Method for predicting residual life of intelligent barrier machine of toll station based on ICEEMDAN-ISOMAP-Transformer

Abstract

The invention discloses an ICEEMDAN-ISOMAP-Transform-based method for predicting the residual life of an intelligent barrier machine of a highway toll station. The method comprises the following steps: 1, collecting vibration signals of a bearing in the intelligent barrier machine in the vertical direction and the horizontal direction; 2, the stability and accuracy of signal decomposition are improved through an ICEEMDAN algorithm; 3, performing time domain and frequency domain feature extraction work on the denoised vibration signals; 4, performing feature dimensionality reduction and reconstruction on the multi-dimensional time domain and frequency domain features by adopting an ISOMAP algorithm; and 5, taking the dimensionality-reduced and reconstructed features as the input of a Transform model, and realizing the prediction of the residual life of the intelligent barrier machine through model training. According to the method, how to establish the nonlinear relation between the original vibration signal and the residual life curve of the intelligent barrier machine is emphatically considered from the data driving direction, especially the work of signal noise reduction and multi-dimensional feature reconstruction, and the prediction performance of the Transform model can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of fault prediction and health management of mechanical equipment, and specifically is a method for predicting the remaining life of an intelligent barrier machine at a toll station based on ICEEMDAN-ISOMAP-Transformer. Background Art

[0002] With the significant increase in traffic volume on highways, the operation and management of highways are facing unprecedented challenges. Especially at toll stations, due to the frequent use of smart barrier machines, the physical wear and aging of equipment is extremely rapid. Equipment failure not only affects vehicle traffic efficiency and increases operating costs, but may also cause traffic congestion, bringing many inconveniences to highway operation and management.

[0003] The replacement and maintenance methods of traditional smart barrier machines require a lot of money. From purchasing new equipment, installation and commissioning to daily maintenance, each link requires professional technicians and high expenses. The remaining life prediction of smart barriers can help decision makers determine whether existing equipment should be repaired, upgraded or directly replaced, and plan these operations to minimize operational interruptions and costs. At present, the research on the remaining life prediction of smart barrier machines at highway toll stations has not effectively integrated and applied signal processing, data analysis and deep learning technology methods to this field. The existing physical models have the shortcomings of inaccuracy and unbalanced generalization ability, and are also limited by the complexity of equipment failure mechanisms and the diversity of mechanical structures.

[0004] Therefore, a new technical solution is needed to solve the above problems. Summary of the invention

[0005] Purpose of the invention: The present invention aims to provide a method for predicting the remaining service life of an intelligent barrier machine with high prediction accuracy and strong generalization ability. The method provides accurate remaining service life prediction results for highway managers through monitoring data of the vibration of the bearing of the intelligent barrier machine, so as to ensure the stability and continuity of highway operation.

[0006] Technical solution: A method for predicting the remaining life of intelligent barriers at toll stations based on ICEEMDAN-ISOMAP-Transformer, including the following steps:

[0007] (1) Collect vibration signals from barrier machines;

[0008] (2) Use the ICEEMDAN algorithm to restore the real vibration signal;

[0009] (3) The key time domain and frequency domain features of the calculated real vibration signal;

[0010] (4) Using the ISOMAP algorithm to reduce the dimension and reconstruct the key time domain features and frequency domain features;

[0011] (5) Build the Transformer model;

[0012] (6) Input the processed feature data into the Transformer model for training;

[0013] (7) Use the trained prediction model to predict the RUL of the intelligent barrier machine.

[0014] Furthermore, the vibration signal of the barrier machine collected in step (1) is the vibration data of the rolling bearing in the barrier machine in the vertical and horizontal directions collected by the dynamic signal collector.

[0015] Furthermore, step (2) uses the ICEEMDAN algorithm to perform modal decomposition and signal denoising on the vibration signal collected in step (1) to restore the real vibration signal, specifically:

[0016] 2.1) Add the special noise component separated by EMD algorithm to the original signal, that is,

[0017] x (i) =x+β 0 E 1 [W (i) ]

[0018] Where x represents the initial signal; β 0 The coefficient of white noise added to the original signal is used to control the intensity of the noise during the signal decomposition process; E 1 [W (i) ] represents special noise; W (i) Indicates adding the i-th white noise in the signal to help separate the different modes of the signal;

[0019] 2.2) Calculate the local average value of the transformed signal, extract the first residual, and perform the decomposition analysis of the first residual into the intrinsic mode function, that is,

[0020]

[0021] IMF 1 =xR 1

[0022] Among them, R 1 represents the first-order residual; I represents the number of sampling points of the signal, that is, the signal length; M[x (i) ] indicates the signal x (i) The i-th mode function obtained after modal decomposition; IMF 1 Indicates the first IMF value;

[0023] 2.3) Calculate the second IMF value

[0024]

[0025] IMF 2 =R 1 -R 2

[0026] Among them, R 2 represents the second-order residual; β 1 represents the noise coefficient used in the second IMF decomposition step; E 2 [W (i) ] represents the second group of white noise added in the decomposition process; IMF 2 Indicates the second IMF value;

[0027] 2.4) Repeat the above steps and keep using the formula until the Kth IMF value is found.

[0028]

[0029] IMF k =R k-1 -R k

[0030] Among them, R k represents the k-order residual; β k-1 represents the noise coefficient used in the k-th IMF decomposition step; E k [W (i) ] represents the second group of white noise added in the decomposition process; N represents the total number of IMFs decomposed in this algorithm; IMF k represents the kth IMF value, and the IMF components are obtained by iterative calculation and decomposition;

[0031] 2.5) For the decomposed IMF components, the Spearman correlation coefficient is used to evaluate the monotonic relationship between the two variables, that is,

[0032]

[0033] Among them, ρ N represents the Spearman correlation coefficient between the Nth IMF component and the original signal; X t represents the original signal value at time t; represents the value of the Nth IMF component at time t; and They represent the average values ​​of the original signal and the IMF component respectively;

[0034] The components with high correlation are extracted and reconstructed, and the reconstructed signal is the restored real vibration signal.

[0035] Furthermore, the specific steps of step (3) are as follows:

[0036] 3.1) According to the expert prior knowledge, the representative time domain feature parameters are extracted from the full life cycle signal data of the rolling bearing of the barrier machine; the time domain features are divided into two categories: dimensioned and dimensionless, and the mean Standard Deviation Square root amplitude RMS Peak-to-peak value is a dimensional characteristic; skewness Kurtosis Peak Index Margin index Waveform indicators Pulse Indicator It is a dimensionless feature; a total of 11 time domain features can be obtained;

[0037] 3.2) Based on the expert’s prior knowledge, the frequency domain is converted by fast Fourier transform to obtain 12 frequency domain features:

[0038] Furthermore, the specific steps of step (4) are as follows:

[0039] 4.1) Draw a histogram density curve and a correlation coefficient heat map of all the features obtained in step (3), and pre-screen the features according to the correlation coefficient heat map;

[0040] 4.2) Use the ISOMAP algorithm to reduce the dimension of the pre-screened features and reconstruct them. The ISOMAP algorithm process is as follows:

[0041] Assume that the number of sample data is n, and the high-dimensional space sample set is X = {x i |i=1,2,...,n},X∈R N , x i and x j For any two samples in the sample set; calculate sample x i and x j The Euclidean distance d e (x i , x j ), when x i and x j When the distance between them is less than the k-neighborhood of i and j, then x is considered i and x jThe k-neighborhood refers to the neighborhood formed by the k nearest neighbors of each data point in the high-dimensional space; the neighboring points are connected by lines to establish a weighted neighborhood graph G of high-dimensional data, where the edges of G are neighboring edges e ij The weight of the neighboring edge is the edge connecting the data points in the constructed weighted neighborhood graph, and the edge length of G is equal to x i and x j The Euclidean distance d e (x i , x j );

[0042] Calculate the geodesic distance d in the neighborhood graph G G (x i , x j ), the shortest path between two points on the neighboring graph G is used to approximate the geodesic distance between two points on the manifold structure, and the Floyds algorithm or Dijkstras algorithm is used to measure any two samples x i and x j The shortest distance between E (x i , x j ), if x i and x j If there is a connection between G (x i , x j ) has an initial value of d E (x i , x j ), otherwise let d G (x i , x j )=∞, where d G (x i , x j ) is,

[0043]

[0044] The geodesic distance matrix D is obtained from the shortest paths of all points in the graph G. G ={d G (x i , x j )), combined with the above formula to determine the optimal field parameter k,

[0045] k = argmin(1-ρ 2 D(x)D(y))

[0046] Where ρ is the parameter of the optimal neighborhood parameter k, which is a hyperparameter used to determine the neighborhood size of each point when constructing the k-nearest neighbor graph; D(x) represents the geodesic distance matrix in high-dimensional space, where the element d of the matrix is x (xi , x j ) represents the sample point x i and x j The geodesic distance between them; D(y) represents the distance matrix in the low-dimensional space, where the elements d of the matrix y (x i , x j ) represents the sample point y after mapping in the low-dimensional space i and j The Euclidean distance between

[0047] Construct a d-dimensional embedding and apply the MDS method to the geodesic distance matrix D G , minimize the objective function to obtain the low-dimensional coordinate Y,

[0048]

[0049] Where ω(Y) represents the objective function for dimensionality reduction, which represents the error of the high-dimensional space X reconstructed in the low-dimensional space Y; τ is the operation matrix operator, which is used to convert the distance matrix into a form suitable for eigendecomposition; D Y is the Euclidean distance matrix between sample points after dimensionality reduction, and H is the unit matrix;

[0050] Characteristic decomposition τ(D G ), take the first d eigenvalues ​​λ 1 ,λ 2 , ..., λ d and the eigenvector V 1 , V 2 , ..., V d And form a matrix U = (u 1 ,u 2 , ..., u d ), and arrange them from large to small according to the eigenvalues ​​to construct a d-dimensional embedding that retains the characteristics of the original data.

[0051]

[0052] The dimension d is an implicit parameter in the low-dimensional embedding space.

[0053] 4.3) The features reconstructed in 4.2) are processed by cumulative feature processing and normalization. The formula is as follows:

[0054]

[0055] In the formula, F jn refers to the jth feature f j (i) Cumulative transformation results in n samples, where N is the total number of features.

[0056] Furthermore, 4.3) normalizes the accumulated feature x, and the formula is as follows:

[0057]

[0059] Furthermore, the model constructed in step (5) is a Transformer model with a module stacking architecture containing a multi-head attention mechanism. The encoder of the model is composed of multiple identical layers stacked together, and each layer has two sub-layers, including:

[0060] The first sub-layer is the Multi-Head Self-Attention mechanism, i.e.

[0061] MultiHead(Q,K,V)=Concat(Head 1 , ..., Head N )W O

[0062] Head i =Attention(QW i Q , KW i K , VW i V )

[0063] Where i = 1, 2, ..., N represents the number of heads in the multi-head attention; W O represents the final projection matrix of the entire multi-head attention head; W i Q and W i K The shape is (d model , d K ), W i V The shape is (d model , d V ), all three of which are learnable parameter matrices, d model Represents the dimension of the entire sequence, d K =d V =d model / N;

[0064] The second sublayer is the Position Wise Feed-Forward Network, which is

[0065] FFN=max(0,XW 1 +b 1 )W 2 +b2

[0066] Where X represents the input sequence; W 1 and W 2 is the learnable parameter matrix of the position feedforward network; b 1 and b 2 are the corresponding bias terms respectively.

[0067] Furthermore, each of the sub-layers adopts residual connection, and layer normalization is applied immediately after the addition calculation of the residual connection; an encoder-decoder attention layer is inserted into the two sub-layers of the decoder, and the layer queries the output from the previous decoder, while the key and value come from the output of the entire encoder; the attention of the decoder adopts masking to retain the autoregressive property, and a matrix with lower triangular elements initialized to 0 and upper triangular elements all negative infinity is added to the attention matrix. After softmax normalization, it ensures that the prediction depends only on the generated output units.

[0068] Beneficial effects: The present invention has the following significant effects: 1. The present invention attempts for the first time to integrate technologies such as signal processing, data analysis and deep learning, and proposes a data-driven remaining life prediction method for intelligent barrier machine equipment at highway toll stations, and obtains relatively excellent prediction results; 2. The multi-algorithm fusion model proposed in the present invention avoids the limitations of many previous models, avoids the limitations of building physical models with unbalanced accuracy and generalization capabilities, and the limitations of the fusion of physical and data-driven models that require online measurement data for updating; 3. The present invention does not need to consider the complexity of the equipment failure mechanism and the diversity of mechanical structures, and only needs to use the bearing vibration data of the barrier machine to accurately predict the remaining life; 4. The present invention also takes into account the variability of working conditions, and experiments are carried out on the same equipment with inconsistent working conditions on the data set, proving that the fusion model proposed in the present invention has strong learning and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is the overall flow chart of the present invention.

[0070] Figure 2 A schematic diagram of a vibration signal of a certain data set used in the present invention.

[0071] Figure 3 This is a schematic diagram of the modal components of the original vibration signal after being decomposed using the ICEEMDAN algorithm in the present invention.

[0072] Figure 4 It is a histogram density curve diagram of all features under certain conditions of the present invention.

[0073] Figure 5It is a heat map of correlation coefficients of all features and RUL feature values ​​under certain conditions of the present invention.

[0074] Figure 6 It is a line graph of residual variance and reconstruction error, domain parameters and feature dimensions under certain conditions of the present invention.

[0075] Figure 7 This is the Transformer prediction model architecture used in the present invention.

[0076] Figure 8 This is a diagram showing the remaining life prediction effect of the present invention under working condition 1.

[0077] Fig. 9 This is a diagram showing the remaining life prediction effect of the present invention under working condition 2.

[0078] Fig.10 This is a diagram showing the remaining life prediction effect of the present invention under working condition 3. DETAILED DESCRIPTION

[0079] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods.

[0080] like Figure 1 As shown, the implementation process of a method for predicting the remaining life of an intelligent barrier machine based on ICEEMDAN-ISOMAP-Transformer includes the following steps:

[0081] Step 1: Collect the vertical and horizontal vibration signals of the rolling bearing in the barrier machine through the dynamic signal collector. Install a portable dynamic signal collector in the barrier machine and set the relevant parameters. For example, set the sampling frequency f = 25.6Hz, the sampling interval Δt = 1min, and the sampling time t = 1.28s each time. Every time the barrier machine starts working, the collected vibration signal is stored in the form of csv or database, where the first column of the data table is the horizontal vibration signal, and the second column is the vertical vibration signal. The files are named in the order of sampling time.

[0082] Step 2: Use the ICEEMDAN algorithm to perform modal decomposition and signal denoising on the original vibration signal;

[0083] The present invention is experimentally demonstrated using an open source dataset from Xi'an Jiaotong University. Figure 2 It is the vibration signal diagram of a bearing in the horizontal and vertical directions under some working conditions of the data set.

[0084] The first step is to add the special noise part x separated by the EMD algorithm to the original signal. (i) =x+β 0 E 1 [W (i)]; the second is to calculate the local average of the transformed signal, extract the first residual and decompose it into the intrinsic mode function

[0085]

[0086] Then calculate the second-order residual

[0087]

[0088] and the second IMF value

[0089] IMF 2 =R 1 -R 2

[0090] Repeat the previous calculation process until the Kth residual is found

[0091]

[0092] and IMF valueIMF k =R k-1 -R k .

[0093] When using the ICEEMDAN algorithm for experiments, by setting the noise standard deviation to 0.2, the number of signal realizations to 100, and the maximum number of screening iterations to 5000, the modal components of the original vibration signal after decomposition are obtained. Figure 3 As shown. Next is the problem of component selection. The present invention proposes to use the Spearman correlation coefficient to extract IMF components, and based on multiple groups of experiments, select the components with correlation coefficient values ​​higher than 0.35 to reconstruct the IMF components to achieve noise reduction.

[0094] Step 3: Calculate the key time domain features and frequency domain features of the denoised vibration signal;

[0095] Table 1 and Table 2 are feature calculation tables in the time domain and frequency domain, respectively. There are 11 time domain features and 12 frequency domain features. Considering the vibration signals in both horizontal and vertical directions, 46 features are initially constructed as input. Figure 4 It is a histogram density curve of all characteristics of a certain type of bearing under a certain working condition.

[0096] Table 1 Time domain characteristic indicators

[0097]

[0098] Table 2 Frequency domain characteristic indicators

[0099]

[0100]

[0101] Step 4: Reduce the dimension and reconstruct the initial features;

[0102] First, a variety of features are quickly pre-screened through the correlation coefficient heat map to eliminate features with little correlation with RUL. The heat map is as follows Figure 5 shown.

[0103] Secondly, the ISOMAP algorithm is used to perform nonlinear dimensionality reduction on the pre-screened features. First, the domain graph G is constructed. Assuming that the number of sample data is n, the high-dimensional space sample set is X = {x i |i=1,2,...,n},X∈R N x i and x j For any two samples. Calculate sample x i and x j The Euclidean distance d e (x i , x j ), when x i and x j If the distance between them is less than the k-domain of i and j, then x is considered i and x j Adjacent, lines connect adjacent points, and establish a weighted domain graph G of high-dimensional data, where the edges of G are adjacent edges e ij The weight of x is equal to i and x j The Euclidean distance d e (x i , x j );

[0104] Then, calculate the geodesic distance d in the domain graph G G (x i , x j ). Considering that the calculation of geodesic distance is very difficult, the shortest path between two points on the neighboring graph G is used to approximate the geodesic distance between two points on the popular structure. According to the Floyds algorithm or Dijkstra algorithm, any two samples x are measured. i and x j The shortest distance between E (x i , x j ), if x i and x j If there is a connection between G (x i , x j ) has an initial value of d E (x i , x j ), otherwise let d G (xi , x j )=∞; Next, construct the d-dimensional embedding and apply the MDS method to the geodesic distance matrix D G , minimize the objective function to obtain the low-dimensional coordinate Y.

[0105] When using this algorithm for feature dimensionality reduction, the most important thing is to pay attention to the optimal field parameter k = argmin (1-ρ 2 D(x)D(y)) and the feature dimension n after dimensionality reduction, because the size of these two parameter values ​​determines different dimensionality reduction effects. Figure 6 It is a schematic diagram for determining the optimal k and n. When making a selection, the parameter value with the smallest residual variance is given priority, followed by comparing which conditions have the smallest reconstruction error to determine the optimal domain parameters and the feature dimension after dimensionality reduction. In the specific implementation steps, it can be seen that when k = 26, n = 5, the residual variance is 0.009419, and when k = 23, n = 6, the residual variance is 0.0094206. The residual variances of these two cases are almost the same and tend to the minimum value. At this time, the reconstruction error is compared and it can be seen that the reconstruction error of the latter is 760.790, which is smaller than the former's 777.938. Therefore, the optimal domain parameter k = 23 and the feature dimension n = 6 after dimensionality reduction can be determined.

[0106] Step 5: Use cumulative feature processing and normalization processing on the denoised features;

[0107] First, cumulative feature processing is performed, and the formula is

[0108]

[0109] Because the use of the barrier machine is a natural process, the cumulative transformation can be used to accumulate and scale the features in the time series at the same time, so that the features have better trend and better characterize the degradation information of the bearing; then the accumulated features are normalized, and the formula is:

[0110]

[0111] To eliminate the influence of dimension and improve the stability of calculation, accelerate the convergence of model and improve the generalization ability.

[0112] Step 6: Build a Transformer model with a modular stacking architecture containing a multi-head attention mechanism;

[0113] The present invention uses a prediction model of Transformer architecture to accurately predict the RUL of the intelligent barrier machine. The model structure is as follows: Figure 7 This prediction model mainly uses the core mechanism of multi-head attention and the encoder-decoder architecture to process sequence data and capture long-distance dependencies in time series.

[0114] The encoder and decoder of Transformer are composed of a stack of self-attention modules. The embedded representations of the input sequence and the target sequence are encoded with positions and then input into the encoder and decoder respectively. The encoder of Transformer is composed of multiple identical layers, each of which has two sublayers.

[0115] The first sublayer is a multi-head self-attention mechanism, expressed as MultiHead(Q, K, V) = Concat(Head 1 , ..., Head N )W o , Head i =Attention(QW i Q , KW i K , VW i V ), where i = 1, 2, ..., N represents the number of heads in the multi-head attention; Q, K, V represent query, key, and value respectively; W i Q and W i K The shape is (d model , d K ), W i V The shape is (d model , d V ), all three of which are learnable parameter matrices; d model Represents the dimension of the entire sequence, d K =d V =d model / N.

[0116] Position encoding is to enable the model to understand the position information in the sequence. It uses a combination of sine and cosine functions to ensure that the embedding vector of each position is unique. The second sublayer of each Transformer layer is a position-based feedforward network, which is expressed as FFN(X)=max(0, XW 1 +b 1 )W 2 +b 2 , each sub-layer uses residual connection, and layer normalization needs to be applied immediately after the residual connection addition calculation. The purpose of residual connection and layer normalization is to solve the gradient disappearance problem in deep network training.

[0117] The decoder part of the model is also similar, but the decoder has an encoder-decoder attention layer inserted in two of its sublayers. What is special about this layer is that the query comes from the output of the previous decoder, while the key and value come from the output of the entire encoder. The decoder's attention is then masked to preserve the autoregressive property. This is done by initializing a matrix with the lower triangular elements being 0 and the upper triangular elements being negative infinity, and adding it to the attention matrix. After softmax normalization, it ensures that the prediction depends only on the generated output units.

[0118] In terms of hyperparameter setting, the prediction model uses the Adam optimizer for model optimization, with an initial learning rate of 0.05. After every 60 training iterations, the learning rate is multiplied by 0.1, and the model is trained 180 times in total. The number of heads in the multi-head self-attention mechanism is the feature dimension n after dimensionality reduction in step 4, and the number of encoder layers is set to 2.

[0119] Step 7: Input the processed feature data into the Transformer model for training;

[0120] The preprocessed and feature-extracted intelligent barrier bearing vibration data is input into the constructed Transformer model for training. During the training process, the model learns the mapping relationship from the input sequence to the output sequence and adjusts the model parameters by optimizing the loss function.

[0121] Step 8: Use the trained prediction model to predict the life of other categories of equipment to test the generalization ability of the model.

[0122] After the training is completed, the model is used to predict the RUL of equipment data that did not participate in the training to test the generalization ability of the model. The prediction results of the model will be compared with the actual equipment maintenance and replacement records to verify the accuracy and practicality of the model. The evaluation indicators of the prediction results are RMSE, MAE, sMAPE and NSE. The evaluation results are shown in Table 3. Combining the evaluation indicators of the remaining life prediction effect of different equipment under 3 different working conditions, it can be seen that the Transformer model is 49.76%, 53.45%, 57.30% and 2.80% higher than the LSTM model in RMSE, MAE, sMAPE and NSE indicators. The comparison diagram of the results of the three types of prediction tasks is shown in the figure. Figure 8-10 , the three figures correspond to three different working conditions 1, 2, and 3. It is worth noting that the method proposed in the present invention only learns the training set data of working condition 2 and then predicts the test set data of the three working conditions, while the LSTM model uses the aforementioned signal processing and feature engineering methods of the present invention to learn based on the training set of the current working condition and predict the test set under the working condition.

[0123] It can be concluded from the evaluation indicators that the present invention can effectively model vibration data and fully explore the relationship between the implicit features in the vibration data and the remaining life. The prediction method proposed in the invention has high prediction accuracy and is of great significance to the actual equipment operation and maintenance.

[0124] Table 3 Comparison of RUL prediction effects

[0125]

Claims

1. A method for predicting the remaining life of intelligent barriers at toll stations based on ICEEMDAN-ISOMAP-Transformer, characterized in that: The following steps are involved: (1) Collect vibration signals from barrier machines; (2) Use the ICEEMDAN algorithm to restore the real vibration signal; (3) The key time domain and frequency domain features of the calculated real vibration signal; (4) Using the ISOMAP algorithm to reduce the dimension and reconstruct the key time domain features and frequency domain features; (5) Build the Transformer model; (6) Input the processed feature data into the Transformer model for training; (7) Use the trained prediction model to predict the RUL of the intelligent barrier machine.

2. The remaining life prediction method of the intelligent barrier machine at a toll station according to claim 1 is characterized in that: The vibration signal of the barrier machine collected in step (1) is the vibration data of the rolling bearing in the barrier machine in the vertical and horizontal directions collected by the dynamic signal collector.

3. The remaining life prediction method of the intelligent barrier machine at a toll station according to claim 1 is characterized in that: Step (2) uses the ICEEMDAN algorithm to perform modal decomposition and signal denoising on the vibration signal collected in step (1) to restore the real vibration signal. The specific steps are as follows: 2.1) Add the special noise component separated by EMD algorithm to the original signal, that is, x (i) =x+β0E1[W (i) ] Where x represents the initial signal; β0 represents the coefficient of white noise added to the original signal, which is used to control the intensity of noise during the signal decomposition process; E1[W (i) ] represents special noise; W (i) Indicates adding the i-th white noise in the signal to help separate the different modes of the signal; 2.2) Calculate the local average value of the transformed signal, extract the first residual, and perform the decomposition analysis of the first residual into the intrinsic mode function, that is, IMF1=x-R1 Among them, R1 represents the first-order residual; I represents the number of sampling points of the signal, that is, the signal length; M[x (i) ] indicates the signal x (i) The i-th mode function obtained after modal decomposition; IMF1 represents the first IMF value; 2.3) Calculate the second IMF value IMF2=R1-R2 Where R2 represents the second-order residual; β1 represents the noise coefficient used in the second IMF decomposition step; E2[W (i) ] represents the second group of white noise added in the decomposition process; IMF2 represents the second IMF value; 2.4) Repeat the above steps and use the formula until the Kth IMF value is obtained. IMF k =R k-1 -R k Among them, R k represents the k-order residual; β k-1 represents the noise coefficient used in the k-th IMF decomposition step; E k [W (i) ] represents the second group of white noise added in the decomposition process; N represents the total number of IMFs decomposed in this algorithm; IMF k represents the kth IMF value, and the IMF components are obtained by iterative calculation and decomposition; 2.5) For the decomposed IMF components, the Spearman correlation coefficient is used to evaluate the monotonic relationship between the two variables, that is, Among them, ρ N represents the Spearman correlation coefficient between the Nth IMF component and the original signal; X t represents the original signal value at time t; represents the value of the Nth IMF component at time t; and They represent the average values ​​of the original signal and the IMF component respectively; The components with high correlation are extracted and reconstructed, and the reconstructed signal is the restored real vibration signal.

4. The method for predicting the remaining life of the intelligent barrier machine at a toll station according to claim 1 is characterized in that: The specific steps of step (3) are as follows: 3.1) Extract time domain characteristic parameters from the real vibration signal of the barrier machine, including mean, standard deviation, root square amplitude, root mean square, peak-to-peak value, skewness, kurtosis, peak index, margin index, waveform index and pulse index; 3.2) Perform frequency domain conversion through fast Fourier transform to obtain frequency domain features corresponding to the time domain features in 3.1).

5. The remaining life prediction method of the intelligent barrier machine at a toll station according to claim 1 is characterized in that: The specific steps of step (4) are as follows: 4.1) Based on the features obtained in step (3), namely, the frequency domain features and the time domain features, a histogram density curve of each feature and a correlation coefficient heat map of all features are drawn, and the features are pre-screened based on the correlation coefficient heat map; 4.2) Use ISOMAP algorithm to reduce the dimension of pre-screened features and reconstruct them; 4.3) The features reconstructed in step 4.2) are processed by cumulative feature processing and normalization.

6. The method for predicting the remaining life of the intelligent barrier machine at a toll station according to claim 5 is characterized in that: The ISOMAP algorithm flow of step 4.2) is as follows: Assume that the number of sample data is n, and the high-dimensional space sample set is X = {x i |i=1,2,...,n},X∈R N , x i and x j For any two samples in the sample set; calculate sample x i and x j The Euclidean distance d e (x i , x j ), when x i and x j When the distance between them is less than the k-neighborhood of i and j, then x is considered i and x j The k-neighborhood refers to the neighborhood formed by the k nearest neighbors of each data point in the high-dimensional space; the neighboring points are connected by lines to establish a weighted neighborhood graph G of high-dimensional data, where the edges of G are neighboring edges e ij The weight of the neighboring edge is the edge connecting the data points in the constructed weighted neighborhood graph, and the edge length of G is equal to x i and x j The Euclidean distance d e (x i , x j ); Calculate the geodesic distance d in the neighborhood graph G G (x i , x i ), the shortest path between two points on the neighboring graph G is used to approximate the geodesic distance between two points on the manifold structure, and the Floyds algorithm or Dijkstras algorithm is used to measure any two samples x i and x j The shortest distance between E (x i , x j ), if x i and x j If there is a line between them, then d G (x i , x j ) has an initial value of d E (x i , x j ), otherwise let d G (x i , x j )=∞, where d G (x i , x j ) is, The geodesic distance matrix D is obtained from the shortest paths of all points in the graph G. G ={d G (x i , x j )), combined with the above formula to determine the optimal neighborhood parameter k, k=arg min(1-ρ 2 D(x)D(y)) Where ρ is the parameter of the optimal neighborhood parameter k, which is a hyperparameter used to determine the neighborhood size of each point when constructing the k-nearest neighbor graph; D(x) represents the geodesic distance matrix in high-dimensional space, where the element d of the matrix is x (x i , x j ) represents the sample point x i and x j The geodesic distance between them; D(y) represents the distance matrix in the low-dimensional space, where the elements d of the matrix y (x i , x j ) represents the sample point y after mapping in the low-dimensional space i and j The Euclidean distance between Construct a d-dimensional embedding and apply the MDS algorithm to the geodesic distance matrix D G , minimize the objective function to obtain the low-dimensional coordinate Y, Where ω(Y) represents the objective function for dimensionality reduction, which represents the error of the high-dimensional space X reconstructed in the low-dimensional space Y; τ is the operation matrix operator, which is used to convert the distance matrix into a form suitable for eigendecomposition; D Y is the Euclidean distance matrix between sample points after dimensionality reduction, and H is the unit matrix; Characteristic decomposition τ(D G ), take the first d eigenvalues ​​λ1, λ2, ..., λ d and eigenvectors V1, V2, ..., V d The matrix U = (u1, u2, ..., u d ), and arrange them from large to small according to the eigenvalues ​​to construct a d-dimensional embedding that retains the characteristics of the original data The dimension d is an implicit parameter in the low-dimensional embedding space.

7. The method for predicting the remaining life of the intelligent barrier machine at a toll station according to claim 5 is characterized in that: The cumulative feature processing in claim 5 (4.3) is as follows: In the formula, F jn refers to the jth feature f j (i) Cumulative transformation results in n samples, where N is the total number of features.

8. The method for predicting the remaining life of the intelligent barrier machine at a toll station according to claim 5 is characterized in that: In claim 5, 4.3) normalizes the accumulated feature x, and the formula is as follows:

9. The method for predicting the remaining life of an intelligent barrier machine at a toll station according to claim 1, characterized in that: The model constructed in step (5) is a Transformer model with a modular stacking architecture containing a multi-head attention mechanism. The encoder of this model is composed of multiple identical layers stacked together, and each layer has two sub-layers, including: The first sub-layer is the multi-head self-attention mechanism, i.e. MultiHead(Q,K,V)=Concat(Head1,...,Head N )W O Head i =Attention(QW i Q ,KW i K ,VW i V ) Where i = 1, 2, ..., N represents the number of heads in the multi-head attention; W O represents the final projection matrix of the entire multi-head attention head; W i Q and W i k The shape is (d model , d K ), W i V The shape is (d model , d V ), all three of which are learnable parameter matrices, d model Represents the dimension of the entire sequence, d K =d V =d model / N; The second sublayer is a position-based feedforward network, i.e. FFN = max(0, XW1+b1)W2+b2. Where X represents the input sequence; W1 and W2 are the learnable parameter matrices of the position feedforward network; b1 and b2 are the corresponding bias terms.

10. The method for predicting the remaining life of the intelligent barrier machine at a toll station according to claim 9, characterized in that: Each of the sub-layers uses a residual connection, and layer normalization is applied immediately after the residual connection addition calculation; An encoder-decoder attention layer is inserted between the two sublayers of the decoder. This layer queries the output from the previous decoder, while the keys and values ​​come from the output of the entire encoder. The decoder's attention uses masking to retain the autoregressive property. A matrix with lower triangular elements initialized to 0 and upper triangular elements all negative infinity is added to the attention matrix. After softmax normalization, it ensures that the prediction depends only on the generated output units.

Images