Train key component online health monitoring method based on multi-sensor signal cooperation

Abstract

The invention discloses a train key component online health monitoring method based on multi-sensor signal cooperation, and the method comprises the steps: selecting an indirect signal as an edge end train online abnormal state early warning index based on the correlation analysis of multi-sensor signals, and diagnosing whether the state of a train key component is abnormal or not; the vibration signals serve as signals for further diagnosis at the cloud, state features of the vibration signals are extracted at the edge end, online optimization of sparse coding is carried out, and the vibration signals are transmitted to the cloud to serve as diagnosis signal features for further diagnosis; real-time training precision and sample complexity are integrated into online estimation of a model training process, optimal features are transferred from a source domain-edge end to a target domain-cloud end, and the cloud end diagnoses the type and grade of a fault by means of the complex operation capability of the cloud end; a dynamic aggregation strategy is provided, the features of a prediction model are determined, optimization of nodes is enhanced, and a feature network is made to adapt to and complete an online prediction task.

Description

Technical Field

[0001] The present invention relates to the technical field of train detection devices, and in particular to an online health monitoring method for key train components based on multi-sensor signal coordination. Background Art

[0002] Rolling bearings are common vulnerable parts in mechanical equipment, and their health is crucial to the normal operation and life of the equipment. Therefore, the health monitoring and fault diagnosis of rolling bearings have always been a hot topic of research.

[0003] Existing rolling bearing health monitoring methods mainly rely on direct signals, such as vibration signals, which are easily interfered by strong noise, affecting the accuracy of fault feature extraction and diagnosis; the amount of vibration signal data is large, requiring complex feature extraction and signal processing, and the computational cost is high; the characteristics of vibration signals are different under different working conditions, and different models need to be established for different working conditions; it is difficult to handle nonlinear and non-stationary signals. In addition, traditional monitoring methods usually require a lot of computing resources and are not suitable for real-time online monitoring. Especially in industrial environments, bearing failures can lead to serious consequences, including equipment damage and production interruptions. Therefore, a method that can accurately and timely monitor the health of bearings is needed to achieve predictive maintenance and reduce unexpected downtime.

[0004] At present, the online monitoring framework of rolling bearings can be divided into two categories:

[0005] (1) Direct signal-based monitoring framework: This type of framework mainly uses vibration signals for fault diagnosis and prediction, and vibration signals contain rich fault information.

[0006] (2) Fault diagnosis and prediction based on indirect signals: perform composite processing of indirect signals such as force, torque, and current, build a multi-signal tensor model, and perform fault diagnosis and state prediction based on signal tensors. There is a correlation between indirect signals and vibration signals, which can be used as auxiliary prediction signals.

[0007] Specifically include:

[0008] (1) Analysis technology based on vibration signals: Analyze the time domain, frequency domain and time-frequency domain characteristics of the vibration signal to identify the fault type and fault degree of the rolling bearing.

[0009] (2) Indirect signal analysis technology: Use indirect signals such as force, torque, and current to perform fault diagnosis and overcome problems such as vibration signal noise interference and high computational complexity.

[0010] (3) Acoustic emission technology: By monitoring the acoustic emission signals generated by the bearing during operation and analyzing the characteristics of the signals, early faults of the bearing can be identified.

[0011] (4) Temperature monitoring technology: By monitoring the temperature changes of the bearing, the health status of the bearing can be diagnosed based on the principle that the temperature rise may be related to the severity of the fault.

[0012] (5) Oil analysis technology: Evaluate the wear condition and potential failure of bearings by analyzing the wear particles, contaminants and chemical composition in the bearing lubricant.

[0013] (6) Model-based monitoring technology: Utilize the dynamic model of the bearing to diagnose faults by comparing the differences between actual operating parameters and model predicted parameters.

[0014] (7) Data-driven technology: Use algorithms such as support vector machine (SVM), random forest, decision tree, K nearest neighbor (KNN) to classify features to identify faults. Use deep learning models such as convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM) to automatically extract features and perform fault diagnosis.

[0015] (8) Multi-sensor data fusion technology: Combining data from multiple sensors (such as vibration, sound, temperature, etc.) to improve the accuracy and reliability of fault diagnosis through data fusion technology.

[0016] The technical solution framework of the prior art mainly covers the following steps:

[0017] Step 1: Multi-sensor signal acquisition.

[0018] Use vibration sensors to collect vibration signals of rolling bearings, such as acceleration signals, velocity signals, etc.; use force sensors, torque sensors, current sensors, etc. to collect indirect signals of rolling bearings; fuse vibration signals and indirect signals, for example, construct a manifold learning model for multi-signal fusion.

[0019] Step 2: Feature extraction.

[0020] Extract time domain features of vibration signals, such as mean, variance, kurtosis, skewness, etc.; extract frequency domain features of vibration signals, such as spectrum, power spectrum, cepstrum, etc.; extract time and frequency domain features of vibration signals, such as wavelet transform, Hilbert-Huang transform, etc.; use adaptive sparse representation technology to extract features of vibration signals and indirect signals, such as K-SVD algorithm; use MKL for multi-core learning to extract features of vibration signals and indirect signals.

[0021] Step 3: Model training.

[0022] Use machine learning algorithms such as support vector machines, random forests, and decision trees to train fault diagnosis models, and use deep learning algorithms such as CNN, RNN, and LSTM to train fault diagnosis models; use federated learning algorithms to perform distributed model training to protect data privacy; dynamically adjust model parameters based on real-time data and prediction results to improve the model's adaptability and prediction accuracy.

[0023] Step 4: Troubleshooting.

[0024] Use the trained model to diagnose rolling bearing faults and identify the fault type and fault severity; generate early warning information based on the diagnosis results, such as fault level, fault location, fault development trend, etc.; perform health management based on the diagnosis results, such as formulating maintenance plans and predicting remaining service life.

[0025] Disadvantages of existing technology:

[0026] (1) Temperature and acoustics-based fault detection techniques, such as time series analysis or artificial intelligence-based methods, are effective means of predicting component anomalies in advance. For univariate time series prediction, Gaussian process regression and autoregressive integrated moving average are effective. However, these methods have strict constraints on adaptive assumptions and are difficult to apply in practice. Neural networks are the most popular due to their nonlinear capabilities and scalability. However, as the network deepens, its computational complexity also increases. Different rolling bearings may involve different operating conditions, and their temperature and acoustic variation patterns are also different.

[0027] (2) In vibration-based technologies, the fault characteristics of the machine easily bury direct and indirect signals. Direct signals are easily interfered by noise, which affects the monitoring accuracy. They consume a lot of computing resources and are not suitable for real-time online monitoring. This makes it difficult for most data-driven signal processing technologies to process nonlinear and non-stationary vibration signals used for online health detection.

[0028] In view of the above-mentioned deficiencies in the prior art, the present invention is proposed. Summary of the invention

[0029] The technical problems to be solved by the present invention are:

[0030] Since time-varying vibration signals (direct signals for short) are often interfered with and overwhelmed by strong noise, most data-driven signal processing techniques (such as deep learning) find it difficult to instantly process nonlinear and non-stationary vibration signals used for online health monitoring, while the fault characteristics of the machine are contained in both direct and indirect signals.

[0031] In view of the shortcomings of the prior art, the present invention models the highly complex dependency structure between direct signals and indirect signals, and searches for indirect signals with traceability and strong correlation with direct signals for state identification and fault diagnosis. State identification and early warning are performed at the edge based on indirect signals with low computational complexity, while more accurate fault diagnosis and state diagnosis are performed based on time-consuming vibration signals in the cloud. An aggregate learning framework for online state prediction and fault diagnosis is proposed by combining the edge and the cloud, and the optimization target of the model is determined, so that online updates have sufficient adaptability to online diagnosis tasks.

[0032] In order to achieve the above object, the present invention adopts the following technical solutions:

[0033] An online health monitoring method for key train components based on multi-sensor signal coordination, specifically comprising:

[0034] Based on the correlation analysis of multi-sensor signals, indirect signals are selected as early warning indicators of abnormal train status online in the offline stage;

[0035] The indirect signal is a signal that is processed at the edge and is strongly correlated with the vibration signal, and an online warning of abnormal train status is performed based on the indirect signal;

[0036] The edge end transmits the vibration signal features to the cloud end for use as diagnostic signal features. The vibration signal is used as a signal for further diagnosis in the cloud end. Its state features are extracted at the edge end, and its sparse coding is optimized online, and transmitted to the cloud end for use as diagnostic signal features.

[0037] Integrate real-time training accuracy and sample complexity into the online estimation of the model training process, and transfer the optimal features from the source domain (edge) to the target domain (cloud). The cloud can diagnose the type and level of faults with its complex computing capabilities.

[0038] A dynamic aggregation strategy is adopted to determine the characteristics of the prediction model and enhance the optimization of nodes, so that the feature network can adapt to complete the online prediction task, providing a new advantageous tool for the current online monitoring based on the coordination of indirect signals and diagnostic signal diagnosis.

[0039] The preferred technical solution is:

[0040] The correlation analysis based on multi-sensor signals selects indirect signals as early warning indicators of abnormal train status online in the offline stage, specifically including:

[0041] Phase space reconstruction: Time-delay reconstruction of vibration signals and lightweight indirect signals is performed to convert them into state space. State space reconstruction is used to reveal the nonlinear dynamic characteristics of the signal and mine the hidden information in the time series signal.

[0042] Correlation analysis of signal time series: The correlation between the vibration signal and the lightweight indirect signal after phase space reconstruction is analyzed using the Copula Entropy (CE) function. The Copula entropy function can effectively describe the correlation relationship and quantify the degree of dependence between the vibration signal and the lightweight indirect signal.

[0043] According to the CE value between the vibration signal and the indirect signal, the correlation between the vibration signal and the indirect signal is estimated;

[0044] Through dimensionality reduction and denoising, lightweight signals with high correlation with vibration signals are selected as auxiliary prediction signals to improve fault diagnosis performance.

[0045] A further preferred technical solution is:

[0046] The phase space reconstruction is to reconstruct the vibration signal and the lightweight signal with time delay and convert them into the state space, which specifically includes:

[0047] According to Takens’ embedding theorem, a time series X = [X1, X2, ..., X L The i-th element X in ] i , is reconstructed in state space as:

[0048] X i =(x(i),x(i+τ),…,x(i+(m-1)τ)) (1)

[0049] Where X is a time series, L means that there are L elements in the time series X, i = 1, 2, λ..., L, i is a natural number, Xi is the i-th point in the state space, and x(i) is the number of elements that constitute the state vector X i A component of , m is the embedding dimension, and τ is the time delay;

[0050] The embedding dimension is estimated by pseudo nearest neighbor (FNN), which examines the distance variation between neighboring points in the state space;

[0051] When the embedding dimension changes, the distance between FNNs does not change, indicating that the current embedding dimension is sufficient;

[0052] The square of the Euclidean distance between the rth nearest neighbors of a coordinate vector is given by:

[0053]

[0054] in, is the square of the Euclidean distance between the rth nearest neighbors of the coordinate vector, t is the time; x(t+kτ) is the value of the original time series at time t+kτ, x (r)(t+kτ) is the value of the rth nearest neighbor corresponding to that moment;

[0055] The criterion for determining the embedding parameter m is called false neighbors:

[0056]

[0057] Among them, R tol is a given threshold;

[0058] Then, based on a geometric framework, the average displacement is applied to determine the time delay τ;

[0059] Based on the reconstructed x(t), calculate the phase point and its neighboring points The average distance S m (τ):

[0060]

[0061] Where N is the total number of data points;

[0062] When the embedding parameter m is given:

[0063]

[0064] x i is the i-th data point in the time series. As τ increases, S m (τ) linear growth saturates, and the value corresponding to the end of its linear range is the optimal delay. m When the slope of (τ) decreases to less than 40% of the initial value, it is the optimal value of the delay time.

[0065] The preferred technical solution is:

[0066] The correlation analysis of the signal time series uses the Copula function to analyze the correlation between the vibration signal and the lightweight signal, specifically:

[0067] The signal time series is a statistically independent measure defined by copula;

[0068] X={X1,X2,...,X n} is a random variable with copula density c(u) and marginal distribution u;

[0069] The copula entropy (CE) of X is defined as:

[0070] H c (X) = -I(X) = -∫ u c(u)ln c(u)du (6)

[0071] The joint entropy H(X) of multiple random variables is equal to the sum of marginal entropy and CE:

[0072] H(X)=∑ i H(X i )+H c (X) (7)

[0073] According to the CE value between the vibration signal and the indirect signal, the optical signal with a higher correlation with the vibration signal is selected as the auxiliary prediction signal.

[0074] The preferred technical solution is:

[0075] The processing of the indirect signal strongly correlated with the vibration signal at the edge and performing fault detection and early warning includes:

[0076] Collect indirect signals and construct training sets and valid sets;

[0077] Adaptive sparse representation of indirect signals and fault diagnosis based on multi-core extreme learning machines;

[0078] Steps for troubleshooting at the edge:

[0079] First, the time series of the signal is considered through phase space reconstruction;

[0080] During the online monitoring process, the signals from the sensors are soft-thresholded to select the signal data stream for training;

[0081] At each time t, the sparse representation of the signal is realized through adaptive sparse coding, and the signal is trained and predicted;

[0082] For time t in the training phase, all information before time t is available, and the epoch is estimated dynamically. Input the model together with the target value to train the model and update the model weights;

[0083] In the prediction stage, Input the multi-core ELM model to generate prediction values If the predicted value If the soft threshold is exceeded, is the predicted value of time step t, which is recorded for early warning, and the vibration signal is further processed and predicted in the cloud; other prior values ​​are used as training samples for future training.

[0084] The preferred technical solution is:

[0085] The fault diagnosis based on multi-core extreme learning machine includes: adaptive sparse representation, kernel function selection, and weight optimization, wherein:

[0086] Adaptive sparse representation: Due to the sparsity of the signal, the diagnostic performance will rely on the sparse representation of the indirect signal, on which the indirect signal features are used for dictionary learning; the generalized minimax concave (GMC) is selected as the penalty function, and the K-singular value decomposition (K-SVD) is used to improve the matrix composition on the sample signal; the soft threshold algorithm is used to denoise the signal and enhance the fault features to reduce the calculation time;

[0087] Based on a single hidden layer neural network, an adaptive multi-kernel extreme learning machine is used for fault diagnosis:

[0088]

[0089] Among them, β j is the output weight, ω j is the input weight, s j is the bias, g( ) is the activation function, x i is the input vector;

[0090] In order to reduce the uncertainty of a single kernel, a multi-kernel ELM (MKL) is adopted, and the optimal kernel is assumed to be a base kernel of a linear combination of a set of kernels, and the optimal combination coefficients and structural parameters of the classifier are jointly learned by maximizing the margin;

[0091]

[0092] Among them, Ω(x, x k )=C1K poly (X, x k )+C2K RBF (X, x k )+C3K wave (X, x k ), C i , i=1, 2, 3 are the weights corresponding to the basis functions, ∑C i =1. The base kernel functions are:

[0093]

[0094] Among them, K poly (X, x k ) is the polynomial kernel function, K RBF (X, x k ) is the radial basis function kernel, K wave (X, x k ) is the wavelet kernel function, C0 is a constant used to control the smoothness of the kernel function; σ is a parameter used to control the width of the kernel function; a is the scale parameter of the wavelet transform; S is the regularization term coefficient; I is the unit matrix;

[0095] Online weight optimization: In order to improve the performance of MKL, a model combining particle swarm optimization (PSO) and genetic algorithm (GA) is used to optimize the above parameters C0, a, C i , σ, S are optimized, and the individual optimal value (pbest) and the global optimal value (gbest) are searched according to the best fitness function;

[0096] The fitness function is defined by the root mean square error (RMSE) of the signal verification set; the parameter group is optimized by the particle swarm algorithm, and the speed and position of the particles are updated using the formula; the individual optimal value and the global optimal value are updated by genetic operation; the optimal parameter set is obtained after the predetermined number of iterations is reached.

[0097] The preferred technical solution is:

[0098] The edge end transmits the vibration signal characteristics to the cloud for use as diagnostic signal characteristics, including:

[0099] On the edge computing platform, features and enhanced nodes are constructed based on local signals and combined into nodes through pseudo-inverse calculation;

[0100] Perform online optimization of adaptive sparse coding on feature signals, extract their state features, send enough feature nodes to the server domain, and evaluate the corresponding weights of the nodes;

[0101] Then the weights are optimized until the model converges, and an aggregation model for predicting the state of rolling bearings is established.

[0102] The preferred technical solution is:

[0103] The edge transmits the vibration signal features to the cloud. It is necessary to integrate the penalty into the width learning system and add the manifold regularization term, including:

[0104] The dynamic distribution adaptive penalty adopts the maximum mean difference (MMD) method, which is specifically:

[0105] MMD is a penalty used to measure and minimize the distance between source and target domain data:

[0106]

[0107] where F is a class of functions defined as the set of unit balls in the reproducing kernel Hilbert space; for two independent distributions p and q, if the maximum mean difference (MMD) between them is 0, then we can detect that there is no difference between them.

[0108] To integrate the penalty into the generalized learning system (BLS), the MMD is calculated and projected considering the given projection vector:

[0109]

[0110] D S is the dataset of the source domain, D T is the dataset of the target domain;

[0111] The original data is converted into mapping features through weight transformation; for the input data X, the i-th group of mapping features is calculated using the linear transformation function,

[0112]

[0113] Among them, W ei and β ei is a random weight with appropriate dimension; n is the number of feature node groups; is a linear transformation function;

[0114] The mapped features are then transformed into enhanced nodes through nonlinear transformation:

[0115]

[0116] in, and are random weights with appropriate dimension; ξ j ( ) is a nonlinear transformation function;

[0117] Where A=[Z n |H m ]=[a(x1) T , ..., a(x s+t ) T ] T , A∈R N×(nk+m) , N is the number of samples, nk+m is the total feature nodes;

[0118] Then minimize the squared loss prediction error:

[0119]

[0120] Among them, Y is the labeled sample corresponding to the input data X; β is the weight coefficient; θ is the regularization parameter to compensate for model error and complexity;

[0121] β=(A T A+θI)-1A T Y(18)

[0122] In particular, we can get β * =lim θ→0 (θI+A T A) -1 A T; When the number of data and model structure are fixed, β * And send it to the cloud for model aggregation;

[0123] The Manifold Regularization Framework (MRF) algorithm integrates a large number of unlabeled samples and a small number of labeled samples to explore the manifold features hidden in the data and express them as a loss function;

[0124] The manifold regularized MRF is specifically:

[0125] The MRF algorithm can explore the various characteristics of the data and express them as a loss function:

[0126]

[0127] in, is the predicted value; Tr(·) is the trajectory of the Laplace operator matrix L, which is obtained by formula (8):

[0128] L=DW (20)

[0129] Among them, the degree matrix D={D ii}, W={w (i，j)} is the adjacency matrix of the similarity between the source domain and the target domain data, which is calculated by the following formula:

[0130]

[0131] Among them, N k (x) is the k-neighboring value of sample x, i = 1…l+u, j = 1…l+u, l is the number of labeled samples, u is the number of unlabeled samples; σ is a hyperparameter used to control the scale in similarity calculation; x i is the data vector of the i-th sample;

[0132] According to formulas (20) and (21), (5) can be written as:

[0133]

[0134] The preferred technical solution is:

[0135] Integrate real-time training accuracy and sample complexity into the online estimation of the model training process, and transfer the optimal features from the source domain (edge) to the target domain (cloud). The cloud can diagnose the type and level of faults with its complex computing capabilities, including:

[0136] On the edge computing platform, features and enhanced nodes are constructed based on local signals, and combined into nodes through pseudo-inverse calculation, and vibration signal features are extracted through adaptive sparse representation;

[0137] Based on manifold regularized MRF and dynamic distribution adaptive penalty MMD, the adaptive transfer learning algorithm is used for the diagnosis of cloud computing platform;

[0138] Send enough feature nodes to the server domain and evaluate the corresponding weights of the nodes;

[0139] Then the weights are optimized until the model converges, and an aggregate model for predicting the state of rolling bearings is established;

[0140] In the prediction phase, the sample X(t-lu) arrives at time t-lu and is input into the model to generate At subsequent times, X(t-lu) is repeatedly input into the model to generate These predicted values ​​should be equal to the true values. The error between the predicted values ​​and the true values ​​can be used to calculate the prediction error vector as feedback for estimating the training parameters.

[0141] The preferred technical solution is:

[0142] The dynamic aggregation strategy determines the characteristics of the prediction model and enhances the optimization of nodes to adapt the feature network to complete the online prediction task, including:

[0143] The cloud computing server receives enough parameters β, takes the time delay of the vibration signal as the main model, calculates the maximum weight, and divides the remaining weights equally among other models to form a new fault diagnosis model.

[0144] The generalization performance of the generated model is verified based on the validation set, and the weights of the model received from the source domain are adjusted until the sovereignty value is reduced to convergence.

[0145] Beneficial effects:

[0146] (1) An online health monitoring framework that adapts to the coordination of multi-sensor signal diagnosis can monitor the health status of rolling bearings in a timely and accurate manner. Through edge-cloud computing coordination, indirect signals with low computing power requirements can be used to provide real-time online warning of train operation status. At the same time, diagnostic signals with high computing complexity can be used in the cloud for more accurate fault diagnosis. The combination of the two fully utilizes the computing power of the edge-cloud and improves monitoring efficiency.

[0147] (2) An analysis model based on the copula entropy function was established to prove the strong correlation between the indirect signal and the vibration signal. The indirect signal was determined as an auxiliary prediction signal for the vibration signal and as an early warning indicator for abnormal train status online.

[0148] (3) Through edge computing, indirect signals are used for online fault diagnosis or prediction, reducing computing costs and improving computing efficiency.

[0149] (4) During the online monitoring process, the characteristic parameters are dynamically estimated based on the multi-kernel function to extract the nonlinear and non-stationary components in the multi-sensor signals, which improves the diagnostic efficiency and performance of the ELM and helps to detect early failures of rolling bearings in a timely manner.

[0150] (5) The edge transmits the vibration signal to the cloud as a diagnostic signal feature for further fault diagnosis. Based on the real-time multi-sensor signal and period prediction results, the feature matrix and optimized weights are aggregated on the cloud platform to improve the diagnostic performance and efficiency of the rolling bearing, making it sufficiently adaptive to the online prediction task.

[0151] (6) Based on the adaptive representation of the selected signal, the training parameters are dynamically estimated, and personalized learning is achieved by adopting a combination of sample complexity and real-time prediction error, thereby achieving personalized training. The cloud uses its complex computing capabilities to diagnose the type and level of faults and improve the prediction accuracy of the model.

[0152] (7) An online aggregation learning framework is proposed on the cloud computing platform to determine the optimization objectives of the model, improve the diagnostic performance and efficiency of rolling bearings, and improve its adaptability to online diagnostic tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0153] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0154] Figure 1 It is a schematic diagram of fault diagnosis of multi-source signal processing through edge cloud collaboration according to the method of the present invention. DETAILED DESCRIPTION

[0155] Attached Figure 1 Schematic diagram of fault diagnosis for multi-source signal processing through edge cloud collaboration. Figure 1 As shown, the present invention is implemented by the following technical solutions:

[0156] An online health monitoring method for key train components based on multi-sensor signal coordination, specifically comprising:

[0157] Based on the correlation analysis of multi-sensor signals, indirect signals are selected as early warning indicators of abnormal train status online in the offline stage;

[0158] Specifically include:

[0159] Phase space reconstruction: Time-delay reconstruction of vibration signals and lightweight indirect signals is performed to convert them into state space. State space reconstruction is used to reveal the nonlinear dynamic characteristics of the signal and mine the hidden information in the time series signal.

[0160] The phase space reconstruction is to reconstruct the vibration signal and the lightweight signal with time delay and convert them into the state space, which specifically includes:

[0161] According to Takens’ embedding theorem, a time series X = [X1, X2, ..., X L ], can be reconstructed into a state space;

[0162] X i =(x(i),x(i+τ),…,x(i+(m-1)τ)) (1)

[0163] Where X is a time series, i is a natural number, and X i is the i-th point in the state space, x(i) is the state vector X i A component of , m is the embedding dimension, and τ is the time delay;

[0164] The embedding dimension is estimated by false nearest neighbor (FNN), which examines the distance variation between neighboring points in the state space;

[0165] When the embedding dimension changes, the distance between FNNs does not change, indicating that the current embedding dimension is sufficient;

[0166] The square of the Euclidean distance between the rth nearest neighbors of a coordinate vector is given by:

[0167]

[0168] in, is the square of the Euclidean distance between the rth nearest neighbors of the coordinate vector, and t is the time;

[0169] The criterion for determining the embedding parameter m is called false neighbors:

[0170]

[0171] Among them, R tol is a given threshold;

[0172] Then, based on a geometric framework, the average displacement is applied to determine the time delay τ;

[0173] Based on the reconstructed x(t), calculate the phase point and its neighboring points The average distance S m (τ):

[0174]

[0175] Where N is the total number of data points;

[0176] When the embedding parameter m is given:

[0177]

[0178] As τ increases, S m (τ) linear growth saturates, and the value corresponding to the end of its linear range is the optimal delay. m When the slope of (τ) decreases to less than 40% of the initial value, it is the optimal value of the delay time.

[0179] Correlation analysis of signal time series: The correlation between the vibration signal and the lightweight indirect signal after phase space reconstruction is analyzed using the Copula Entropy (CE) function. The Copula entropy function can effectively describe the correlation relationship and quantify the degree of dependence between the vibration signal and the lightweight indirect signal.

[0180] The correlation analysis of the signal time series uses the Copula function to analyze the correlation between the vibration signal and the lightweight signal, specifically:

[0181] The signal time series is a statistically independent measure defined by copula;

[0182] X={X1,X2,...,X n} is a random variable with copula density c(u) and marginal distribution u;

[0183] The copula entropy (CE) of X is defined as:

[0184] H c (X) = -I(X) = -∫ u c(u)ln c(u)du (6)

[0185] The joint entropy H(X) of multiple random variables is equal to the sum of marginal entropy and CE:

[0186] H(X)=∑ i H(X i )+H c (X) (7)

[0187] According to the CE value between the vibration signal and the indirect signal, the optical signal with a higher correlation with the vibration signal is selected as the auxiliary prediction signal.

[0188] According to the CE value between the vibration signal and the indirect signal, the correlation between the vibration signal and the indirect signal is estimated;

[0189] Through dimensionality reduction and denoising, lightweight signals with high correlation with vibration signals are selected as auxiliary prediction signals to improve fault diagnosis performance.

[0190] Wherein, the indirect signal is a signal strongly correlated with the vibration signal processed at the edge, and an online warning of abnormal train status is performed based on the indirect signal;

[0191] The processing of the indirect signal strongly correlated with the vibration signal at the edge and performing fault detection and early warning includes:

[0192] Collect indirect signals and construct training sets and valid sets;

[0193] Adaptive sparse representation of indirect signals and fault diagnosis based on multi-core extreme learning machines;

[0194] Steps for troubleshooting at the edge:

[0195] First, the time series of the signal is considered through phase space reconstruction;

[0196] During the online monitoring process, the signals from the sensors are soft-thresholded to select the signal data stream for training;

[0197] At each time t, the sparse representation of the signal is realized through adaptive sparse coding, and the signal is trained and predicted;

[0198] For time t in the training phase, all information before time t is available, and the epoch is estimated dynamically. Sample X t-lu ={X (t-lu-lv) , …, X (t-lu)}, Input the model together with the target value to train the model and update the model weights;

[0199] In the prediction stage, Input the multi-core ELM model to generate prediction values If the predicted value If the soft threshold is exceeded, is the predicted value of time step t, which is recorded for early warning, and the vibration signal is further processed and predicted in the cloud; other prior values ​​are used as training samples for future training.

[0200] The fault diagnosis based on multi-core extreme learning machine includes: adaptive sparse representation, kernel function selection, and weight optimization, wherein:

[0201] Adaptive sparse representation: Due to the sparsity of the signal, the diagnostic performance will rely on the sparse representation of the indirect signal, on which the indirect signal features are used for dictionary learning; the generalized minimax concave (GMC) is selected as the penalty function, and the K-singular value decomposition (K-SVD) is used to improve the matrix composition on the sample signal; the soft threshold algorithm is used to denoise the signal and enhance the fault features to reduce the calculation time;

[0202] Based on a single hidden layer neural network, an adaptive multi-kernel extreme learning machine is used for fault diagnosis:

[0203]

[0204] Among them, β j is the output weight, ω j is the input weight, s j is the bias, g( ) is the activation function, x i is the input vector;

[0205] In order to reduce the uncertainty of a single kernel, a multi-kernel ELM (MKL) is adopted, and the optimal kernel is assumed to be a base kernel of a linear combination of a set of kernels, and the optimal combination coefficients and structural parameters of the classifier are jointly learned by maximizing the margin;

[0206]

[0207] Among them, Ω(X, x k )=C1K poly (X, x k )+C2K RBF (X, x k )+C3K wave (X, x k ), C i , i=1, 2, 3 are the weights corresponding to the basis functions, ∑C i =1. The base kernel functions are:

[0208]

[0209] Among them, K poly (X, x k ) is the polynomial kernel function, K RBF (X, x k ) is the radial basis function kernel, K wave (X, x k ) is the wavelet kernel function, C0 is a constant used to control the smoothness of the kernel function; σ is a parameter used to control the width of the kernel function; a is the scale parameter of the wavelet transform;

[0210] Online weight optimization: In order to improve the performance of MKL, a model combining particle swarm optimization (PSO) and genetic algorithm (GA) is used to optimize the above parameters C0, a, C i , σ, S are optimized, and the individual optimal value (pbest) and the global optimal value (gbest) are searched according to the best fitness function;

[0211] The fitness function is defined by the root mean square error (RMSE) of the signal verification set; the parameter group is optimized by the particle swarm algorithm, and the speed and position of the particles are updated using the formula; the individual optimal value and the global optimal value are updated by genetic operation; the optimal parameter set is obtained after the predetermined number of iterations is reached.

[0212] The edge end transmits the vibration signal features to the cloud end for use as diagnostic signal features. The vibration signal is used as a signal for further diagnosis in the cloud end. Its state features are extracted at the edge end, and its sparse coding is optimized online, and transmitted to the cloud end for use as diagnostic signal features.

[0213] The edge end transmits the vibration signal characteristics to the cloud for use as diagnostic signal characteristics, including:

[0214] On the edge computing platform, features and enhanced nodes are constructed based on local signals and combined into nodes through pseudo-inverse calculation;

[0215] Perform online optimization of adaptive sparse coding on feature signals, extract their state features, send enough feature nodes to the server domain, and evaluate the corresponding weights of the nodes;

[0216] Then the weights are optimized until the model converges, and an aggregation model for predicting the state of rolling bearings is established.

[0217] in,

[0218] The edge transmits the vibration signal features to the cloud. It is necessary to integrate the penalty into the width learning system and add the manifold regularization term, including:

[0219] The dynamic distribution adaptive penalty adopts the maximum mean difference (MMD) method, which is specifically:

[0220] MMD is a penalty used to measure and minimize the distance between source and target domain data:

[0221]

[0222] where F is a class of functions defined as the set of unit balls in the reproducing kernel Hilbert space; for two independent distributions p and q, if the maximum mean difference (MMD) between them is 0, then we can detect that there is no difference between them.

[0223] To integrate the penalty into the generalized learning system (BLS), the MMD is calculated and projected considering the given projection vector:

[0224]

[0225] D S is the dataset of the source domain, D T is the dataset of the target domain;

[0226] The original data is converted into mapping features through weight transformation; for the input data X, the i-th group of mapping features is calculated using the linear transformation function,

[0227]

[0228] Among them, W ei and β ei is a random weight with appropriate dimension; n is the number of feature node groups; is a linear transformation function;

[0229] The mapped features are then transformed into enhanced nodes through nonlinear transformation:

[0230]

[0231] in, and are random weights with appropriate dimension; ξ j ( ) is a nonlinear transformation function;

[0232] Where A=[Z n |H m ]=[a(x1) T , ..., a(x s+t ) T ] T , A∈R N×(nk+m) , N is the number of samples, nk+m is the total feature nodes;

[0233] Then minimize the squared loss prediction error:

[0234]

[0235] Among them, Y is the labeled sample corresponding to the input data X; β is the weight coefficient; θ is the regularization parameter to compensate for model error and complexity;

[0236] β=(A T A+θI) -1 A T Y (18)

[0237] In particular, we can get β *=lim θ→0 (θI+A T A) -1 A T ; When the number of data and model structure are fixed, β * And send it to the cloud for model aggregation;

[0238] The Manifold Regularization Framework (MRF) algorithm integrates a large number of unlabeled samples and a small number of labeled samples to explore the manifold features hidden in the data and express them as a loss function;

[0239] The manifold regularized MRF is specifically:

[0240] The MRF algorithm can explore the various characteristics of the data and express them as a loss function:

[0241]

[0242] in, is the predicted value; Tr(·) is the trajectory of the Laplace operator matrix L, which is obtained by formula (8):

[0243] L=DW (20)

[0244] Among them, the degree matrix D={D ii}, W={w (i，j)} is the adjacency matrix of the similarity between the source domain and the target domain data, which is calculated by the following formula:

[0245]

[0246] Among them, N k (x) is the k-neighboring value of sample x, i = 1…l+u, j = 1…l+u, l is the number of labeled samples, u is the number of unlabeled samples; σ is a hyperparameter used to control the scale in similarity calculation; x i is the data vector of the i-th sample;

[0247] According to formulas (20) and (21), (5) can be written as:

[0248]

[0249] Integrate real-time training accuracy and sample complexity into the online estimation of the model training process, and transfer the optimal features from the source domain (edge) to the target domain (cloud). The cloud can diagnose the type and level of faults with its complex computing capabilities, including:

[0250] On the edge computing platform, features and enhanced nodes are constructed based on local signals, and combined into nodes through pseudo-inverse calculation, and vibration signal features are extracted through adaptive sparse representation;

[0251] Based on manifold regularized MRF and dynamic distribution adaptive penalty MMD, the adaptive transfer learning algorithm is used for the diagnosis of cloud computing platform;

[0252] Send enough feature nodes to the server domain and evaluate the corresponding weights of the nodes;

[0253] Then the weights are optimized until the model converges, and an aggregate model for predicting the state of rolling bearings is established;

[0254] In the prediction stage, sample X (t-lu) Arrival time t-lu, input to the model generation In the subsequent time, X (t -lu) Repeated input to model generation These predicted values ​​should be equal to the true values. The error between the predicted values ​​and the true values ​​can be used to calculate the prediction error vector as feedback for estimating the training parameters.

[0255] A dynamic aggregation strategy is adopted to determine the characteristics of the prediction model and enhance the optimization of nodes, so that the feature network can adapt to complete the online prediction task, providing a new advantageous tool for the current online monitoring based on the coordination of indirect signals and diagnostic signal diagnosis.

[0256] The dynamic aggregation strategy determines the characteristics of the prediction model and enhances the optimization of nodes to adapt the feature network to complete the online prediction task, including:

[0257] The cloud computing server receives enough parameters β, takes the time delay of the vibration signal as the main model, calculates the maximum weight, and divides the remaining weights equally among other models to form a new fault diagnosis model.

[0258] The generalization performance of the generated model is verified based on the validation set, and the weights of the model received from the source domain are adjusted until the sovereignty value is reduced to convergence.

[0259] The terminal proposed in this embodiment and the method for behavior recognition proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to any of the above embodiments, and this embodiment has the same beneficial effects as executing the behavior recognition method.

[0260] In addition, an embodiment of the present application also proposes a computer-readable storage medium, which may be a non-volatile computer-readable storage medium, on which a behavior recognition program is stored, and when the behavior recognition program is executed by a processor, the method of the present application as described above is implemented.

[0261] The various embodiments of the electronic device and computer-readable storage medium of the present application can all refer to the various embodiments of the behavior recognition method of the present application, which will not be repeated here.

[0262] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0263] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0264] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling an electronic device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0265] The described embodiments are only a part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

Claims

1. A method for online health monitoring of key train components based on multi-sensor signal coordination, characterized in that: Specifically include: Based on the correlation analysis of multi-sensor signals, indirect signals are selected as early warning indicators of abnormal train status online in the offline stage; Wherein, the indirect signal is a signal strongly correlated with the vibration signal processed at the edge, and an online warning of abnormal train status is performed based on the indirect signal; The edge end transmits the vibration signal features to the cloud end for use as diagnostic signal features. The vibration signal is used as a signal for further diagnosis in the cloud end. Its state features are extracted at the edge end, and its sparse coding is optimized online, and transmitted to the cloud end for use as diagnostic signal features. Integrate real-time training accuracy and sample complexity into the online estimation of the model training process, and transfer the optimal features from the source domain (edge) to the target domain (cloud). The cloud can diagnose the type and level of faults with its complex computing capabilities. A dynamic aggregation strategy is adopted to determine the characteristics of the prediction model and enhance the optimization of nodes, so that the feature network can adapt to complete the online prediction task, providing a new advantageous tool for the current online monitoring based on the coordination of indirect signals and diagnostic signal diagnosis.

2. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 1 is characterized in that: The correlation analysis based on multi-sensor signals selects indirect signals as early warning indicators of abnormal train status online in the offline stage, specifically including: Phase space reconstruction: Reconstruct vibration signals and lightweight indirect signals with time delay and convert them into state space. State space reconstruction is used to reveal the nonlinear dynamic characteristics of signals and mine hidden information in time series signals. Correlation analysis of signal time series: The correlation between the vibration signal and the lightweight indirect signal after phase space reconstruction is analyzed using the Copula Entropy (CE) function. The Copula entropy function can effectively describe the correlation relationship and quantify the degree of dependence between the vibration signal and the lightweight indirect signal. According to the CE value between the vibration signal and the indirect signal, the correlation between the vibration signal and the indirect signal is estimated; Through dimensionality reduction and denoising, lightweight signals with high correlation with vibration signals are selected as auxiliary prediction signals to improve fault diagnosis performance.

3. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 2 is characterized in that: The phase space reconstruction is to reconstruct the vibration signal and the lightweight signal with time delay and convert them into the state space, which specifically includes: According to Takens’ embedding theorem, a time series X = [X1, X2, …, X L The i-th element X in ] i is reconstructed in state space as: X i =(x(i),x(i+τ),…,x(i+(n-1)τ)) (1) Where X is a time series, L means that there are L elements in the time series X, i = 1, 2, ..., L, i is a natural number, X i is the i-th point in the state space, x(i) is the state vector X i A component of , m is the embedding dimension, and τ is the time delay; The embedding dimension is estimated by pseudo nearest neighbor (FNN), which examines the distance variation between neighboring points in the state space; When the embedding dimension changes, the distance between FNNs does not change, indicating that the current embedding dimension is sufficient; The square of the Euclidean distance between the rth nearest neighbors of a coordinate vector is given by: in, is the square of the Euclidean distance between the rth nearest neighbors of the coordinate vector, t is the time, x(t+kτ) is the value of the original time series at time t+kτ, x (r) (t+kτ) is the value of the rth nearest neighbor corresponding to that moment; The criterion for determining the embedding parameter m is called false neighbors: Among them, R tol is a given threshold; Then, based on a geometric framework, the average displacement is applied to determine the time delay τ; Based on the time series x(t) reconstructed in phase space, calculate the phase point and its neighboring points The average distance S m (τ): Where N is the total number of data points; When the embedding parameter m is given: x i is the i-th data point in the time series. As τ increases, S m (τ) linear growth saturates, and the value corresponding to the end of its linear range is the optimal delay; when the curve S m When the slope of (τ) decreases to less than 40% of the initial value, it is the optimal value of the delay time.

4. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 2 is characterized in that: The correlation analysis of the signal time series uses the Copula entropy function to analyze the correlation between the vibration signal and the lightweight signal, specifically: The signal time series is a statistically independent measure defined by copula entropy; is a random variable with copula density c(u) and marginal distribution u; The copula entropy (CE) is defined as: in, is the copula entropy; is the mutual information, where copula entropy is defined as the negative of the mutual information; Joint Entropy of Multiple Random Variables Equal to the sum of marginal entropy and CE: According to the CE value between the vibration signal and the indirect signal, the lightweight signal with higher correlation with the vibration signal is selected as the auxiliary prediction signal.

5. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 1 is characterized in that: The processing of the indirect signal strongly correlated with the vibration signal at the edge and performing fault detection and early warning includes: Based on the collected multi-sensor signals, a training set and a valid set are constructed from indirect signals that are strongly correlated with vibration signals; Adaptive sparse representation of indirect signals and fault diagnosis based on multi-core extreme learning machines; Steps for troubleshooting at the edge: First, during the online monitoring process, the signal from the sensor is soft-thresholded to select the signal data stream for training; At each time t, the sparse representation of the signal is realized through adaptive sparse coding, and the signal is trained and predicted; For time t in the training phase, all information before time t is available, and the epoch is estimated dynamically. Input the model together with the target value to train the model and update the model weights; In the prediction stage, Input the multi-core ELM model to generate prediction values If the predicted value If the soft threshold is exceeded, is the predicted value of time step t, which is recorded for early warning. The vibration signal and the strongly correlated indirect signal feature vector are transmitted to the cloud for further processing and prediction; other prior values ​​are used as training samples for future training; The fault diagnosis based on multi-core extreme learning machine includes adaptive sparse representation, kernel function selection, and weight optimization: Adaptive sparse representation: Due to the sparsity of the signal, the diagnostic performance will rely on the sparse representation of the indirect signal, on which the indirect signal features are used for dictionary learning; the generalized minimax concave (GMC) is selected as the penalty function, and the K-singular value decomposition (K-SVD) is used to improve the matrix composition on the sample signal; the soft threshold algorithm is used to denoise the signal and enhance the fault features to reduce the calculation time; Based on a single hidden layer neural network, an adaptive multi-kernel extreme learning machine is used for fault diagnosis: Among them, β j is the output weight, ω j is the input weight, s j is the bias, g() is the activation function, x i is the input vector; In order to reduce the uncertainty of a single kernel, a multi-kernel ELM (MKL) is adopted, and the optimal kernel is assumed to be a base kernel of a linear combination of a set of kernels, and the optimal combination coefficients and structural parameters of the classifier are jointly learned by maximizing the margin; Among them, Ω(χ,x k )=C1K poly (χ,x k )+C2K RBF (χ,x k )+C3K wave (χ,x k ), C i , i=1,2,3 are the weights corresponding to the base kernel function, ∑C i =1. The base kernel functions are: Among them, K poly (χ,x k ) is the polynomial kernel function, K RBF (χ,x k ) is the radial basis function kernel, K wave (χ,x k ) is the wavelet kernel function, C0 is the polynomial kernel function parameter, which is used to control the smoothness of the kernel function; σ is the parameter used to control the width of the kernel function; a is the scale parameter of the wavelet transform; S is the regularization term coefficient; I is the unit matrix; Online weight optimization: In order to improve the performance of MKL, a model combining particle swarm optimization (PSO) and genetic algorithm (GA) is used to optimize the above parameters C0, a, C i , σ, S are optimized, and the individual optimal value (pbest) and the global optimal value (gbest) are searched according to the best fitness function; The fitness function is defined by the root mean square error (RMSE) of the signal verification set; the parameter group is optimized by the particle swarm algorithm, and the speed and position of the particles are updated using the formula; the individual optimal value and the global optimal value are updated by genetic operation; the iteration reaches the predetermined number of iterations to obtain the optimal parameter set.

6. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 1 is characterized in that: The edge end transmits the vibration signal characteristics to the cloud for use as diagnostic signal characteristics, including: On the cloud computing platform, based on the multi-sensor signal vector features transmitted by the edge, the feature nodes and enhanced nodes are constructed in combination with the historical signals on the cloud, and then combined into nodes through pseudo-inverse calculation; Perform online optimization of adaptive sparse coding on feature signals, extract their state features, send enough feature nodes to the server domain, and evaluate the corresponding weights of the nodes; Then the weights are optimized until the model converges, and a feature vector aggregation model for predicting the state of rolling bearings is established.

7. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 6 is characterized in that: The edge end transmits the vibration signal features to the cloud. It is necessary to use the penalty term to map the source domain multi-sensor signal features transmitted by the edge end and integrate them into the width learning system, and add the manifold regularization term, including: The dynamic distribution adaptive penalty adopts the maximum mean difference (MMD) method, which is specifically: MMD is a penalty used to measure and minimize the distance between source and target domain data: where F is a class of functions defined as the set of unit balls in the reproducing kernel Hilbert space; for two independent distributions p and q, if the maximum mean difference (MMD) between them is 0, then we can detect that there is no difference between them. Based on the penalty term, the source domain signal feature map is integrated into the target domain width learning system (BLS), the MMD is calculated and the given projection vector is considered for projection: D S is the dataset of the source domain, D T is the dataset of the target domain; The original data is converted into mapping features through weight transformation; for the input data X, the i-th group of mapping features is calculated using the linear transformation function, Among them, W ei and β ei is a random weight with appropriate dimension; n is the number of feature node groups; is a linear transformation function; The mapped features are then transformed into enhanced nodes through nonlinear transformation: in, and are random weights with appropriate dimension; ξ j () is a nonlinear transformation function; Where A=[Z n ∣H m ]=[a(x1) T ,…,a(x s+t ) T ] T ,A∈R N×(nk+m) , N is the number of samples, nk+m is the total feature nodes; Then minimize the squared loss prediction error: Among them, Y is the labeled sample corresponding to the input data X; β is the weight coefficient; θ is the regularization parameter to compensate for model error and complexity; β=(A T A+θI) -1 A T Y (18) Where I represents the identity matrix; In particular, we can get β * =lim θ→0 (θI+A T A) -1 A T ; When the number of data and model structure are fixed, β * And send it to the cloud for model aggregation; The Manifold Regularization Framework (MRF) algorithm integrates a large number of unlabeled samples and a small number of labeled samples to explore the manifold features hidden in the data and express them as a loss function; The manifold regularized MRF is specifically: The MRF algorithm can explore the various characteristics of the data and express them as a loss function: in, is the predicted value; Tr(·) is the trajectory of the Laplace operator matrix L, which is obtained by formula (8): L=DW (20) Among them, the degree matrix is the adjacency matrix of the similarity between the source domain and the target domain data, calculated by the following formula: Among them, N k (x) is the k-neighboring value of sample x, i = 1…l+u, j = 1…l+u, l is the number of labeled samples, u is the number of unlabeled samples; σ is a hyperparameter used to control the scale in similarity calculation; x i is the data vector of the i-th sample; According to formulas (20) and (21), (5) can be written as: in, is the predicted value.

8. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 1 is characterized in that: Integrate real-time training accuracy and sample complexity into the online estimation of the model training process, and transfer the optimal multi-sensor signal features from the source domain (edge) to the target domain (cloud). The cloud can diagnose the type and level of faults with its complex computing capabilities, including: Based on manifold regularized MRF and dynamic distribution adaptive penalty MMD, the adaptive transfer learning algorithm is used for the diagnosis of cloud computing platform; On the cloud computing platform, based on the multi-sensor signal vector features transmitted by the edge, the feature nodes and enhanced nodes are constructed in combination with the historical signals on the cloud, and then combined into nodes through pseudo-inverse calculation; Send enough feature nodes to the server domain and evaluate the corresponding weights of the nodes; Then the weights are optimized until the model converges, and an aggregate model for predicting the state of rolling bearings is established; In the prediction stage, sample X (t-lu) Arrival time t-lu, input to the model generation In the following time, X (t-lu) Repeated input to model generation These predicted values ​​should be equal to the true values. The error between the predicted values ​​and the true values ​​can be used to calculate the prediction error vector as feedback for estimating the training parameters.

9. The method for online health monitoring of key train components based on multi-sensor signal coordination according to claim 1 is characterized in that: The dynamic aggregation strategy determines the characteristics of the prediction model and enhances the optimization of nodes to adapt the feature network to complete the online prediction task, including: The cloud computing server receives enough parameters β, takes the time delay of the vibration signal as the main model, calculates the maximum weight, and divides the remaining weights equally among other models to form a new fault diagnosis model. The generalization performance of the generated model is verified based on the validation set, and the weights of the model received from the source domain are adjusted until the sovereignty value is reduced to convergence.

10. An electronic device, characterized in that: include: A memory, a processor, and an online health monitoring method program stored in the memory and executable on the processor, wherein when the online health monitoring method program is executed by the processor, the online health monitoring method for key components of a train based on multi-sensor signal coordination as described in any one of claims 1 to 9 is implemented.

Images