the matrix Then the matrix It is the best approximation moment of H with rank r, and the calculation formula is:

[0119]

[0120] Where σi represents the singular value of the matrix A; r represents the rank of the matrix H, T represents the transpose; ui represents the i-th eigenvector of the matrix AAT; vi represents the i-th eigenvector of the matrix AT A;

[0121] The matrix The anti-diagonal elements in are added and averaged to obtain the denoised state monitoring quantity.

[0122] S4, using CEEMDAN to decompose the degradation feature quantity into intrinsic mode components and residual components, where the intrinsic mode components reflect the global degradation trend, and the residual components reflect the local fluctuations caused by uncertainty. The degradation process of the wire-controlled steering system is essentially a nonlinear non-smooth sequence that follows the global degradation trend and is accompanied by local fluctuations. The local fluctuations are caused by uncertain working conditions. CEEMDAN decomposition adds an IMF component with auxiliary noise after EMD decomposition, instead of directly adding the Gaussian white noise signal to the original signal, which solves the problem that white noise always exists in the intrinsic mode components obtained by EMD. Therefore, CEEMDAN is used to decompose the degradation feature quantity into global degradation trends and local fluctuations, and predict them separately. Specifically:

[0123] S41, Gaussian white noise ε0E1(ω i (t)) is added to the signal to be decomposed x(t), and then empirical mode decomposition is performed to obtain the first eigenmode component IMF1 i (t), for IMF1 i(t) is averaged to obtain the first-order eigenmode component IMF1(t), whose formula is:

[0124]

[0125] Where N is the number of eigenmode components, ω i (t) is the white noise added for the ith time, ε i is the standard deviation of white noise;

[0126] S42, calculate the first-order residual component r1(t), the formula is:

[0127] r1(t)=x(t)-IMF1(t)

[0128] S43, continue to add Gaussian white noise to r1(t), and the new signal r1(t)+ε1E1(ω i (t)) to perform empirical mode decomposition and obtain the second-order eigenmode component:

[0129]

[0130] Among them, E1(r1(t)+ε1E1(w i )) represents an empirical mode decomposition;

[0131] S44, and so on, we get the kth order residual r k (t), the formula is:

[0132] r k (t) = r k-1 (t)-IMF k (t)

[0133] S45, calculate the k+1th order eigenmode component as:

[0134]

[0135] The signal is decomposed cyclically until the residual obtained is a monotonic function and cannot be further decomposed. The final residual component formula is:

[0136]

[0137] The obtained eigenmode components It reflects the global degradation trend, and the residual component r(t) reflects the local fluctuation caused by uncertainty.

[0138] S5, global degradation trend modeling is performed by inputting the eigenmode components into the MHA-BiLSTM model to predict the global degradation trend. For mechanical components, the degradation information contained in the monitoring signal changes during its service life and usually becomes more obvious as the degradation increases. The purpose of the attention module is to prompt the network to learn the hidden knowledge in the time series data that is closely related to the degradation process and to suppress redundant information in the feature learning process. Reflected in the internal calculation of the structure, the attention matrix is ​​constructed by assigning different weights to each incoming degradation data, thereby achieving matrix reconstruction, specifically:

[0139] S51, build a multi-head attention module, specifically:

[0140] S511, assuming that the tensor input to the module is input, after the linear transformation layer linear, map three tensors of the same size as the input: Q, K, V. These three tensors are put into the scaled dot product attention calculation part to obtain the attention weights and the reconstructed attention matrix, and the calculation formula is:

[0141]

[0142] S512, divide the tensors Q, K, and V into h equal-sized fragments in the same direction (usually the last dimension of the tensor), and calculate the attention matrix by placing each fragment of the tensor in a different subspace. After the calculation is completed, the matrices of these subspaces are stitched together in the division direction to obtain the multi-head attention matrix, which is calculated as follows:

[0143] Multihead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0144] Among them, head i is the attention matrix calculated in the i subspace, Concat is the concatenation operation of these sub-attention matrices in the split direction, and W O is the parameter matrix for linear transformation; the structure of the multi-head attention module is as follows Figure 2 shown.

[0145] S52, the multi-head attention module is embedded into the LSTM unit structure based on BiLSTM. Considering the good characteristics of the multi-head attention mechanism for sequence learning, the multi-head attention module is embedded into the LSTM unit structure based on BiLSTM. Considering the good characteristics of the multi-head attention mechanism for sequence learning, the multi-head attention module is embedded into the LSTM unit structure based on BiLSTM. The unit structure is as follows Figure 3 As shown, the calculation formula is as follows:

[0146] f(t)=MultiHead(σ(w f x(t)+w f h(t)+b f ))

[0147] i(t)=MultiHead(σ(w i x(t)+w i h(t-1)+b i ))

[0148] o(t)=σ(w o x(t)+w o h(t-1)+b o )

[0149]

[0150] h(t)=o(t)⊙tanh(c(t))

[0151] Among them, f(t), i(t), and o(t) are the outputs of the forget gate, input gate, and output gate respectively. Since the forget gate and input gate structure contain multi-head attention modules, the outputs f(t) and i(t) become tensors after attention reconstruction and participate in the calculation of the current state c(t) and output state h(t); x(t) is the input layer at time t, that is, the state monitoring quantity obtained after singular value decomposition in step S3; h(t-1) is the hidden layer at time t-1; h(t) is the hidden layer at time t; w f 、w i 、w o 、w c are the weight matrices of the forget gate, input gate, output gate, and hidden layer respectively; b f , b i , b o , b c are the forget gate, input gate, output gate, and bias term of the hidden layer respectively; σ and tanh are both activation functions, σ is the sigmoid function, and tanh is the hyperbolic tangent function; in the BiLSTM structure, the attention module exists in both directions of the information flow to adjust the weight of the degraded information;

[0152] S53, when training the prediction model, the root mean square error function is selected, and the formula is as follows:

[0153]

[0154] in, Represents the predicted value of the eigenmode component in the i-th data to be predicted, y irepresents the actual value of the intrinsic mode component in the i-th data to be predicted, n is the number of data, the network is initialized by randomly initializing weights and biases, and the gradient descent algorithm is used to update the parameters until the loss function converges to obtain the final prediction model, that is, the global degradation trend model;

[0155] S54, inputting the new state monitoring value into the global degradation trend model after preprocessing, and outputting the eigenmode component reflecting the global degradation trend.

[0156] S6, local fluctuation modeling is performed by inputting the residual component into the hidden Markov model to predict local fluctuations, specifically:

[0157] S61, the data sequence r={r1, r2, ..., r n} is represented as a set of observed states, and the data sequence of the state monitoring quantity obtained in step S3 is x={x1,x2,…,x n} is represented as a set of hidden states, where n represents the number of state values;

[0158] S62, the model parameters are estimated using the Baum-Welch algorithm, the formula is as follows:

[0159] λ=(A,B,P)

[0160] Among them, λ is the local fluctuation prediction model, A={a ij} N×N is the state transfer matrix, B is the probability of output observation, and P is the probability of the initial state;

[0161] The training parameters P and A are solved by the Baum-Welch algorithm to make the expected P(r t |λ) is the maximum; calculate the output observation probability B, the formula is as follows:

[0162]

[0163] Among them, j represents the state, K represents the number of Gaussian distributions, φ jl represents the weight of the lth Gaussian component, μ jl and Σ jl are the mean and covariance values ​​respectively;

[0164] S63, using the Viterbi algorithm to find the optimal hidden state sequence x corresponding to the observation sequence t , specifically:

[0165] S631, the calculation formula is as follows:

[0166] δ1(i)=p i bi (r1),1≤i≤N

[0167] ψ1(i)=0,1≤i≤N

[0168] Where N is the number of states, p i and b i is the i-th element of the model parameters P and B, δ1(i) and ψ1(i) are intermediate variables;

[0169] S632, recursively calculate the maximum probability of all single paths and their nodes, and the calculation formula is as follows:

[0170]

[0171] Where N is the number of states, δ t (i) is the total number of δ at the last moment. t-1 (i) with element a in A ij and element b in B i (r t ) product, where 1≤j≤N, ψ t (i) is to make δ t-1 (i)a ij The index of the maximum correspondence;

[0172] S633, terminate the iteration, the calculation formula is as follows:

[0173]

[0174] Where N is the number of states and P* is the number of all ψ t The maximum value of the elements in (i), x T * is to make ψ t (i) the index of the maximum correspondence;

[0175] S634, perform optimal path backtracking to obtain the final optimal hidden state sequence, the calculation formula is as follows:

[0176]

[0177] Among them, x t * is the optimal hidden state sequence;

[0178] S64, input the new state monitoring value into the local fluctuation prediction model after preprocessing, and output the local fluctuation.

[0179] S7, the prediction output of the global degradation trend model and the prediction output of the local fluctuation model are integrated to realize the prediction of the steering system degradation process, specifically:

[0180] S71, using weighted fusion method to process the prediction output of global degradation trend model and uncertainty model to achieve degradation process modeling;

[0181] S72, prediction time t k The remaining life L k , the calculation formula is as follows:

[0182] L k = inf{l k :X(l k +t k )>w}

[0183] Where X is the degradation data value at a certain moment, w is the failure threshold, l k Indicates the number of remaining life moments.

[0184] The degradation data is decomposed into global degradation trend and local fluctuation through Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). Accordingly, MHA-BiLSTM is used to predict the global degradation trend, and the hidden Markov model is used to predict the local fluctuation caused by uncertainty. Finally, the prediction output is fused to realize the life prediction of the steer-by-wire system.

[0185] It is understandable that the specific embodiments described above are only used to explain the relevant inventions, rather than to limit the inventions. It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. Multiple technical solutions in the same embodiment, as well as multiple technical solutions in different embodiments, can be arranged and combined to form new technical solutions without contradictions or conflicts. All equivalent structural transformations made using the contents of the specification and drawings of the present invention, directly or indirectly used in other related technical fields, are similarly included in the protection scope of the present invention.

Claims

1. A method for predicting the life of a steer-by-wire system, characterized in that: The steps include: S1, determining a degradation feature that affects the life of the steer-by-wire system, and collecting a state monitoring quantity and a degradation feature quantity related to the degradation feature; S2, analyzing the correlation between the state monitoring quantity and the remaining life, calculating the correlation between each state monitoring quantity and the degradation characteristic quantity, and selecting the state monitoring quantity that can reflect the aging trend of the wire control steering system; S3, using singular value decomposition to reduce the noise of the filtered state monitoring quantity; S4, using CEEMDAN to decompose the degradation feature into eigenmode components and residual components, where the eigenmode components reflect the global degradation trend and the residual components reflect the local fluctuations caused by uncertainty; S5, global degradation trend modeling is performed by inputting the eigenmode components into the MHA-BiLSTM model to predict the global degradation trend; S6, local fluctuation modeling by inputting the residual component into a hidden Markov model to predict local fluctuations; S7, the prediction output of the global degradation trend model and the prediction output of the local fluctuation model are integrated to achieve the prediction of the steering system degradation process.

2. The life prediction method of a steer-by-wire system according to claim 1, characterized in that: The step S1 specifically includes the following steps: The state monitoring quantity of the degradation characteristics of the steer-by-wire system is collected. Assuming that the steer-by-wire system has n state monitoring quantities and monitors t time moments in total, the state monitoring quantity x is expressed as: x={x1,x2,...,x n } Among them, x i =[x i1 ,x i2 ,...,x it ] T , i = 1, 2, ..., n, that is, the state monitoring quantity x can be expressed as: Determine the degradation characteristic quantity corresponding to the state monitoring quantity of the wire control steering system as the accuracy of the steering system, collect the degradation characteristic quantity, and express it as: Y = [y1, y2, ..., y t ] T Among them, yj represents the accuracy value of the wire-controlled steering system at time j, j=1, 2,..., t.

3. The method for predicting the life of a steer-by-wire system according to claim 2, characterized in that: The step S2 specifically includes the following steps: S21, calculate the correlation between each state monitoring quantity and the remaining life, and the calculation formula of the correlation coefficient R is: Among them, x i is the i-th state monitoring sequence, is the state monitoring sequence x i The mean of y and x i The corresponding degenerate characteristic quantity, is the mean of the time series y; S22, screening state monitoring quantities with strong correlation: when the correlation coefficient is closer to 1, the correlation between the state monitoring quantity and the degradation characteristic quantity is stronger, based on which the critical range of the correlation coefficient r is determined to be 0.9 to 1, and the state monitoring quantity within this range is selected.

4. The method for predicting the life of a steer-by-wire system according to claim 1, characterized in that: The step S3 specifically comprises the following steps: S31, based on the phase space reconstruction theory, the state monitoring quantity is constructed into an m×n Hankel matrix H, and the matrix is ​​subjected to singular value decomposition to obtain the mutually unrelated noise subspace and signal subspace, the formula is: Where U represents an M×N orthogonal matrix; D represents a diagonal element; D=diag(σ1,σ2,σ3......σ n ) represents an M×N diagonal matrix; V represents an N×N orthogonal matrix; S32, find the singular values ​​of the Hankel matrix A, retain the singular values ​​of the diagonal matrix that reflect the effective signal, set the smaller singular values ​​that reflect the noise signal to zero, preliminarily remove the noise in the high-frequency signal, and then perform the inverse process of singular value decomposition to finally obtain the matrix Then the matrix It is the best approximation moment of H with rank r, and the calculation formula is: Where σi represents the singular value of the matrix A; r represents the rank of the matrix H, T represents the transpose; ui represents the i-th eigenvector of the matrix AAT; vi represents the i-th eigenvector of the matrix AT A; The matrix The anti-diagonal elements in are added and averaged to obtain the denoised state monitoring quantity.

5. The method for predicting the life of a steer-by-wire system according to claim 1, characterized in that: The step S4 specifically comprises the following steps: S41, Gaussian white noise ε0E1(ω i (t)) is added to the signal to be decomposed x(t), and then empirical mode decomposition is performed to obtain the first eigenmode component right The average calculation is performed to obtain the first-order eigenmode component IMF1(t), whose formula is: Where N is the number of eigenmode components, ω i (t) is the white noise added for the ith time, ε i is the standard deviation of white noise; S42, calculate the first-order residual component r1(t), the formula is: r1(t)=x(t)-IMF1(t) S43, continue to add Gaussian white noise to r1(t), and the new signal r1(t)+ε1E1(ω i (t)) to perform empirical mode decomposition and obtain the second-order eigenmode component: Among them, E1(r1(t)+ε1E1(w i )) represents an empirical mode decomposition; S44, and so on, we get the kth order residual r k (t), the formula is: r k (t)=r k-1 (t)-IMF k (t) S45, calculate the k+1th order eigenmode component as: The signal is decomposed cyclically until the residual obtained is a monotonic function and cannot be further decomposed. The final residual component formula is: The obtained eigenmode components It reflects the global degradation trend, and the residual component r(t) reflects the local fluctuation caused by uncertainty.

6. The method for predicting the life of a steer-by-wire system according to claim 1, characterized in that: The step S5 specifically comprises the following steps: S51, build a multi-head attention module, specifically: S511, assuming that the tensor input to the module is input, after the linear transformation layer linear, map three tensors of the same size as the input: Q, K, V. These three tensors are put into the scaled dot product attention calculation part to obtain the attention weights and the reconstructed attention matrix, and the calculation formula is: S512, divide the tensors Q, K, and V into h equal-sized fragments in the same direction (usually the last dimension of the tensor), and calculate the attention matrix by placing each fragment of the tensor in a different subspace. After the calculation is completed, the matrices of these subspaces are stitched together in the division direction to obtain the multi-head attention matrix, which is calculated as follows: Multihead(Q,K,V)=Concat(head1,head2,…,head h )W O Among them, head i is the attention matrix calculated in the i subspace, Concat is the concatenation operation of these sub-attention matrices in the split direction, and W O is the parameter matrix for linear transformation; S52, embed the multi-head attention module into the LSTM unit structure based on BiLSTM, and its calculation formula is as follows: f(t)=MultiHead(σ(w f x(t)+w f h(t)+b f )) i(t)=MultiHead(σ(w i x(t)+w i h(t-1)+b i )) o(t)=σ(w o x(t)+w o h(t-1)+b o ) h(t)=o(t)⊙tanh(c(t)) Among them, f(t), i(t), and o(t) are the outputs of the forget gate, input gate, and output gate respectively. Since the forget gate and input gate structure contain multi-head attention modules, the outputs f(t) and i(t) become tensors after attention reconstruction and participate in the calculation of the current state c(t) and output state h(t); x(t) is the input layer at time t, that is, the state monitoring quantity obtained after singular value decomposition in step S3; h(t-1) is the hidden layer at time t-1; h(t) is the hidden layer at time t; w f 、w i 、w o 、w c are the weight matrices of the forget gate, input gate, output gate, and hidden layer respectively; b f 、b i 、b o 、b c are the forget gate, input gate, output gate, and bias term of the hidden layer respectively; σ and tanh are both activation functions, σ is the sigmoid function, and tanh is the hyperbolic tangent function; in the BiLSTM structure, the attention module exists in both directions of the information flow to adjust the weight of the degraded information; S53, when training the prediction model, the root mean square error function is selected, and the formula is as follows: in, Represents the predicted value of the eigenmode component in the i-th data to be predicted, y i represents the actual value of the intrinsic mode component in the i-th data to be predicted, n is the number of data, the network is initialized by randomly initializing weights and biases, and the gradient descent algorithm is used to update the parameters until the loss function converges to obtain the final prediction model, that is, the global degradation trend model; S54, inputting the new state monitoring value into the global degradation trend model after preprocessing, and outputting the eigenmode component reflecting the global degradation trend.

7. The method for predicting the life of a steer-by-wire system according to claim 3, characterized in that: The step S6 specifically comprises the following steps: S61, the data sequence r={r1, r2, ..., r n } is represented as a set of observed states, and the data sequence of the state monitoring quantity obtained in step S3 is x={x1,x2,…,x n } is represented as a set of hidden states, where n represents the number of state values; S62, the model parameters are estimated using the Baum-Welch algorithm, the formula is as follows: λ=(A,B,P) Among them, λ is the local fluctuation prediction model, A={a ij } N×N is the state transfer matrix, B is the probability of output observation, and P is the probability of the initial state; The training parameters P and A are solved by the Baum-Welch algorithm to make the expected P(r t |λ) is the maximum; calculate the output observation probability B, the formula is as follows: Among them, j represents the state, K represents the number of Gaussian distributions, φ jl represents the weight of the lth Gaussian component, μ jl and Σ jl are the mean and covariance values ​​respectively; S63, using the Viterbi algorithm to find the optimal hidden state sequence x corresponding to the observation sequence t , specifically: S631, the calculation formula is as follows: δ1(i)=p i b i (r1),1≤i≤N ψ1(i)=0,1≤i≤N Where N is the number of states, p i and b i is the i-th element of the model parameters P and B, δ1(i) and ψ1(i) are intermediate variables; S632, recursively calculate the maximum probability of all single paths and their nodes, and the calculation formula is as follows: Where N is the number of states, δ t (i) is the total number of δ at the last moment. t-1 (i) with element a in A ij and element b in B i (r t ) product, where 1≤j≤N, ψ t (i) is to make δ t-1 (i)a ij The index of the maximum correspondence; S633, terminate the iteration, the calculation formula is as follows: Where N is the number of states and P* is the number of all ψ t The maximum value of the elements in (i), x T * is to make ψ t (i) the index of the maximum correspondence; S634, perform optimal path backtracking to obtain the final optimal hidden state sequence, the calculation formula is as follows: Among them, x t * is the optimal hidden state sequence; S64, input the new state monitoring value into the local fluctuation prediction model after preprocessing, and output the local fluctuation.

8. The method for predicting the life of a steer-by-wire system according to claim 1, characterized in that: The step S7 specifically includes the following steps: S71, using weighted fusion method to process the prediction output of global degradation trend model and uncertainty model to achieve degradation process modeling; S72, prediction time t k The remaining life L k , the calculation formula is as follows: L k =inf{l k :X(l k +t k )>w} Where X is the degradation data value at a certain moment, w is the failure threshold, l k Indicates the number of remaining life moments.

Images