Health state assessment method for giant magnetostrictive bar

Abstract

The invention discloses a health state assessment method for a giant magnetostrictive bar. The method comprises the following steps: firstly, obtaining output displacement data, vibration data and loss data of the giant magnetostrictive bar; performing parameter optimization on the improved MPE method by using an AE algorithm, then effectively extracting multi-scale features of displacement data, vibration data and loss data by using selected parameters, converting the multi-scale features into gray matrixes, and respectively mapping the gray matrixes to R, G and B color channels to obtain a three-dimensional image feature set of the giant magnetostrictive bar; according to the method, CapsNet is improved, two convolution layers and a pooling layer are added in front of a capsule layer to compress the size of an image, meanwhile, the integrity of image features is kept as much as possible, an obtained three-dimensional image data set is input into the built CapsNet network for training, and a giant magnetostrictive bar detection model is obtained. According to the method, the health state of the giant magnetostrictive bar is evaluated, and the method plays a guiding role in development and application of transducers.

Description

Technical Field

[0001] The invention relates to the technical field of transducers, and in particular to a health status assessment method for a giant magnetostrictive rod. Background Art

[0002] Giant magnetostrictive rod (GMM) is the core material for making giant magnetostrictive transducers. Giant magnetostrictive rods will produce magnetostrictive strain under alternating excitation. The reliability index of the product is an important reference information for formulating maintenance plans and update cycles for the product. In order to allow the giant magnetostrictive transducer to work stably and permanently, it is necessary to study and evaluate the reliability of the giant magnetostrictive rod. However, the reliability of the giant magnetostrictive rod is affected by many factors, such as workload, temperature, etc., and in order to reduce the loss of the giant magnetostrictive rod, it is usually cut to reduce eddy current loss. Stress concentration will occur at the cutting seam, which will inevitably cause its mechanical strength to decrease, which brings requirements for the fault diagnosis of the giant magnetostrictive rod. By collecting the output data of the giant magnetostrictive rod, its state detection can be realized. In the present invention, the acceleration data, displacement data and loss data of the giant magnetostrictive rod are collected. If the health status assessment of the giant magnetostrictive rod is performed based on the 1-dimensional time domain signal, it will lead to problems such as difficult feature extraction and high computational overhead. Converting time domain signals into image data and using the powerful image recognition capabilities of neural networks to perform fault identification is a feasible route. Therefore, some scholars have proposed image coding methods to map one-dimensional data to the image field, and then use the powerful image processing capabilities of neural networks for classification, including Markov Transition Field (MTF), Gram angle field, recursive graph method, relative position matrix method (RPM) and Hilbert-Huang transform. However, traditional image coding methods are difficult to fully extract fault features, which affects the accuracy of health status assessment of giant magnetostrictive rods, and at the same time brings problems such as large training samples, complex network model structure, and high computational and time costs. In addition, the traditional multi-scale permutation entropy method ignores the information of adjacent coarse-grained sequences during the coarse-graining process.

[0003] Capsule Network (CapsNet) is a new neural network architecture proposed by famous scholar Geoffrey Hinton and others in 2017. The network innovatively uses vectors as output, solving the challenges faced by convolutional neural networks in processing spatial hierarchical relationships and posture changes in images. However, there is a problem that the capsule layer generates too many parameters.

[0004] Glossary:

[0005] AE optimization algorithm: Alpha evolution algorithm, is a new meta-heuristic algorithm (intelligent optimization algorithm), inspired by the alpha operator update solution with adaptive step size.

[0006] Multiscale permutation entropy method: An entropy measure based on the chi-square statistic.

[0007] CapsNet network: capsule neural network. Summary of the invention

[0008] The purpose of the invention is to provide a method for evaluating the health status of a giant magnetostrictive rod.

[0009] To achieve the above object, the technical solution of the present invention is as follows:

[0010] A method for evaluating the health status of a giant magnetostrictive rod comprises the following steps:

[0011] Step 1: Use particle swarm optimization to optimize the three optimization parameters of the multi-scale permutation entropy method: embedding dimension m, scale factor s and delay time t, and obtain the optimized multi-scale permutation entropy method;

[0012] Step 2, collecting acceleration signals, displacement signals and loss signals of the magnetostrictive rod to form an original data set;

[0013] Step 3, using the optimized multi-scale permutation entropy method to extract features from the original data set, with the acceleration signal as the R channel, the displacement signal as the G channel, and the loss signal as the B channel, to obtain a 3-channel feature map data set of the giant magnetostrictive rod;

[0014] Step 4: Add a pooling layer before the capsule layer of the CapsNet network, and add two convolutional layers before the pooling layer; the two convolutional layers and one pooling layer are used to compress the image size to obtain an optimized CapsNet network;

[0015] Step 5: Input the 3-channel feature map data set of the giant magnetostrictive rod into the optimized CapsNet network for training, and obtain the classification model after the training is completed;

[0016] Step 6. The acceleration signal, displacement signal and loss signal of the giant magnetostrictive rod to be evaluated are extracted using the optimized multi-scale permutation entropy method to obtain a 3-channel feature map data set, and then the 3-channel feature map data set is input into the classification model to obtain a health status assessment.

[0017] Further improvement, in step 1, the particle swarm algorithm is an AE optimization algorithm, and the specific steps of the AE optimization algorithm are as follows:

[0018] Step 3.1: Initialization,

[0019] X i =lb+(ub-lb)·rand(0,1,[1,D]),i=1,2,...,N(1.51)

[0020] Among them, X i represents the i-th candidate solution, lb and ub represent the lower bound and upper bound of the search space respectively, rand represents a random number uniformly distributed in the range [0,1], N represents the number of solutions; D represents the number of columns of generated random numbers;

[0021] Step 3.2: Construct the alpha operator. The set of candidate solutions is defined as the candidate matrix X. The subset of the candidate matrix is ​​the evolution matrix. During the evolution process, each solution X in the current candidate matrix is ​​not directly evolved. Instead, X is sampled with replacement operation to obtain the evolution matrix to be evolved:

[0022]

[0023] E represents the evolution matrix, → represents the replacement sampling operation, and N represents that the operation will be performed N times;

[0024] In the alpha operator, multiple steps of extracting and utilizing evolutionary information are performed simultaneously in one operator. The mathematical model is shown in the following formula:

[0025]

[0026] Where P is the adaptive basis vector, which determines the starting point of evolution; diagonal represents a function obtained by sampling by replacing D multiplied candidate solutions; where B is a K*D matrix obtained by sampling without replacement; A represents a D-order square matrix; represents the evolutionary solution of t+1 iterations, Δr represents the step size, θ represents the control parameter of the control differential vector, i.e., the adaptive step size, W i and L i represents the sampled solution obtained by sampling from the candidate matrix X;

[0027] P is generated by A or B itself and does not participate in evolution, which will result in no connection between P in each generation and information loss. Therefore, the concept of evolutionary path is introduced to increase the correlation between consecutive iterative steps. a and P b represents two independent evolutionary lines, then:

[0028]

[0029] Among them, c a Indicates P a The learning rate, c b Indicates Pb The learning rate; represents the evolutionary path at time t;

[0030] Step 3.3: Construct boundary constraints using the distance halving method,

[0031]

[0032] Among them, E i,j Indicates the boundary, ub indicates the upper bound of the search, and lb indicates the lower bound of the search;

[0033] Step 3.4: Greedy selection passes the evolved solution to the next generation in a simple way without any additional operations, introducing the greedy selection strategy into the AE optimization algorithm.

[0034]

[0035] represents the candidate solution at time t+1, represents the boundary at time t+1, Represents the boundary value at time t+1;

[0036] Step 3.5: Select the fitness function and construct the mean square error of each permutation entropy under different scale optimization parameters as the fitness function of the AE algorithm to evaluate the optimization results of the algorithm. The fitness function is as follows:

[0037]

[0038] Where n represents the nth fault state, m represents the total number of fault states, and H p (i) represents the permutation entropy value of the nth fault state, and Fitness represents the permutation entropy mean square error value; the permutation entropy mean square error values ​​at different scales are calculated, and the embedding dimension m of the multi-scale permutation entropy method corresponding to the maximum permutation entropy mean square error value, the scale factor s and the delay time t value, that is, the parameter combination with the best separability, is obtained, that is, the optimized multi-scale permutation entropy method.

[0039] As a further improvement, in step 2, the acceleration signal, displacement signal and loss signal are collected and obtained by an acceleration sensor, a laser displacement sensor and a BH sensor respectively.

[0040] For further improvement, the specific steps of step three are as follows:

[0041] Step 2.1: The acceleration signal, displacement signal and loss signal are respectively formed into time series of length n. For the time series of length n, X = {x i ,i=1,2,...,n} Use the root mean square value instead of the average value to reduce the feature loss caused by the coarse-grained process:

[0042]

[0043] Where s is the scale factor, j = 1, 2, ... [N / s], [N / s] means rounding down, x i represents the original signal subsequence, i represents the i-th data point; represents the average value of the original signal subsequence, y (s) Represents new sequences under different coarse-graining; represents the coarse-grained sequence at scale s; N represents the number of coarse-grained groups at each scale;

[0044] Step 2.2: Reconstruct the phase space of each coarse-grained sequence to obtain a new reconstructed sequence Y l(s) ,

[0045] Y l(s) ={y l(s) ,y l+1(s) ,…y (l+(m-1)λ)(s)} (1.48)

[0046] Where m is the embedding dimension, λ is the time delay, l is the reconstruction component; y (l+(m-1)λ)(s) represents the l+(m-1)λ(s)th coarse-grained subsequence;

[0047] Step 2.3: Reconstruct the sequence Y obtained in step 2.2 l(s) Arrange in ascending order, each coarse-grained sequence has a new sequence s(v)=(l 1 ,l 2 ,…l m ), where v = 1, 2, ..., V, V ≤ m! ; then the probability of various permutations at different scales is counted; m! represents the factorial number of permutation combinations at scale m, V represents the maximum number of coarse-grained sequences, l m Represents a subsequence sorted in ascending order;

[0048] Step 2.4: Calculate the entropy of the probability of various permutations at different scales and obtain the permutation entropy value.

[0049]

[0050] Among them, P v is the probability of the vth symbol appearing; H p (m) represents the entropy value of scale m;

[0051] Step 2.5: From step 2.4, we know that H p When (m) reaches the maximum value ln(m!), For H p (m) is normalized,

[0052]

[0053] H p is the final permutation entropy value of the signal, and H p The value is inversely proportional to the order of the signal, that is, the more chaotic the signal, the smaller the H p The larger the value, the greater the fault level.

[0054] H p Together with the original time series signal, it forms the feature matrix Y.

[0055]

[0056] Among them, [H p1 ,H p1 ,H p1 ,...,H ps ] is the multi-scale permutation entropy value of the signal, x 1 x 2 ...x n is the original sequence, Y is the newly constructed feature matrix, n is the number of sample points, s is the permutation entropy scale, and y is the eigenvalue of signal X after feature extraction; y ns represents the feature extraction value of the nth data under the scale s, x n represents the nth value of the original time series, H ps Represents the permutation entropy value at the sth scale;

[0057] Step 4.2: Perform standardization so that the feature matrix Y can be mapped to the grayscale matrix domain.

[0058]

[0059] Among them, y norm is the standardized eigenvalue, X max is the maximum value in the original sequence X, X min is the minimum value in the original sequence X;

[0060] Step 4.3: Set the resolution r, r of the generated image 2 ≤n×s,

[0061] Step 4.4: Read the Y matrix by column and place it by column to obtain the reconstructed feature matrix C.

[0062]

[0063] in [·]for Round up, c is the representation after rearranging y, c krRepresents the kth row and rth column data point after reconstruction;

[0064] Step 4.5: Perform grayscale value mapping and map the reconstructed feature matrix C to the grayscale value range:

[0065]

[0066] g i,j represents the gray value after mapping, c i,j represents the value of the i-th row and j-th column of matrix C, c max Represents the maximum value in matrix C;

[0067] Where i = 1, 2, 3 ... k; j = 1, 2, 3 ... r; and the gray value feature matrix G is obtained,

[0068]

[0069] Step 4.6: Perform a three-dimensional stacking of the obtained acceleration signal, displacement signal and loss signal gray value feature matrices to obtain a three-channel feature map data set of the giant magnetostrictive rod.

[0070] F m =[G a ; G v ; G w ](1.64)

[0071] Among them, F m The three-channel characteristic diagram of the giant magnetostrictive rod, G a Represents the gray value feature matrix of the acceleration signal, G v Represents the gray value feature matrix of the displacement signal, G w Gray value feature matrix representing the loss signal.

[0072] For further improvement, the specific steps of step 4 are as follows:

[0073] Step 5.1: Build convolutional layers and pooling layers.

[0074]

[0075] in, is the jth element of the lth layer; M j is the jth convolution region of the feature map of the l-1th layer; is the element in the i-th element of the i-1 layer; is the weight matrix corresponding to the convolution kernel; is the bias term; f(·) is the ReLU activation function;

[0076] Y (i,j) =max (p,q) (X(i+p,j+q) ) (1.66)

[0077] Where i and j represent the horizontal and vertical positions of the pixels in the output feature map, respectively; p and q represent the horizontal and vertical positions of the pixels in the window, respectively; Y (i,j) Represents the output matrix after maximum pooling; performs maximum pooling on the pooling layer after convolution, thereby reducing the number of parameters and the amount of calculation;

[0078] Step 5.2: Initialize the digital capsule layer,

[0079]

[0080] Among them, v i is the instantiation vector corresponding to category j; s i is the deep feature vector; c ij is the coupling coefficient, which indicates the connection strength between the capsules in the i-th layer and the capsules in the j-th layer; M ij represents the transformation matrix between the capsule in the i-th layer and the capsule in the j-th layer; u i is the i-th output vector of the initial capsule layer, c ij The update of the intermediate variable b ij Calculated by Softmax function;

[0081] Step 5.3: b ij The initial value is 0, which is determined by v i With the prediction vector Updated by the stateful routing algorithm, the dynamic routing algorithm iterates 3 times.

[0082]

[0083] Step 5.4: Use the margin loss function Margin as the loss function L of the CapsNet network k ,

[0084] L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 (1.69)

[0085] Among them, T k is the classification indicator function. When category k exists, T k =1, otherwise T k =0;||v k|| is the probability that the input is identified as category k; m + and m - are the upper and lower bounds respectively; λ is the proportional coefficient, and the loss function L k When the minimum value is obtained, the classification model is obtained after the training is completed.

[0086] The advantages of the present invention are as follows:

[0087] 1. Therefore, the present invention proposes to use an improved MPE method to extract the features of the displacement signal, acceleration signal and loss signal of the magnetostrictive rod respectively, obtain three feature sets, and then map them to the three color channels of R, G, and B respectively, to obtain a three-dimensional color image feature, construct an image feature set that can fully reflect the fault information of the giant magnetostrictive rod, and then use the CapsNet network for classification, thereby realizing the health status assessment of the giant magnetostrictive rod and guiding the development and application of the transducer.

[0088] 2. The present invention proposes an improved method, that is, the root mean square value is used instead of the average value in the coarse-graining step, so as to further fully extract the characteristic information of the sequence; since the calculation of MPE requires the selection of the embedding dimension, delay time and scale factor, they are often determined by artificial selection in the past, which is disadvantageous to the feature extraction. Therefore, the present invention proposes to use the AE algorithm to optimize these parameters. The AE algorithm is updated through the internal speed, and the principle is simpler, the parameters are fewer, and the implementation is easier. By constructing the mean square error value of each permutation entropy at different scales as the fitness function of the algorithm, the optimization result of the algorithm is evaluated, and the optimal embedding dimension, delay time and scale factor are obtained, so as to fully extract the multi-scale features of the three characteristic signals.

[0089] 3. In the present invention, in order to avoid too many parameters generated by the capsule layer, two convolution layers and one pooling layer are added before the capsule layer, which effectively compresses the image size without causing too much information loss. The obtained feature data set is input into the constructed capsule network for training to obtain an offline detection model, thereby realizing real-time health status assessment of giant magnetostrictive rods. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 This is a flow chart of the health status assessment method of giant magnetostrictive rods based on the improved multi-scale permutation entropy image encoding method and CapsNet network;

[0091] Figure 2 To improve the MPE image coding method;

[0092] Figure 3 It is the gray value mapping process;

[0093] Figure 4 It is the optimization process of AE algorithm;

[0094] Figure 5 To improve the CapsNet network structure. DETAILED DESCRIPTION

[0095] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. The preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0096] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification are only for describing specific implementation purposes and are not intended to limit the present invention.

[0097] Figure 1 FIG. 1 is a brief flow chart of a method for evaluating the health status of giant magnetostrictive materials based on three-channel image encoding of improved MPE and CapsNet corresponding to this embodiment. Figure 1 As shown, the implementation of this embodiment may include:

[0098] Step 1: Collect the acceleration signal, displacement signal and loss signal of the giant magnetostrictive rod.

[0099] In this embodiment, step 1 may include:

[0100] A vibration test platform for giant magnetostrictive rods was built. Acceleration signals, displacement signals and loss signals were collected respectively by accelerometers, laser displacement sensors and BH sensors to obtain the original data sets.

[0101] Step 2: The collected acceleration signal, displacement signal and loss signal contain a lot of noise, which makes it impossible to effectively identify the need for further feature extraction. In this step, a multi-scale permutation entropy method is used to extract features. Step 2 may include:

[0102] Step 2.1: Perform a coarse-graining operation on the sequence to obtain sequences at multiple scales. However, since MPE simply calculates the average value of each coarse-grained subsequence during the coarse-graining process, the information between the two sub-coarse-grained signals is ignored. Therefore, the present invention proposes to use the root mean square value instead of the average value, which can reduce the feature loss caused by the coarse-graining process.

[0103]

[0104] Where s is the scale factor, j = 1, 2, ... [N / s], [N / s] means rounding down, xi represents the original signal subsequence, represents the average value of the original signal subsequence, y (s) Represents new sequences under different coarse-graining;

[0105] Step 2.2: Reconstruct the phase space of each coarse-grained sequence to obtain a new reconstructed sequence.

[0106] Y l(s) ={y l(s) ,y l+1(s) ,…y (l+(m-1)λ)(s)} (1.48)

[0107] Where m is the embedding dimension, λ is the time delay, and l is the reconstruction component;

[0108] Step 2.3: Arrange the phase space sequence obtained in step 2.2 in ascending order. Each coarse-grained sequence has a new sequence s(v) = (l 1 ,l 2 ,…l m ), where v = 1, 2, ..., V, V ≤ m! ; then count the probabilities of various arrangements at different scales;

[0109] Step 2.4: Calculate the entropy of the probability of various permutations at different scales and obtain the permutation entropy value.

[0110]

[0111] Among them, P v is the probability of the vth symbol appearing.

[0112] Step 2.4: From step 2.4, we know that H p When (m) reaches the maximum value ln(m!), For H p (m) is normalized,

[0113]

[0114] H p Then it is the final permutation entropy value of the signal, and H p The value is inversely proportional to the order of the signal, that is, the more chaotic the signal, the smaller the H p The larger the value, the greater the fault level. p The larger the value.

[0115] Step 3: To improve MPE, three key parameters still need to be confirmed: embedding dimension m, scale factor s and delay time t. The selection of these three parameters is closely related to the ability of improving MPE to extract multi-scale features of signals. To this end, the present invention uses a more advanced AE optimization algorithm to optimize these three parameters.

[0116] In this embodiment, step 3 may include:

[0117] Step 3.1: Initialization,

[0118] X i =lb+(ub-lb)·rand(0,1,[1,D]),i=1,2,…,N(1.51)

[0119] Among them, X i represents a candidate solution, lb and ub represent the lower bound and upper bound of the search space respectively, rand represents a random number uniformly distributed in the range [0,1], and N represents the number of solutions;

[0120] Step 3.2: Construct the alpha operator. The set of candidate solutions is defined as the candidate matrix, and the evolution matrix is ​​a subset of the candidate matrix. In the evolution process, instead of directly evolving each solution X in the current candidate matrix, a sampling with replacement operation is performed to obtain the evolution matrix to be evolved.

[0121]

[0122] In this operator, multiple steps of extracting and utilizing evolutionary information are performed simultaneously in one operator, and its mathematical model is shown in the following formula:

[0123]

[0124] Among them, P is the adaptive basis vector, which determines the starting point of evolution; diagonal represents a function, which is obtained by sampling the candidate solution multiplied by D. Among them, B is a K*D matrix, which is obtained by sampling without replacement.

[0125] There are two evolutionary routes for alpha operators.

[0126]

[0127] Among them, c a Indicates P a The learning rate, c b Indicates P b The learning rate

[0128] Step 3.3: Construct boundary constraints using the distance halving method,

[0129]

[0130] Step 3.4: Greedy selection passes the evolved solution to the next generation in a simple way without any additional operations, and introduces the greedy selection strategy into the AE algorithm.

[0131]

[0132] Step 3.5: Select the fitness function and construct the mean square error of each permutation entropy at different scales as the fitness function of the algorithm to evaluate the optimization result of the algorithm. The fitness function is as follows:

[0133]

[0134] Where i represents the i-th fault state, m represents the total number of fault states, and H p (i) represents the permutation entropy value of the i-th fault state. By calculating the maximum mean square error of the permutation entropy under different scales, the parameter combination with the best separability can be obtained;

[0135] Step 4: After obtaining the three parameters of the improved multi-scale permutation entropy through the AE algorithm, feature extraction is performed on the acceleration signal, displacement signal and loss signal to obtain the feature matrix. The acceleration signal is used as the R channel, the displacement signal is used as the G channel, and the loss signal is used as the B channel to obtain the three-channel feature map data set of the giant magnetostrictive rod.

[0136] In this embodiment, step 4 may include:

[0137] Step 4.1: Use the selected parameters m, t, s to calculate the time series X = {x i , i=1,2,...,n}Use the method in Section 2.1 to calculate and obtain H p , and then combine it with the original time series signal to form a feature matrix M,

[0138]

[0139]

[0140] Among them, [H p1 ,H p1 ,H p1 ,...,H ps ] is the multi-scale permutation entropy value of the signal, X is the original sequence, Y is the newly constructed feature matrix, n is the number of sample points, s is the permutation entropy scale, and y is the eigenvalue of signal X after feature extraction.

[0141] Step 4.2: Normalize it so that it can be mapped to the grayscale matrix domain.

[0142]

[0143] Among them, y norm is the standardized eigenvalue, X max is the maximum value of signal X, X minis the minimum value of signal X;

[0144] Step 4.3: Set the resolution r, r of the generated image 2 ≤n×s,

[0145] Step 4.4: Read the Y matrix by column and place it by column to obtain the reconstructed feature matrix.

[0146]

[0147] in [·]for Round off, c is the representation after rearranging y.

[0148] Step 4.5: Perform grayscale value mapping and map the C matrix to the grayscale value range:

[0149]

[0150] Where i = 1, 2, 3 ... k; j = 1, 2, 3 ... r; and the gray value feature matrix is ​​obtained,

[0151]

[0152] Step 4.6: Stack the obtained gray value feature matrices of acceleration signal, displacement signal and loss signal by 3 to obtain a 3-channel feature map data set of the giant magnetostrictive rod.

[0153] F m =[G a ; G v ; G w ](1.64)

[0154] Among them, F m The three-channel characteristic diagram of the giant magnetostrictive rod, G a Represents the gray value feature matrix of the acceleration signal, G v Represents the gray value feature matrix of the displacement signal, G w Gray value feature matrix representing the loss signal.

[0155] Step 5: After obtaining the 3-channel feature map data set of the giant magnetostrictive rod, the large-size image will cause the capsule network to generate too many parameters and reduce the training speed. Therefore, the present invention adds 2 convolutional layers before the capsule layer to compress the image size while maintaining the integrity of the image features as much as possible. The CapsNet network is built, and the constructed three-dimensional image feature set is input into the network for training to obtain a classification model, which has realized the health status assessment of real-time data.

[0156] In this embodiment, step 5 may include:

[0157] Step 5.1: Build convolutional layers and pooling layers.

[0158]

[0159] Among them, X j l is the jth element of the lth layer; M j is the jth convolution region of the feature map of the l-1th layer; for the elements therein; is the weight matrix corresponding to the convolution kernel; is a bias term. f(·) is an activation function, and the present invention uses the ReLU function.

[0160] Y (i,j) =max (p,q) (X (i+p,j+q) ) (1.66)

[0161] Among them, i and j represent the positions of pixels in the output feature map, and p and q represent the positions of pixels in the window. The pooling layer after convolution is subjected to maximum pooling to reduce the number of parameters and the amount of calculation.

[0162] Step 5.2: Initialize the digital capsule layer,

[0163]

[0164] Among them, v i is the instantiation vector corresponding to category j; s i is the deep feature vector; c ij is the coupling coefficient; M ij is the transformation matrix; u i is the i-th output vector of the initial capsule layer. ij The update of the intermediate variable b ij This is calculated using the Softmax function.

[0165] Step 5.3: b ij The initial value is 0, which is determined by v i With the prediction vector Updated through the stateful routing algorithm. The dynamic routing algorithm generally iterates 3 times.

[0166]

[0167] Step 5.4: Use the margin loss function Margin as the loss function of CapsNet.

[0168] L k =T k max(0,m + -||v k ||)2 +λ(1-T k )max(0,||v k ||-m - ) 2 (1.69)

[0169] Among them, T k is the classification indicator function. When category k exists, T k =1, otherwise T k =0;||v k || is the probability that the input is identified as category k; m + and m - are the upper and lower bounds respectively; λ is the proportionality coefficient.

[0170] Step 5.5: Through the built CapsNet training, an offline model for evaluating the health status of giant magnetostrictive rods is obtained, and the offline model is used to perform real-time health status evaluation on giant magnetostrictive rods.

[0171] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for evaluating the health status of a giant magnetostrictive rod, characterized in that: The steps include: Step 1: Use particle swarm optimization to optimize the three optimization parameters of the multi-scale permutation entropy method: embedding dimension m, scale factor s and delay time t, and obtain the optimized multi-scale permutation entropy method; Step 2, collecting acceleration signals, displacement signals and loss signals of the magnetostrictive rod to form an original data set; Step 3, using the optimized multi-scale permutation entropy method to extract features from the original data set, with the acceleration signal as the R channel, the displacement signal as the G channel, and the loss signal as the B channel, to obtain a 3-channel feature map data set of the giant magnetostrictive rod; Step 4: Add a pooling layer before the capsule layer of the CapsNet network, and add two convolutional layers before the pooling layer; the two convolutional layers and one pooling layer are used to compress the image size to obtain an optimized CapsNet network; Step 5: Input the 3-channel feature map data set of the giant magnetostrictive rod into the optimized CapsNet network for training, and obtain the classification model after the training is completed; Step 6. The acceleration signal, displacement signal and loss signal of the giant magnetostrictive rod to be evaluated are extracted using the optimized multi-scale permutation entropy method to obtain a 3-channel feature map data set, and then the 3-channel feature map data set is input into the classification model to obtain a health status assessment.

2. The method for evaluating the health status of a magnetostrictive rod according to claim 1, characterized in that: In step 1, the particle swarm algorithm is an AE optimization algorithm, and the specific steps of the AE optimization algorithm are as follows: Step 3.1: Initialization, X i =lb+(ub-lb)·rand(0,1,[1,D]),i=1,2,...,N (1.51) Among them, X i represents the i-th candidate solution, lb and ub represent the lower bound and upper bound of the search space respectively, rand represents a random number uniformly distributed in the range [0,1], N represents the number of solutions; D represents the number of columns of generated random numbers; Step 3.2: Construct the alpha operator. The set of candidate solutions is defined as the candidate matrix X. The subset of the candidate matrix is ​​the evolution matrix. During the evolution process, each solution X in the current candidate matrix is ​​not directly evolved. Instead, X is sampled with replacement operation to obtain the evolution matrix to be evolved: E represents the evolution matrix, → represents the replacement sampling operation, and N represents that the operation will be performed N times; In the alpha operator, multiple steps of extracting and utilizing evolutionary information are performed simultaneously in one operator. The mathematical model is shown in the following formula: Where P is the adaptive basis vector, which determines the starting point of evolution; diagonal represents a function obtained by sampling by replacing D multiplied candidate solutions; where B is a K*D matrix obtained by sampling without replacement; A represents a D-order square matrix; represents the evolutionary solution of t+1 iterations, Δr represents the step size, θ represents the control parameter of the control differential vector, i.e., the adaptive step size, and W i and L i represents the sampled solution obtained by sampling from the candidate matrix X; P is generated by A or B itself and does not participate in evolution, which will result in no connection between P in each generation and information loss. Therefore, the concept of evolutionary path is introduced to increase the correlation between consecutive iterative steps. a and P b represents two independent evolutionary lines, then: Among them, c a Indicates P a The learning rate, c b Indicates P b The learning rate; represents the evolutionary path at time t; Step 3.3: Construct boundary constraints using the distance halving method, Among them, E i,j Indicates the boundary, ub indicates the upper bound of the search, and lb indicates the lower bound of the search; Step 3.4: Greedy selection passes the evolved solution to the next generation in a simple way without any additional operations, introducing the greedy selection strategy into the AE optimization algorithm. represents the candidate solution at time t+1, represents the boundary at time t+1, Represents the boundary value at time t+1; Step 3.5: Select the fitness function and construct the mean square error of each permutation entropy under different scale optimization parameters as the fitness function of the AE algorithm to evaluate the optimization results of the algorithm. The fitness function is as follows: Where n represents the nth fault state, m represents the total number of fault states, and H p (i) represents the permutation entropy value of the nth fault state, and Fitness represents the permutation entropy mean square error value; the permutation entropy mean square error values ​​at different scales are calculated, and the embedding dimension m of the multi-scale permutation entropy method corresponding to the maximum permutation entropy mean square error value, the scale factor s and the delay time t value, that is, the parameter combination with the best separability, is obtained, that is, the optimized multi-scale permutation entropy method.

3. The method for evaluating the health status of a magnetostrictive rod according to claim 1, characterized in that: In the step 2, the acceleration signal, the displacement signal and the loss signal are collected and obtained by the acceleration sensor, the laser displacement sensor and the BH sensor respectively.

4. The method for evaluating the health status of a magnetostrictive rod according to claim 1, characterized in that: The specific steps of step three are as follows: Step 2.1: The acceleration signal, displacement signal and loss signal are respectively formed into time series of length n. For the time series of length n, X = {x i ,i=1,2,...,n} Use the root mean square value instead of the average value to reduce the feature loss caused by the coarse-graining process: Where s is the scale factor, j = 1, 2, ... [N / s], [N / s] means rounding down, x i represents the original signal subsequence, i represents the i-th data point; represents the average value of the original signal subsequence, y (s) Represents new sequences under different coarse-graining; represents the coarse-grained sequence at scale s; N represents the number of coarse-grained groups at each scale; Step 2.2: Reconstruct the phase space of each coarse-grained sequence to obtain a new reconstructed sequence Y l(s) , AND l(s) ={and l(s) ,and l+1(s) ,…and (l+(m-1)λ)(s) } (1.48) Where m is the embedding dimension, λ is the time delay, l is the reconstruction component; y (l+(m-1)λ)(s) represents the l+(m-1)λ(s)th coarse-grained subsequence; Step 2.3: Reconstruct the sequence Y obtained in step 2.2 l(s) Arrange in ascending order, each coarse-grained sequence has a new set of sequences s(v) = (l1, l2, ... l m ), where v = 1, 2, ..., V, V ≤ m! ; then the probability of various permutations at different scales is counted; m! represents the factorial number of permutation and combination types at scale m, V represents the maximum number of coarse-grained sequences, l m Represents a subsequence sorted in ascending order; Step 2.4: Calculate the entropy of the probability of various permutations at different scales and obtain the permutation entropy value. Among them, P v is the probability of the vth symbol appearing; H p (m) represents the entropy value of scale m; Step 2.5: From step 2.4, we know that H p When (m) reaches the maximum value ln(m!), For H p (m) is normalized, H p is the final permutation entropy value of the signal, and H p The value is inversely proportional to the order of the signal, that is, the more chaotic the signal, the smaller the H p The larger the value, the greater the fault level. H p Together with the original time series signal, it forms the feature matrix Y. Among them, [H p1 ,H p1 ,H p1 ,...,H ps ] is the multi-scale permutation entropy value of the signal, x1x2...x n is the original sequence, Y is the newly constructed feature matrix, n is the number of sample points, s is the permutation entropy scale, and y is the eigenvalue of signal X after feature extraction; y ns represents the feature extraction value of the nth data under the scale s, x n represents the nth value of the original time series, H ps Represents the permutation entropy value at the sth scale; Step 4.2: Perform standardization so that the feature matrix Y can be mapped to the grayscale matrix domain. Among them, y norm is the standardized eigenvalue, X max is the maximum value in the original sequence X, X min is the minimum value in the original sequence X; Step 4.3: Set the resolution r, r of the generated image 2 ≤n×s, Step 4.4: Read the Y matrix by column and place it by column to obtain the reconstructed feature matrix C. in [·]for Round up, c is the representation after rearranging y, c kr Represents the kth row and rth column data point after reconstruction; Step 4.5: Perform grayscale value mapping and map the reconstructed feature matrix C to the grayscale value range: g i,j represents the gray value after mapping, c i,j represents the value of the i-th row and j-th column of matrix C, c max Represents the maximum value in matrix C; Where i = 1, 2, 3 ... k; j = 1, 2, 3 ... r; and the gray value feature matrix G is obtained, Step 4.6: Perform a three-dimensional stacking of the obtained acceleration signal, displacement signal and loss signal gray value feature matrices to obtain a three-channel feature map data set of the giant magnetostrictive rod. F m =[G a ;G v ;G w ] (1.64) Among them, F m The three-channel characteristic diagram of the giant magnetostrictive rod, G a Represents the gray value feature matrix of the acceleration signal, G v Represents the gray value feature matrix of the displacement signal, G w Gray value feature matrix representing the loss signal.

5. The method for evaluating the health status of a magnetostrictive rod according to claim 1, characterized in that: The specific steps of step 4 are as follows: Step 5.1: Build convolutional layers and pooling layers. Among them, X j l is the jth element of the lth layer; M j is the jth convolution region of the feature map of the l-1th layer; is the element in the i-th element of the i-1 layer; is the weight matrix corresponding to the convolution kernel; is the bias term; f(·) is the ReLU activation function; AND (i,j) =max (p,q) (X (i+p,j+q) ) (1.66) Where i and j represent the horizontal and vertical positions of the pixels in the output feature map, respectively; p and q represent the horizontal and vertical positions of the pixels in the window, respectively; Y (i,j) Represents the output matrix after maximum pooling; performs maximum pooling on the pooling layer after convolution, thereby reducing the number of parameters and the amount of calculation; Step 5.2: Initialize the digital capsule layer, Among them, v i is the instantiation vector corresponding to category j; s i is the deep feature vector; c ij is the coupling coefficient, which indicates the connection strength between the capsules in the i-th layer and the capsules in the j-th layer; M ij represents the transformation matrix between the capsule in the i-th layer and the capsule in the j-th layer; u i is the i-th output vector of the initial capsule layer, c ij The update of the intermediate variable b ij Calculated by Softmax function; Step 5.3: b ij The initial value is 0, which is determined by v i With the prediction vector Updated by the stateful routing algorithm, the dynamic routing algorithm iterates 3 times. Step 5.4: Use the margin loss function Margin as the loss function L of the CapsNet network k , L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 (1.69) Among them, T k is the classification indicator function. When category k exists, T k =1, otherwise T k =0;||v k || is the probability that the input is identified as category k; m + and m - are the upper and lower bounds respectively; λ is the proportional coefficient, and the loss function L k When the minimum value is obtained, the classification model is obtained after the training is completed.

Images