A damage diagnosis method with waveform adaptation and self-balancing

Abstract

The present invention discloses a waveform-adaptive and self-balancing damage diagnosis method, which comprises the following steps: S1: acquiring an original vibration acceleration data set of a vibration signal of an equipment; S2: performing sliding interval sampling on the original vibration acceleration data set to obtain a sampled data set; S3: dividing the sampled data set into an unbalanced part and a balanced part; S4: performing sample expansion on the unbalanced part using an improved one-dimensional generative network to obtain an expanded part; S5: performing adaptive modal decomposition on a new data set formed by the balanced part and the expanded part to obtain a decomposed waveform data set; S6: obtaining multiple time-frequency spectra based on the decomposed waveform data set using Hilbert transform; S7: splicing the multiple time-frequency spectra to generate image samples; and S8: performing classification diagnosis on the image samples.

Description

Technical Field

[0001] The present invention relates to the technical field of damage diagnosis, and in particular to a waveform adaptive and self-balancing damage diagnosis method. Background Art

[0002] With the development of the Industrial Internet, condition monitoring of mechanical equipment has entered the era of big data. This data, characterized by massive amounts of data, a wide variety of data, and rapid updates, presents new challenges for mechanical damage diagnosis. Take wind turbines, for example. As large-scale mechanical equipment, they are widely used worldwide. However, wind turbines are located in remote locations, making manual inspection and troubleshooting difficult, which poses certain challenges to equipment operation and maintenance. To ensure the safe and efficient operation of wind farms, effective damage detection and diagnosis methods are required. Therefore, data analysis and fault damage prediction for wind turbines have broad application scenarios.

[0003] In real-world applications, wind turbines rarely remain in their normal state, and damage data is often lost or incomplete when a fault occurs. This makes collecting fault samples difficult. The number of samples in the normal category far exceeds that in the other category. This imbalanced data problem is widespread in industry, and eliminating the negative impact of sample imbalance on model building is particularly important.

[0004] For the original waveform samples generated by equipment sensors, the traditional method of manual classification and screening data is inefficient and highly subjective. In addition, the status monitoring data of in-service equipment has the characteristics of large span, multiple features, and high noise. It is necessary to find a reasonable method to eliminate non-relevant features while minimizing the loss of information contained in the original indicators to achieve the purpose of comprehensive analysis of the waveform. Summary of the Invention

[0005] The object of the present invention is to provide a waveform adaptive and self-balancing damage diagnosis method to solve the problems of data sample imbalance and low fault resolution in the prior art.

[0006] The technical solution of the present invention to solve the above technical problems is as follows:

[0007] The present invention provides a waveform adaptive and self-balancing damage diagnosis method, the waveform adaptive and self-balancing damage diagnosis method comprising:

[0008] S1: Obtain the original vibration acceleration dataset of the equipment vibration signal;

[0009] S2: performing sliding interval sampling on the original vibration acceleration data set to obtain a sampled data set;

[0010] S3: dividing the sampled data set into an unbalanced part and a balanced part;

[0011] S4: using an improved one-dimensional generative network to perform sample expansion on the unbalanced part to obtain an expanded part;

[0012] S5: performing adaptive modal decomposition on a new data set formed by the balanced part and the expanded part to obtain a decomposed waveform data set;

[0013] S6: Obtain multiple time-frequency spectra using Hilbert transform according to the decomposed waveform data set;

[0014] S7: splicing the multiple time-frequency spectra to generate an image sample;

[0015] S8: Using a neural network to perform classification diagnosis on the image samples.

[0016] Optionally, step S2 includes:

[0017] Starting from the first data point of the original vibration acceleration data set, every l slide The number of data interval points is l new , obtaining several segmented data sets, wherein the several shorter data sets are the sampled data sets.

[0018] Optionally, the number Count of the sampled data sets is:

[0019]

[0020] Among them, l old is the length of the original vibration acceleration data set, l new is the length of the split data set, l slide The number of data points sampled for the sliding interval.

[0021] Optionally, in step S4, the improved one-dimensional generative network includes a generative network and a discriminant network, the generative network is used to balance the unbalanced part to obtain a one-dimensional waveform; the discriminator is used to process the one-dimensional waveform to obtain optimization parameters to optimize the generative network; the generative network includes a noise input layer, a first one-dimensional deconvolution layer, a fourth batch normalization layer, a first ReLU activation function layer, a second one-dimensional deconvolution layer, a fifth batch normalization layer, a second ReLU activation function layer, a third one-dimensional deconvolution layer, a sixth batch normalization layer, a third ReLU activation function layer, a fourth one-dimensional deconvolution layer, a hyperbolic tangent activation function layer and a one-dimensional waveform output layer connected in sequence;

[0022] The discriminant network includes a one-dimensional waveform input layer, a first one-dimensional convolution layer, a first Leaky activation function layer, a second one-dimensional convolution layer, a first batch normalization layer, a second Leaky activation function layer, a third one-dimensional convolution layer, a second batch normalization layer, a third Leaky activation function layer, a fourth one-dimensional convolution layer, a third batch normalization layer, a fourth Leaky activation function layer and a fifth one-dimensional convolution layer, which are connected in sequence. The fifth one-dimensional convolution layer is used to output the optimization parameters.

[0023] Optionally, the cost function L of the generator network G for:

[0024]

[0025] The cost function L of the discriminant network D for:

[0026]

[0027] The joint cost function of the generative network and the discriminative network is:

[0028]

[0029] Where E represents the expected function, μ represents the random noise of the input, and P μ (μ) represents the distribution of noise μ, G(μ) represents the waveform generated by the generator network for the input noise, D() represents the output result of the discriminant network, and D w Represents the discriminant network, δ represents a constant, X represents the entire waveform sample space, and P represents the distribution. Indicates the gradient of x(t), |||| p represents the p-order norm, x(t) represents the true waveform of the original unbalanced damage, P re (x(t)) represents the data distribution of the real waveform, Indicates fixed generation of network parameters.

[0030] Optionally, step S5 includes:

[0031] S51: According to the formula Construct i random Gaussian white noise sequences; where g(t) represents the data in the new data set, and its data length is N, σ (i) (t) represents the i-th random Gaussian white noise generated, R(N) represents a random sequence of length N, and snr represents the signal-to-noise ratio;

[0032] S52: According to the formula g(t) (i) =g(t)+ε0E1(σ (i)(t)), i random Gaussian white noise sequences and i original data in the new data set, to obtain i noisy data; where ε0 represents the amplitude of the added noise, E1(σ (i) (t)) represents the first-order modal component generated by the empirical mode decomposition of the i-th Gaussian white noise, g(t) (i) Represents the i-th noise-added data;

[0033] S53: According to the formula r1(t) (i) =g(t) (i) -E1(g(t) (i) ) and formula Get the average residual of i original data in the first iteration; where E1(g(t) (i) ) represents the first-order modal component generated by the empirical mode decomposition of the i-th noise-added data, r1(t) (i) represents the residual of the i-th original data in the first iteration, r1(t) represents the average of the i residuals in the first iteration, that is, the average residual;

[0034] S54: According to the average residual, use the formula IMF1=g(t)-r1(t) to obtain the first-order modal component of the first iteration;

[0035] S55: Repeat steps S52 to S54, using formula g k (t) (i) =r k-1 (t)+ε k-1 E k (σ (i) (t)) and formula IMF k =r k-1 (t)-r k (t), all modal components obtained after K iterations; where r k-1 (t) represents the average residual generated by k-1 iterations, and k = 0, 1...K, E k (σ (i) (t)) represents the kth order modal component generated by empirical mode decomposition of the i-th Gaussian white noise, ε k-1 Indicates the amplitude of the added noise, g k (t) (i) represents the i-th noisy data generated by the k-th iteration; IMF k represents the kth-order modal component of adaptive modal decomposition;

[0036] S56: Output all modal components obtained after the K iterations as the decomposed waveform data set.

[0037] Optionally, step S6 includes:

[0038] S61: IMF of each component in the decomposed waveform data set k Perform fast Fourier transform to obtain the spectrum;

[0039] S62: Take the first five frequency components in the spectrum and perform Hilbert transform to obtain the transformation result;

[0040] S63: Utilize the formula Performing time-frequency spectrum transformation on the transformation result to obtain a time-frequency spectrum.

[0041] Optionally, step S7 includes:

[0042] splicing the multiple time-frequency spectra in sequence longitudinally and normalizing them into an image of size N×N;

[0043] The N×N image is output as the image sample.

[0044] The present invention has the following beneficial effects:

[0045] The present invention can adaptively decompose waveforms and then reconstruct the decomposed waveforms to achieve noise reduction. Furthermore, the present invention uses a time-frequency fusion approach that can intuitively reflect the time domain characteristics of the original signal while also delving into the signal's frequency nature. It also converts one-dimensional samples into two-dimensional images, facilitating subsequent classification and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of the waveform adaptive and self-balancing damage diagnosis method of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the improved one-dimensional generation network of the present invention;

[0048] Figure 3 This is the loss variation diagram of the improved one-dimensional generation network of the present invention. DETAILED DESCRIPTION

[0049] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0050] The present invention provides a waveform adaptive and self-balancing damage diagnosis method, referring to Figure 1 As shown, the waveform adaptation and self-balancing damage diagnosis method includes:

[0051] S1: Obtain the original vibration acceleration dataset of the equipment vibration signal;

[0052] The test subjects of this embodiment of the present invention are ball bearings on the drive end of equipment. The bearings under diagnosis have three damage locations: inner race damage, outer race damage, and ball damage. Each damage type is classified into three equipment load levels, for a total of nine types of fault damage. Based on the above description, the sample set is divided into 10 categories: normal sample sets and nine types of damage sample sets.

[0053] S2: performing sliding interval sampling on the original vibration acceleration data set to obtain a sampled data set;

[0054] Alternatively, the present invention utilizes the formula The original vibration acceleration data set is sampled at a sliding interval, that is, starting from the first data point of the original vibration acceleration data set, every l slide The number of data interval points is l new , and obtain several segmented data sets, wherein the several shorter data sets are the sampled data sets. old is the length of the original vibration acceleration data set, l new is the length of the split data set, l slide The number of data interval points sampled for the sliding interval. slide The value should not be greater than the number of data points in a complete vibration cycle.

[0055] In order to ensure the integrity of the segmented samples and increase the number of samples, the length of the newly generated samples is l new It is set to 3 times of one rotation period of the equipment, and the sliding length l is taken slide 1 / 2 of a rotation period of the equipment;

[0056] For an imbalanced data set with an imbalance greater than 1:4, for a small number of samples, a generative network G is constructed: random noise is input and the generated extended time series waveform data is output; a judgment network D is constructed: a waveform data is input and the probability that the data is the original true waveform is output. The network structures of the generative network and the judgment network are shown in Tables 1, 2 and Figure 2 shown.

[0057] Table 1 Determine the structure of each layer of the network

[0058]

[0059]

[0060] Table 2 Generate network structure at each layer

[0061] Number of network layers Network layer parameter 1 First one-dimensional deconvolution kernel_size=114, stride=1, padding=0 2 Fourth batch normalization feature_num=512 3 First ReLU activation 4 Second 1D deconvolution kernel_size=4, stride=2, padding=1 5 Fifth batch normalization feature_num=256 6 Second ReLU activation 7 The third one-dimensional deconvolution kernel_size=4, stride=2, padding=1 8 Sixth batch normalization feature_num=128 9 The third ReLU activation 10 Fourth 1D deconvolution kernel_size=4, stride=2, padding=1 11 Hyperbolic tangent activation 12 One-dimensional waveform output kernel_size=4, stride=2, padding=1

[0062] S3: dividing the sampled data set into an unbalanced part and a balanced part;

[0063] S4: using an improved one-dimensional generative network to perform sample expansion on the unbalanced part to obtain an expanded part;

[0064] The one-dimensional generative network provided by the present invention includes a generative network and a discriminant network. The generative network is used to balance the unbalanced part to obtain a one-dimensional waveform; the discriminator is used to process the one-dimensional waveform to obtain optimization parameters to optimize the generative network.

[0065] The present invention improves the improved one-dimensional generative network of the present invention based on the correlation function of the original one-dimensional generative network, specifically as follows:

[0066] According to the formula Construct the cost function of the original generative network;

[0067] According to the formula Construct the cost function of the original discriminant network;

[0068] According to the formula Establish a joint cost function of the original generative network and the original discriminative network, and obtain a learned discriminative network;

[0069] According to the formula Optimize and improve the learned discriminant network, that is, the gradient of the discriminant network cannot exceed K, and K is 1;

[0070] According to the formula Construct the cost function of the optimized discriminant network;

[0071] Where E represents the expected function, μ represents the random noise of the input, and P μ (μ) represents the distribution of noise μ, G(μ) represents the waveform generated by the generator network for the input noise, D() represents the output result of the discriminant network, and D w Represents the discriminant network, δ represents a constant, X represents the entire waveform sample space, and P represents the distribution. Indicates the gradient of x(t), |||| p represents the p-order norm, x(t) represents the true waveform of the original unbalanced damage, P re (x(t)) represents the data distribution of the real waveform, Indicates fixed generation of network parameters.

[0072] Using the AdamW optimization algorithm, the original generative network and the optimized discriminative network are cross-trained as follows:

[0073] According to the formula Calculate the gradient of the network with respect to x(t);

[0074] Calculate the first-order momentum m based on the historical gradient t and the second-order momentum v t ;

[0075] m t =β1m t-1 +(1-β1)grad t

[0076]

[0077] Among them, β1, β2∈[0,1) are the exponential decay rates of the current estimate.

[0078] Calculate the descent gradient at the current moment;

[0079] According to the formula Update the descent gradient and repeat the above steps until θ t Convergence, where θ t is the pre-learning parameter vector, α is the model learning rate, ε is the denominator bias term, and γ is the weight attenuation factor.

[0080] By iterating L D And V(G,D) to optimize the parameters, use the generative network G to obtain the generated sample x'(t), and combine the original sample x(t) with x'(t) to form a new original data set G(t).

[0081] refer to Figure 2 As shown, the generation network includes a noise input layer, a first one-dimensional deconvolution layer, a fourth batch normalization layer, a first ReLU activation function layer, a second one-dimensional deconvolution layer, a fifth batch normalization layer, a second ReLU activation function layer, a third one-dimensional deconvolution layer, a sixth batch normalization layer, a third ReLU activation function layer, a fourth one-dimensional deconvolution layer, a hyperbolic tangent activation function layer and a one-dimensional waveform output layer, which are connected in sequence.

[0082] The discriminant network includes a one-dimensional waveform input layer, a first one-dimensional convolution layer, a first Leaky activation function layer, a second one-dimensional convolution layer, a first batch normalization layer, a second Leaky activation function layer, a third one-dimensional convolution layer, a second batch normalization layer, a third Leaky activation function layer, a fourth one-dimensional convolution layer, a third batch normalization layer, a fourth Leaky activation function layer and a fifth one-dimensional convolution layer, which are connected in sequence. The fifth one-dimensional convolution layer is used to output the optimization parameters.

[0083] S5: performing adaptive modal decomposition on a new data set formed by the balanced part and the expanded part to obtain a decomposed waveform data set;

[0084] Optionally, step S5 includes:

[0085] S51: According to the formula Construct i random Gaussian white noise sequences; where g(t) represents the data in the new data set, and its data length is N, σ (i) (t) represents the i-th random Gaussian white noise generated, R(N) represents a random sequence of length N, and snr represents the signal-to-noise ratio;

[0086] S52: According to the formula g(t) (i) =g(t)+ε0E1(σ (i) (t)), i random Gaussian white noise sequences and i original data in the new data set, to obtain i noisy data; where ε0 represents the amplitude of the added noise, E1(σ (i) (t)) represents the first-order modal component generated by the empirical mode decomposition of the i-th Gaussian white noise, g(t) (i) Represents the i-th noise-added data;

[0087] S53: According to the formula r1(t) (i) =g(t) (i) -E1(g(t) (i) ) and formula Get the average residual of i original data in the first iteration; where E1(g(t) (i) ) represents the first-order modal component generated by the empirical mode decomposition of the i-th noise-added data, r1(t) (i) represents the residual of the i-th original data in the first iteration, r1(t) represents the average of the i residuals in the first iteration, that is, the average residual;

[0088] S54: According to the average residual, use the formula IMF1=g(t)-r1(t) to obtain the first-order modal component of the first iteration;

[0089] S55: Repeat steps S52 to S54, using formula g k (t) (i) =r k-1 (t)+ε k-1 E k (σ (i) (t)) and formula IMF k =r k-1 (t)-r k (t), all modal components obtained after K iterations; where r k-1 (t) represents the average residual generated by k-1 iterations, and k = 0, 1...K, E k (σ (i) (t)) represents the kth order modal component generated by empirical mode decomposition of the i-th Gaussian white noise, εk-1 Indicates the amplitude of the added noise, g k (t) (i) represents the i-th noisy data generated by the k-th iteration; IMF k represents the kth-order modal component of adaptive modal decomposition;

[0090] S56: Output all modal components obtained after the K iterations as the decomposed waveform data set.

[0091] S6: Obtain multiple time-frequency spectra using Hilbert transform according to the decomposed waveform data set;

[0092] Optionally, step S6 includes:

[0093] S61: IMF of each component in the decomposed waveform data set k Perform fast Fourier transform to obtain the spectrum;

[0094] S62: Take the first five frequency components in the spectrum and perform Hilbert transform to obtain the transformation result;

[0095] S63: Utilize the formula Performing time-frequency spectrum transformation on the transformation result to obtain a time-frequency spectrum.

[0096] S7: splicing the multiple time-frequency spectra to generate an image sample;

[0097] Optionally, step S7 includes:

[0098] splicing the multiple time-frequency spectra in sequence longitudinally and normalizing them into an image of size N×N;

[0099] The N×N image is output as the image sample.

[0100] S8: Perform classification diagnosis on the image samples.

[0101] The present invention utilizes a neural network to perform classification diagnosis on image samples. The neural network can be any neural network as long as it has a classification function, and the present invention does not impose any specific limitation.

[0102] The present invention uses an improved one-dimensional generative network to perform data self-balancing and augmentation based on the original waveform data points. For the unbalanced data sets commonly seen in the industrial field, the traditional data undersampling or oversampling methods can easily cause the loss of important data information, especially the sample's spatiotemporal information, while reducing the majority class sample data. It also leads to the generation of too much redundant invalid data while increasing the minority class samples. With the development of machine learning, there are methods to improve the loss function to adapt to unbalanced data sets, but when the number of minority class samples is too small or even there are no minority class samples, this method is no longer applicable. At the same time, the classification method based on the loss function does not fundamentally solve the problem of sample imbalance. Based on this, the present invention uses a generative adversarial network to generate data samples that conform to the original rules, thereby improving the accuracy of damage diagnosis under unbalanced conditions, such as Figure 3 As shown, the generative adversarial network constructed by the present invention can be stable and converge within a certain period.

[0103] The present invention can adaptively decompose the waveform and then reconstruct the decomposed waveform to achieve the purpose of noise reduction. The present invention adopts a time-frequency fusion method, which can intuitively reflect the time domain characteristics of the original signal and deeply understand the frequency nature of the signal. At the same time, it converts the one-dimensional sample into a two-dimensional image to facilitate subsequent classification and diagnosis.

[0104] 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, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A waveform adaptive and self-balancing damage diagnosis method, characterized in that: The waveform self-adaptation and self-balancing damage diagnosis method includes: S1: Obtain the original vibration acceleration dataset of the equipment vibration signal; S2: performing sliding interval sampling on the original vibration acceleration data set to obtain a sampled data set; S3: dividing the sampled data set into an unbalanced part and a balanced part; S4: using an improved one-dimensional generative network to perform sample expansion on the unbalanced part to obtain an expanded part; In step S4, the improved one-dimensional generative network includes a generative network and a discriminant network, the generative network is used to balance the unbalanced part to obtain a one-dimensional waveform; the discriminator is used to process the one-dimensional waveform to obtain optimization parameters to optimize the generative network; the generative network includes a noise input layer, a first one-dimensional deconvolution layer, a fourth batch normalization layer, a first ReLU activation function layer, a second one-dimensional deconvolution layer, a fifth batch normalization layer, a second ReLU activation function layer, a third one-dimensional deconvolution layer, a sixth batch normalization layer, a third ReLU activation function layer, a fourth one-dimensional deconvolution layer, a hyperbolic tangent activation function layer and a one-dimensional waveform output layer connected in sequence; The discriminant network includes a one-dimensional waveform input layer, a first one-dimensional convolutional layer, a first Leaky activation function layer, a second one-dimensional convolutional layer, a first batch normalization layer, a second Leaky activation function layer, a third one-dimensional convolutional layer, a second batch normalization layer, a third Leaky activation function layer, a fourth one-dimensional convolutional layer, a third batch normalization layer, a fourth Leaky activation function layer and a fifth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer being used to output the optimization parameters; S5: performing adaptive modal decomposition on a new data set formed by the balanced part and the expanded part to obtain a decomposed waveform data set; S6: Obtain multiple time-frequency spectra using Hilbert transform according to the decomposed waveform data set; S7: splicing the multiple time-frequency spectra to generate an image sample; S8: Perform classification diagnosis on the image samples.

2. The waveform adaptive and self-balancing damage diagnosis method according to claim 1, characterized in that: The step S2 comprises: Starting from the first data point of the original vibration acceleration data set, every The number of data interval points is taken as the length , obtaining several segmented data sets, wherein the several shorter data sets are the sampled data sets.

3. The waveform adaptive and self-balancing damage diagnosis method according to claim 2, characterized in that: The number of sampled data sets for: in, is the length of the original vibration acceleration data set, is the length of the split data set, The number of data points sampled for the sliding interval.

4. The waveform adaptive and self-balancing damage diagnosis method according to claim 1, characterized in that: The cost function of the generative network for: The cost function of the discriminant network for: The joint cost function of the generative network and the discriminative network is: in, represents the expected function, represents the random noise of the input, Represents noise The distribution of represents the waveform generated by the generator network for input noise, Represents the output of the discriminant network, represents the discriminant network, represents a constant, represents the entire waveform sample space, Indicates the distribution, Express Find the gradient, express p Order norm, The true waveform representing the original unbalanced damage, represents the data distribution of the real waveform, Indicates fixed generation of network parameters.

5. The waveform adaptive and self-balancing damage diagnosis method according to claim 1, characterized in that: The step S5 comprises: S51: According to the formula structure A random Gaussian white noise sequence; where Represents the data in the new data set, and its data length is N , Indicates the generated i A random Gaussian white noise, Indicates the length is N A random sequence of snr represents the signal-to-noise ratio; S52: According to the formula 、 A random Gaussian white noise sequence and the new data set The original data is obtained Noised data; among them, represents the amplitude of the added noise, Indicates the i The first-order modal component generated by empirical mode decomposition of Gaussian white noise is Indicates the i Noised data; S53: According to the formula and formula , and get the first iteration The average residual of the original data; Indicates the i The first-order modal component generated by empirical mode decomposition of the noisy data is Indicates the first iteration i The residuals of the original data, In the first iteration i The average of the residuals, that is, the mean residual; S54: Based on the average residual, use the formula , get the first-order modal component of the first iteration; S55: Repeat steps S52 to S54, using the formula and formula ,get K All modal components obtained after iterations; among them, express k -The average residual generated by 1 iteration, and k =0,1… K , Indicates the i The first one generated by empirical mode decomposition of Gaussian white noise k The modal component, represents the amplitude of the added noise, Indicates the k The first iteration generated i Noisy data; represents the adaptive modal decomposition k The first modal component; S56: The K All modal components obtained after iterations are output as the decomposed waveform data set.

6. The waveform adaptive and self-balancing damage diagnosis method according to claim 5, characterized in that: The step S6 comprises: S61: Concentrate the components of the decomposed waveform data set Perform fast Fourier transform to obtain the spectrum; S62: Take the first five frequency components in the spectrum and perform Hilbert transform to obtain the transformation result; S63: Utilize the formula Performing time-frequency spectrum transformation on the transformation result to obtain a time-frequency spectrum.

7. The waveform adaptive and self-balancing damage diagnosis method according to any one of claims 1 to 6, characterized in that: The step S7 comprises: Multiple time-frequency spectra are spliced ​​vertically in sequence and standardized as Size of the image; The An image of size φ is output as the image sample.

Images