Classification method of aero-engine rotor vibration data based on network structure search

Abstract

This invention provides a classification method for aero-engine rotor vibration data based on network structure search, relating to the field of aero-engine technology. The method first collects vibration data of aero-engine rotor systems under various operating conditions and converts it into grayscale images. Then, it uses a convolutional neural network (CNN) with empirically selected parameters and a CNN based on network structure search to perform multi-classification of the vibration data under all operating conditions, comparing the classification results. Next, it uses visualization dimensionality reduction technology to analyze the separability of the original vibration data. Then, it uses the aforementioned two CNNs to classify various types of vibration data with differences greater than a set threshold, comparing the classification results. Finally, based on the classified vibration data, it adds several types of vibration data with differences less than or equal to a set threshold from the original vibration data, classifying them using the aforementioned two CNNs, and comparing the classification results.

Description

Technical Field

[0001] This invention relates to the field of aero-engine technology, and in particular to a classification method for aero-engine rotor vibration data based on network structure search. Background Technology

[0002] With the rapid development of my country's aviation industry, technologies related to engine fault detection and signal diagnosis have attracted increasing attention from researchers. As aero-engines age, they may experience faults due to various reasons, and different operating conditions during normal operation will result in different states. In recent years, how to analyze collected data sequences using vibration data and methods such as deep learning, and then classify the data sequences using the analysis results, has become a major research focus. Since the rotor of an aero-engine exhibits different vibration states under different operating conditions, classifying these vibration states using this vibration data can provide prior knowledge for further fault diagnosis, which has practical significance.

[0003] One common approach is to process the data (e.g., Fourier transform, wavelet transform, time series transform), convert it into images, and then train it using a convolutional neural network. However, the structural parameters of convolutional neural networks are often manually selected, relying on experience. The hyperparameters used in the network structure are not necessarily optimal for the current dataset, and it is not easy to quickly select suitable network parameters when the dataset changes. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a classification method for vibration data of aero-engine rotor system based on network structure search, which addresses the shortcomings of the prior art and realizes the classification of vibration data of aero-engine rotor system.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a classification method for aero-engine rotor vibration data based on network structure search, comprising the following steps:

[0006] Step 1: Collect vibration data of the aero-engine rotor system under I different operating conditions;

[0007] Step 2: Preprocess the collected vibration data of the aero-engine rotor system under different operating conditions and convert the vibration data into grayscale images;

[0008] Step 2.1: Preprocess the vibration data collected under different working conditions;

[0009] The collected vibration data under different working conditions are time series data. The data is set to be collected at the [number]th [period]. The length under this working condition is The original vibration data sequence is , , The vibration data is standardized as shown in the following formula:

[0010] ;

[0011] in, Indicates the first Data from the original vibration data sequence under various working conditions, Represents the standardized version of the first... Data from vibration data sequences under various working conditions Indicates the first The mean of all data within the original vibration data sequence under a certain working condition. Indicates the first The standard deviation of all data within the original vibration data sequence under various working conditions;

[0012] Then the standardized first The vibration data sequence under the following working conditions is represented as follows: ;

[0013] Step 2.2: Convert the standardized vibration data into a grayscale image;

[0014] First, the vibration datasets under all operating conditions are truncated to a length of [missing value]. This data facilitates subsequent signal processing transformations and grayscale image conversions;

[0015] The grayscale image to be converted from the vibration data under each working condition is defined as high-resolution. Width If the size is specified, then each grayscale image has one and only one unique size. OK Column matrix Correspondingly, each vibration data sequence after standardization can be converted into at most [number of data points]. The images include... This indicates rounding down; the specific conversion method is as follows:

[0016] First, the vibration data sequences under each working condition after standardization are... Divided into training set and verification set Each contains several groups of lengths. Vibration data sequence , and This sequence is a subsequence of the original vibration data sequence; first, a Fast Fourier Transform (FFT) is performed on this subsequence, and then a grayscale image is generated based on the transformed result;

[0017] (1) For a set of vibration data sequences ,Pick Perform a Fast Fourier Transform (FFT) on each point to obtain the spectrum, where... Since the sequence obtained by FFT is a symmetric complex sequence, after the original data is processed by FFT to obtain the spectrum, the... The length is sufficient; take the modulus value of each element in the frequency domain sequence and then divide by . Then, it is scaled down to the interval [0, 255] to obtain the frequency domain sequence. ,in, For the frequency domain sequence, the first... One spectrum data;

[0018] (2) In order to embed the frequency domain sequence into the image, it is necessary to generate a set of time-series indices corresponding to the frequency domain sequence: The elements within it are calculated as follows: And it satisfies: And make ;

[0019] in, Indicates to Round to the nearest integer. This represents the scaling factor for the index, and the formula will scale each time-series index value. All reduced to 1 to Within the range;

[0020] (3) Based on the time index and the frequency domain sequence, rearrange the frequency domain sequence to obtain the rearranged frequency domain sequence. The sorting rules are as follows:

[0021] ;

[0022] in, Indicates the rearrangement of the frequency domain sequence at the th The element value at the index, Indicates the original frequency domain sequence at the th The value of the element at the index;

[0023] The generated rearranged frequency domain sequence is transformed into OK Column matrix This will generate a grayscale image.

[0024] Step 3: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to perform I classification on all vibration data under I working conditions, and compare the classification results;

[0025] Step 3.1: Construct a convolutional neural network with empirically selected parameters to perform I-type classification on vibration data under all I working conditions;

[0026] The convolutional neural network with parameters selected based on experience includes six convolutional layers (conv), a normalization layer (BN), a ReLU activation function layer, a pooling layer (MaxPool), and a fully connected layer (Fully Connection). Fallout means flattening the matrix into a vector, and dropout means randomly discarding data to prevent overfitting.

[0027] The structure of the convolutional neural network based on empirically selected parameters is defined as follows:

[0028] conv1: 1-8 @3×3, padding = 1, stride = 1;

[0029] BN + ReLU;

[0030] MaxPool: @2×2, stride = 2;

[0031] conv2: 8-16 @3×3, padding = 1, stride = 1;

[0032] BN + ReLU;

[0033] MaxPool: @2×2, stride = 2;

[0034] conv3_1: 16-32 @3×3, padding = 1, stride = 1;

[0035] BN + ReLU;

[0036] conv3_2: 32-32 @3×3, padding = 1, stride = 1;

[0037] BN + ReLU;

[0038] MaxPool: @2×2, stride = 2;

[0039] conv4_1: 32-64 @3×3, padding = 1, stride = 1;

[0040] BN + ReLU;

[0041] conv4_2: 64-64 @3×3, padding = 1, stride = 1;

[0042] BN + ReLU;

[0043] MaxPool: @2×2, stride = 2;

[0044] Flatten;

[0045] Fully Connection1 + ReLu + dropout;

[0046] Fully Connection2 + ReLU;

[0047] softMax;

[0048] Where conv represents a convolutional layer, BN represents a normalization layer, ReLU represents a ReLU activation function layer, MaxPool represents a pooling layer, Fully Connection represents a fully connected layer, Fallout represents flattening the matrix into a vector, and dropout represents randomly dropping out to prevent overfitting.

[0049] Step 3.2: Construct a convolutional neural network based on network structure search to perform I-classification on vibration data under all I working conditions;

[0050] First, the basic architecture of the convolutional neural network is given, including the definition of the number of convolutional layers, the structure of pooling layers, and the structure and number of fully connected layers. The number of input and output channels of each convolutional layer, the kernel size, and the structure-related hyperparameters of the fully connected layers are obtained through search.

[0051] Then, the search space for hyperparameters is given, and the hyperparameters are restricted to integers within the search space; through random search, a set of hyperparameters will be obtained within the specified search space;

[0052] The convolutional neural network structure is updated using hyperparameters, and the network is trained for a specified number of epochs to obtain the classification accuracy of the network for the dataset under that structure. After training reaches a given number of epochs, new parameters are obtained through random search, and training continues. The accuracy of each search is recorded. After training is completed, the classification accuracy under different network structures can be obtained. The network structure parameters corresponding to the highest accuracy are taken as the desired result.

[0053] The convolutional neural network for network architecture search includes five convolutional layers (conv), a ReLU activation function layer, a maxPool pooling layer, and a fully connected layer. Fallout flattens the matrix into a vector, and dropout randomly discards data to prevent overfitting. The fully connected layer configuration of the convolutional neural network based on network architecture search used in this invention is the same as that of a regular convolutional neural network. The definitions of each convolutional layer are as follows:

[0054] conv1: Input channel 1, output channel search space: [1, 100), kernel scale search space: [1, 10);

[0055] ReLU activation function;

[0056] maxPool: Pooling kernel size: 2×2;

[0057] conv2: The number of input channels is equal to the number of output channels of conv1. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0058] ReLU activation function;

[0059] maxPool: Pooling kernel size: 2×2;

[0060] conv3: The number of input channels is equal to the number of output channels of conv2. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0061] ReLU activation function;

[0062] maxPool: Pooling kernel size: 2×2;

[0063] conv4: The number of input channels is equal to the number of output channels of conv3. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0064] ReLU activation function;

[0065] maxPool: Pooling kernel size: 2×2;

[0066] conv5: The number of input channels is equal to the number of output channels of conv4. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0067] ReLU activation function;

[0068] maxPool: Pooling kernel size: 2×2;

[0069] Step 4: Use visualization dimensionality reduction techniques to analyze the separability of the original vibration data;

[0070] Step 4.1: Use t-distributed random neighbor embedding (t-SNE) to visualize and reduce the dimensionality of the vibration data;

[0071] For each type of image generated from vibration data under different working conditions, the corresponding matrix mentioned in step 2 is applied. Straighten into a vector Using principal component analysis to transform vectors Reduce the dimensionality to 50 and then draw a visual t-SNE diagram;

[0072] Step 4.2: Based on the visualized t-SNE plot, analyze the vibration data where the difference is greater than a set threshold; the difference refers to the Euclidean distance between two data vectors.

[0073] ;

[0074] in, Indicates the first The vector and the first Euclidean distance between vectors Indicates the first The first vector One component;

[0075] Step 5: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to classify various types of vibration data (Data00—Data09+Data13+Data15) with differences greater than a set threshold, and compare the classification results;

[0076] Step 6: Based on the vibration data used in Step 5, add several types of vibration data (Data12, Data14) whose difference from the original vibration data is less than or equal to a set threshold. Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to classify them respectively, and compare the classification results.

[0077] The beneficial effects of adopting the above technical solution are as follows: The classification method for aero-engine rotor vibration data based on network structure search provided by this invention first collects vibration data of the aero-engine rotor system under different operating conditions, then standardizes the vibration data, and then performs a fast Fourier transform to generate a grayscale image. Next, a convolutional neural network based on network structure search is built to search for the most suitable parameters, and the model is trained simultaneously. The convolutional neural network based on network structure search of this invention can automatically select the parameters of each layer of the convolutional neural network. It does not rely on human experience to select parameters, but automatically adjusts the parameters of each layer of the convolutional neural network according to the characteristics of the data itself, so that the training results achieve sufficiently high classification accuracy. Compared with convolutional neural networks based on experience-based parameter selection, even when the data differences are small, the classification accuracy of the convolutional neural network based on network structure search technology is still higher than that of the convolutional neural network based on experience-based parameter selection.

[0078] The classification method for vibration data of aero-engine rotor systems based on network structure search provided by this invention has high classification accuracy, and the network structure does not need to be manually specified based on experience. Compared with traditional convolutional neural networks that select parameters based on experience, it has better classification performance. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the classification process of the convolutional neural network based on network structure search used in this invention;

[0080] Figure 2 A schematic diagram illustrating the generation of a grayscale image from a vibration data sequence provided in an embodiment of the present invention;

[0081] Figure 3 The diagrams provided in this embodiment of the invention are a convolutional neural network structure based on empirical parameter selection and a convolutional neural network structure based on network structure search. (a) is a convolutional neural network structure based on empirical parameter selection, and (b) is a convolutional neural network structure based on network structure search.

[0082] Figure 4 The accuracy obtained on the validation set by using a convolutional neural network with empirically selected parameters to perform 17 classifications on all vibration data, as provided in this embodiment of the invention;

[0083] Figure 5 The embodiments of this invention provide the top ten results of the accuracy of each experiment obtained by using a convolutional neural network based on network structure search technology to perform 17 classifications on all data (the horizontal axis in the figure is the experiment number, and the vertical axis is the classification accuracy obtained in each experiment).

[0084] Figure 6 A schematic diagram of t-SNE visualization of the data differences of 17 types of vibration data to be classified, provided in an embodiment of the present invention;

[0085] Figure 7 This is a t-SNE visualization diagram illustrating the differences between Data10 and Data16 provided in an embodiment of the present invention.

[0086] Figure 8 This is a t-SNE visualization diagram illustrating the data differences between normal conditions (Data09) and different load currents (Data10~Data13) at 5000 rpm, provided in an embodiment of the present invention. (a) shows the difference between normal conditions (Data09) and Data10 (load current 0.1A), (b) shows the difference between normal conditions (Data09) and Data11 (load current 0.2A), (c) shows the difference between normal conditions (Data09) and Data12 (load current 0.3A), and (d) shows the difference between normal conditions (Data09) and Data13 (load current 0.4A).

[0087] Figure 9 This is a t-SNE visualization diagram illustrating the differences between normal conditions (Data09) and UB1 (Data14), UB2 (Data15), and UB3 (Data16) at 5000 rpm, provided in an embodiment of the present invention. (a) shows the differences between normal conditions (Data09) and UB1 (Data14), (b) shows the differences between normal conditions (Data09) and UB2 (Data15), and (c) shows the differences between normal conditions (Data09) and UB3 (Data16).

[0088] Figure 10 A t-SNE visualization diagram illustrating the data differences between load4 (Data13) and UB2 (Data15) at 5000 rpm, provided in an embodiment of the present invention;

[0089] Figure 11 The accuracy obtained on the validation set by using a convolutional neural network with empirically selected parameters to perform 12 classifications on the first 10 working conditions and load4 (Data13) and UB2 (Data15) in an embodiment of the present invention;

[0090] Figure 12 The top ten results of the accuracy of each experiment obtained by using a convolutional neural network based on network structure search to perform 12 classification on the first 10 working conditions and load4 (Data13) and UB2 (Data15) in the embodiment of the present invention (the horizontal axis of the figure is the experiment number, and the vertical axis is the classification accuracy obtained in each experiment).

[0091] Figure 13The accuracy graph of using a convolutional neural network with empirically selected parameters to perform 14 classifications on the validation set for the first 10 working conditions and load3 (Data12), load4 (Data13), UB1 (Data14), and UB2 (Data15) provided in the embodiments of the present invention;

[0092] Figure 14 The top ten results of the accuracy of each experiment obtained by using a convolutional neural network based on network structure search to perform 14 classification on the first 10 working conditions and load3 (Data12), load4 (Data13), UB1 (Data14), and UB2 (Data15) in the embodiment of the present invention (the horizontal axis of the figure is the experiment number, and the vertical axis is the classification accuracy obtained in each experiment). Detailed Implementation

[0093] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0094] In this embodiment, the classification method for aero-engine rotor vibration data based on network structure search includes the following steps:

[0095] Step 1: Collect vibration data of the aero-engine rotor system under 17 different operating conditions. The sampling frequency is 50kHz, and the running time for each operating condition is 1 minute. In this embodiment, the collected vibration data of the aero-engine rotor system includes: vibration data of the rotor system operating at 0 rpm under normal conditions (numbered Data00), 200 rpm under normal conditions (numbered Data01), 400 rpm under normal conditions (numbered Data02), 600 rpm under normal conditions (numbered Data03), 800 rpm under normal conditions (numbered Data04), 1000 rpm under normal conditions (numbered Data05), 2000 rpm under normal conditions (numbered Data06), and 3000 rpm under normal conditions (numbered Data07). Vibration data under normal operating conditions at 4000 rpm (Data08), vibration data under normal operating conditions at 5000 rpm (Data09), vibration data under load at 5000 rpm (load current 0.1A) (Data10), vibration data under load at 5000 rpm (load current 0.2A) (Data11), vibration data under load at 5000 rpm (load current 0.3A) (Data12), vibration data under load at 5000 rpm (load current 0.4A) (Data13), unbalanced operating state 1 at 5000 rpm (Data14), unbalanced operating state 2 at 5000 rpm (Data15), and unbalanced operating state 3 at 5000 rpm (Data16).

[0096] Step 2: Preprocess the collected vibration data of the aero-engine rotor system under different operating conditions, and convert the vibration data into grayscale images, such as... Figure 2 As shown;

[0097] Step 2.1: Preprocess the vibration data collected under different working conditions;

[0098] The collected vibration data under different working conditions are time series data. The data is set to be collected at the [number]th [period]. The length under this working condition is The original vibration data sequence is , , The vibration data is standardized as shown in the following formula:

[0099] ;

[0100] in, Indicates the first Data from the original vibration data sequence under various working conditions, Represents the standardized version of the first... Data from vibration data sequences under various working conditions Indicates the first The mean of all data within the original vibration data sequence under a certain working condition. Indicates the first The standard deviation of all data within the original vibration data sequence under various working conditions;

[0101] Then the standardized first The vibration data sequence under the following working conditions is represented as follows: ;

[0102] Step 2.2: Convert the standardized vibration data into a grayscale image;

[0103] First, the vibration datasets under all operating conditions are truncated to a length of [missing value]. This data facilitates subsequent signal processing transformations and grayscale image conversions;

[0104] The grayscale image to be converted from the vibration data under each working condition is defined as high-resolution. Width If the size is specified, then each grayscale image has one and only one unique size. OK Column matrix Correspondingly, each vibration data sequence after standardization can be converted into at most [number of data points]. The images include... This indicates rounding down; the specific conversion method is as follows:

[0105] First, the vibration data sequences under each working condition after standardization are... Divided into training set and verification set Each contains several groups of lengths. Vibration data sequence , and Clearly, this is a subsequence of the original vibration data sequence. First, a Fast Fourier Transform (FFT) is performed on this subsequence, and then a grayscale image is generated using the transformed result. Compared to the Discrete Fourier Transform (DFT), the FFT uses a divide-and-conquer approach, which greatly simplifies the computation. The specific operation is as follows:

[0106] (1) For a set of vibration data sequences ,Pick Perform a Fast Fourier Transform (FFT) on each point to obtain the spectrum, where... Since the sequence obtained by FFT is a symmetric complex sequence, after the original data is processed by FFT to obtain the spectrum, the... The length is sufficient; take the modulus value of each element in the frequency domain sequence and then divide by . Then, it is scaled down to the interval [0, 255] to obtain the frequency domain sequence. ,in, For the frequency domain sequence, the first... One spectrum data;

[0107] (2) In order to embed the frequency domain sequence into the image, it is necessary to generate a set of time-series indices corresponding to the frequency domain sequence: The elements within it are calculated as follows: And it satisfies: And make ;

[0108] in, Indicates to Round to the nearest integer. This represents the scaling factor for the index, and the formula will scale each time-series index value. All reduced to 1 to Within the range; correspondingly take the first one. Each element constitutes a time-series index set. , Indicates the time series at the 1st The element value at the index; the reason for setting it appropriately. Make This is to prevent index overlap, which could cause an element in the image to be overwritten, resulting in information loss.

[0109] (3) Based on the time index and the frequency domain sequence, rearrange the frequency domain sequence to obtain the rearranged frequency domain sequence. The sorting rules are as follows:

[0110] ;

[0111] in, Indicates the rearrangement of the frequency domain sequence at the th The element value at the index, Indicates the original frequency domain sequence at the th The element value at the index; the meaning of this formula is: rearrange the original frequency domain sequence, and take the first element as the index. The element values ​​are placed into the rearranged sequence. At the time index, it is clear that since the arrangement only occurs at the time index, the length of the rearranged sequence is greater than the original sequence and equal to the image width multiplied by the height.

[0112] It can be seen that the rearranged sequence generated in this way has the following properties:

[0113] (4) Discreteness: It does not include 1~ All integers, therefore the rearranged sequence is only possible if... The positions with values ​​are rearranged, and the positions without values ​​are filled with 1. Therefore, the generated image will show a situation where holes and non-holes alternate.

[0114] (5) Order: The rearranged sequence is rearranged according to the index of the time feature sequence, so the rearranged sequence is ordered.

[0115] (6) Non-negativity: Since the original frequency feature sequence is scaled to the interval [0, 255], and the element values ​​of the original sequence are not changed in the permutation, this is a non-negative sequence.

[0116] The generated rearranged frequency domain sequence is transformed into OK Column matrix This will generate a grayscale image.

[0117] Step 3: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to perform 17 classifications on the vibration data under all 17 working conditions, and compare the classification results;

[0118] By combining network structure search and convolutional neural network (CNN) techniques, a CNN based on network structure search is constructed and compared with a CNN method that uses empirically selected parameters for classification. The classification performance of the two methods is then compared using a test set.

[0119] Step 3.1: Construct a convolutional neural network with empirically selected parameters to perform I-type classification on vibration data under all I working conditions;

[0120] The convolutional neural network with parameters selected based on experience includes six convolutional layers (conv), a normalization layer (BN), a ReLU activation function layer, a pooling layer (MaxPool), and a fully connected layer (Fully Connection). Fallout means flattening the matrix into a vector, and dropout means randomly discarding data to prevent overfitting.

[0121] The structure of the convolutional neural network based on empirical parameter selection used in this invention is defined as follows:

[0122] conv1: 1-8 @3×3, padding = 1, stride = 1;

[0123] BN + ReLU;

[0124] MaxPool:@2×2,stride = 2;

[0125] conv2:8-16 @3×3,padding = 1, stride = 1

[0126] BN + ReLu

[0127] MaxPool:@2×2,stride = 2

[0128] conv3_1:16-32 @3×3,padding = 1,stride = 1;

[0129] BN + ReLu;

[0130] conv3_2:32-32 @3×3,padding = 1,stride = 1;

[0131] BN + ReLu;

[0132] MaxPool:@2×2,stride = 2;

[0133] conv4_1:32-64 @3×3,padding = 1, stride = 1;

[0134] BN + ReLu;

[0135] conv4_2:64-64 @3×3,padding = 1, stride = 1;

[0136] BN + ReLu;

[0137] MaxPool:@2×2,stride = 2;

[0138] Flatten;

[0139] Fully Connection1 + ReLu + dropout;

[0140] Fully Connection2 + ReLu;

[0141] softMax;

[0142] Where conv represents a convolutional layer, BN represents a normalized layer, ReLU represents a ReLU activation function layer, MaxPool represents a pooling layer, Fully Connection represents a fully connected layer, Fallout represents flattening the matrix into a vector, and dropout represents randomly dropping values ​​to prevent overfitting.

[0143] The specific operation of the convolutional layer is as follows: A two-dimensional convolution kernel is used to perform a convolution operation on the input image; that is, the kernel multiplies the corresponding positions of the image input to this layer and then adds them together, with the result serving as the element at the center of the kernel. Since this operation changes the size of the output image, a padding value is set to indicate edge padding during convolution. It should be noted that a grayscale image should be generated, not an RGB color image. Experiments show that using a full-color RGB image does not necessarily produce better results than a grayscale image, because the rearrangement of a full-color image disrupts the original temporal characteristics.

[0144] The activation function layer uses the ReLU function, whose input-output characteristics are as follows:

[0145] ;

[0146] The specific operation of the pooling layer is as follows: within the region (2×2) scanned by the pooling kernel, the maximum value within that region is used to replace the previous value.

[0147] The original value of the region. Each region is iterated through sequentially to obtain the output. The pooling layer stride is 2, and no padding is performed, so the number of data channels is not changed, but the data dimension is changed.

[0148] The dropout layer works by randomly discarding a portion of feature detections from each layer, thus preventing overfitting. This is particularly effective when the sample size is small and the model is large.

[0149] Step 3.2: Construct a convolutional neural network based on network structure search to classify the vibration data under all 17 working conditions into 17 categories, such as... Figure 1 As shown;

[0150] First, the basic architecture of the convolutional neural network is given, including the definition of the number of convolutional layers, the structure of pooling layers, and the structure and number of fully connected layers. The number of input and output channels of each convolutional layer, the kernel size, and the structure-related hyperparameters of the fully connected layers are obtained through search.

[0151] Then, the search space for hyperparameters is given. Since the hyperparameters in this invention are the number of channels and the kernel size, the hyperparameters should also be restricted to integers within the search space. Through random search, a set of hyperparameters will be obtained within the specified search space.

[0152] The convolutional neural network structure is updated using hyperparameters, and the network is trained for a specified number of epochs to obtain the classification accuracy of the network for the dataset under that structure. After training reaches a given number of epochs, new parameters are obtained through random search, and training continues. The accuracy of each search is recorded. After training is completed, the classification accuracy under different network structures can be obtained. The network structure parameters corresponding to the highest accuracy are taken as the desired result.

[0153] The convolutional neural network based on network architecture search includes five convolutional layers (conv), a ReLU activation function layer, a pooling layer (MaxPool), and a fully connected layer. Fallout flattens the matrix into a vector, and dropout randomly discards data to prevent overfitting. The fully connected layer settings of the convolutional neural network based on network architecture search used in this invention are the same as those of a regular convolutional neural network. The definitions of each convolutional layer are as follows: conv1: input channel 1, output channel search space: [1, 100), kernel scale search space: [1, 10);

[0154] ReLU activation function;

[0155] maxPool: Pooling kernel size: 2×2;

[0156] conv2: The number of input channels is equal to the number of output channels of conv1. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0157] ReLU activation function;

[0158] maxPool: Pooling kernel size: 2×2;

[0159] conv3: The number of input channels is equal to the number of output channels of conv2. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0160] ReLU activation function;

[0161] maxPool: Pooling kernel size: 2×2;

[0162] conv4: The number of input channels is equal to the number of output channels of conv3. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0163] ReLU activation function;

[0164] maxPool: Pooling kernel size: 2×2;

[0165] conv5: The number of input channels is equal to the number of output channels of conv4. The output channel search space is [1, 100), and the kernel scale search space is [1, 10).

[0166] ReLU activation function;

[0167] maxPool: Pooling kernel size: 2×2;

[0168] It should be noted that when the kernel size is 1, it is equivalent to only changing the value at the corresponding position, and does not achieve the effect of a true convolutional layer. Therefore, if the kernel size is 1 and the number of output channels remains unchanged, it means that the convolutional layer can be discarded, and the network can be changed by the searched parameters.

[0169] Step 4: Use visualization dimensionality reduction techniques to analyze the separability of the original vibration data;

[0170] Step 4.1: Use t-Distributed Stochastic Neighbor Embedding (t-SNE) to visualize and reduce the dimensionality of the vibration data;

[0171] For each type of image generated from vibration data under different working conditions, the corresponding matrix mentioned in step 2 is applied. Straighten into a vector Using principal component analysis to transform vectors Reduce the dimensionality to 50 and then draw a visual t-SNE diagram;

[0172] Step 4.2: Based on the visualized t-SNE plot, analyze the vibration data where the difference is greater than a set threshold; the difference refers to the Euclidean distance between two data vectors.

[0173] ;

[0174] in, Indicates the first The vector and the first Euclidean distance between vectors Indicates the first The first vector Each component.

[0175] Step 5: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to perform 12 classifications on various vibration data (Data00—Data09+Data13+Data15) with differences greater than a set threshold, and compare the classification results;

[0176] Step 6: Based on the vibration data used in Step 5, add several types of vibration data (Data12, Data14) whose difference from the original vibration data is less than or equal to a set threshold. Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to classify them respectively, and compare the classification results.

[0177] In this embodiment, in step 2, the length of each group of data extracted is 20000 (i.e., ), If the value is 100000, the generated image size will be 250×200.

[0178] In step 3, the vibration datasets under 17 working conditions are imported into a convolutional neural network based on empirically selected parameters and a convolutional neural network based on network structure search for training, respectively. The network structures are as follows: Figure 3 As shown. Using this vibration data, 17 classifications were performed, and the results were as follows: In the convolutional neural network classification based on empirically selected parameters, the accuracy on the test set reached 61%, and using a network search strategy, it reached over 82%, as shown. Figure 4 and Figure 5 As shown, the accuracy achieved by network structure search techniques is significantly higher than that of convolutional neural networks based on empirical parameter selection. However, an accuracy rate exceeding 82% still has certain limitations in practical application, thus necessitating further research on the data.

[0179] In step 4, a visualization of the differences is generated for the 17 types of operating data, as shown below. Figure 6 As shown, the first 10 data categories (Data00-Data09) are completely classifiable, exhibiting significant differences, while the latter 7 data categories (Data10-Data16) are almost completely mixed together. If these 7 data categories are studied separately, as shown... Figure 7 As shown, these seven categories of data are mixed together, making it very difficult to classify them one by one.

[0180] In step 4, the differences between the data under normal conditions at 5000 rpm (Data09) and different load currents are studied, such as... Figure 8 As shown, it can be seen that at 5000rpm, the difference between normal and loaded speeds increases with the increase of current. However, the difference between 5000rpm_load1 (Data10) and 5000rpm_load4 (Data13) is not significant.

[0181] The study examines the differences between the data (Data09) under normal conditions at 5000 rpm and UB1 (Data14), UB2 (Data15), and UB3 (Data16), such as... Figure 9As shown, it can be seen that the difference between UB2 and UB1 is greatest at 5000rpm, while the difference between UB2 and UB1 is smallest.

[0182] The study examines the differences between 5000rpm_load4 (Data13) and UB2 (Data15), such as... Figure 10 As shown, the differences between the two are significant. The results presented in the figure are almost discrete.

[0183] Therefore, in this embodiment, the specific method and experimental results of step 5 are as follows:

[0184] Step 5.1: Reread the original data, retaining only the 12 categories described in Step 4, and regenerate the data images and their labels. Then, import them into the two network models for training.

[0185] Step 5.2: After training, the results are as follows: the convolutional neural network using empirically selected parameters achieves an accuracy of 95% on the test set, while the convolutional neural network using network structure search achieves an accuracy of over 98%. Figure 11 and Figure 12 As shown.

[0186] The specific methods and experimental results of step 6 are as follows:

[0187] Step 6.1: Based on the original 12 data categories, add Data14 and Data16, and use convolutional neural networks based on network structure search and empirical parameter selection to classify the 14 data categories respectively;

[0188] Step 6.2: After training, the results are as follows: the accuracy of the convolutional neural network using empirically selected parameters on the test set is 83%, while the accuracy of the convolutional neural network using network structure search can reach over 92%. Figure 13 and Figure 14 As shown.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A classification method for aero-engine rotor vibration data based on network structure search, characterized in that: Includes the following steps: Step 1: Collect vibration data of the aero-engine rotor system under I different operating conditions; Step 2: Convert the collected vibration data of the aero-engine rotor system under different operating conditions into grayscale images; Step 3: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to perform I classification on all vibration data under I working conditions, and compare the classification results; Step 4: Use visualization dimensionality reduction techniques to analyze the separability of the original vibration data and identify vibration data with differences greater than a set threshold; Step 4.1: Use t-distributed random neighbor embedding (t-SNE) to visualize and reduce the dimensionality of the vibration data; For each type of image generated from vibration data under different working conditions, its corresponding matrix is... Straighten into a vector Using principal component analysis to transform vectors Reduce the dimensionality to 50 and then draw a visual t-SNE diagram; Step 4.2: Based on the visualized t-SNE plot, analyze the vibration data where the difference is greater than a set threshold; the difference refers to the Euclidean distance between two data vectors. ； in, Indicates the first The vector and the first Euclidean distance between vectors Indicates the first The first vector One component; Step 5: Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to classify various types of vibration data with differences greater than a set threshold, and compare the classification results; Step 6: Based on the vibration data used in Step 5, add several types of vibration data whose difference from the original vibration data is less than or equal to a set threshold. Use a convolutional neural network with parameters selected based on experience and a convolutional neural network based on network structure search to classify them respectively, and compare the classification results.

2. The classification method for aero-engine rotor vibration data based on network structure search according to claim 1, characterized in that: The specific method for step 2 is as follows: Step 2.1: Standardize the collected vibration data under different working conditions; Step 2.2: Convert the standardized vibration data into a grayscale image.

3. The classification method for aero-engine rotor vibration data based on network structure search according to claim 2, characterized in that: The specific method for step 2.1 is as follows: The collected vibration data under different working conditions are time series data. The data is set to be collected at the [number]th [period]. The length under this working condition is The original vibration data sequence is , , The vibration data is standardized as shown in the following formula: ； in, Indicates the first Data from the original vibration data sequence under various working conditions, Represents the standardized version of the first... Data from vibration data sequences under various working conditions Indicates the first The mean of all data within the original vibration data sequence under a certain working condition. Indicates the first The standard deviation of all data within the original vibration data sequence under various working conditions; Then the standardized first The vibration data sequence under the following working conditions is represented as follows: .

4. The classification method for aero-engine rotor vibration data based on network structure search according to claim 3, characterized in that: The specific method for step 2.2 is as follows: First, the vibration datasets under all operating conditions are truncated to a length of [missing value]. This data facilitates subsequent signal processing transformations and grayscale image conversions; The grayscale image to be converted from the vibration data under each working condition is defined as high-resolution. Width If the size is specified, then each grayscale image has one and only one unique size. OK Column matrix Correspondingly, each vibration data sequence after standardization can be converted into at most [number of data points]. The images include... This indicates rounding down; the specific conversion method is as follows: First, the vibration data sequences under each working condition after standardization are... Divided into training set and verification set Each contains several groups of lengths. Vibration data sequence , and This sequence is a subsequence of the original vibration data sequence; first, a Fast Fourier Transform (FFT) is performed on this subsequence, and then a grayscale image is generated based on the transformed result; (1) For a set of vibration data sequences ,Pick Perform a Fast Fourier Transform (FFT) on each point to obtain the spectrum, where... Since the sequence obtained by FFT is a symmetric complex sequence, after the original data is processed by FFT to obtain the spectrum, the... The length is sufficient; take the modulus value of each element in the frequency domain sequence and then divide by . Then, it is scaled down to the interval [0, 255] to obtain the frequency domain sequence. ,in, For the frequency domain sequence, the first... One spectrum data; (2) In order to embed the frequency domain sequence into the image, it is necessary to generate a set of time-series indices corresponding to the frequency domain sequence: The elements within it are calculated as follows: And it satisfies: And make ; in, Indicates to Round to the nearest integer. This represents the scaling factor for the index, and the formula will scale each time-series index value. All reduced to 1 to Within the range; (3) Based on the time index and the frequency domain sequence, rearrange the frequency domain sequence to obtain the rearranged frequency domain sequence. The sorting rules are as follows: ； in, Indicates the rearrangement of the frequency domain sequence at the th The element value at the index, Indicates the original frequency domain sequence at the th The value of the element at the index; The generated rearranged frequency domain sequence is transformed into OK Column matrix This will generate a grayscale image.

5. The classification method for aero-engine rotor vibration data based on network structure search according to claim 4, characterized in that: The convolutional neural network with parameters selected based on experience includes six convolutional layers (conv), a normalization layer (BN), a ReLU activation function layer, a pooling layer (MaxPool), and a fully connected layer (Fully Connection). Fallout means flattening the matrix into a vector, and dropout means randomly discarding data to prevent overfitting.

6. The classification method for aero-engine rotor vibration data based on network structure search according to claim 5, characterized in that: The structure of the convolutional neural network based on empirically selected parameters is defined as follows: conv1: 1-8 @3×3, padding = 1, stride = 1; BN + ReLU; MaxPool: @2×2, stride = 2; conv2: 8-16 @3×3, padding = 1, stride = 1; BN + ReLU; MaxPool: @2×2, stride = 2; conv3_1: 16-32 @3×3, padding = 1, stride = 1; BN + ReLU; conv3_2: 32-32 @3×3, padding = 1, stride = 1; BN + ReLU; MaxPool: @2×2, stride = 2; conv4_1: 32-64 @3×3, padding = 1, stride = 1; BN + ReLU; conv4_2: 64-64 @3×3, padding = 1, stride = 1; BN + ReLU; MaxPool: @2×2, stride = 2; Flatten; Fully Connection1 + ReLu + dropout; Fully Connection2 + ReLU; softMax; Where conv represents a convolutional layer, BN represents a normalized layer, ReLU represents a ReLU activation function layer, MaxPool represents a pooling layer, Fully Connection represents a fully connected layer, Fallout represents flattening the matrix into a vector, and dropout represents randomly dropping values ​​to prevent overfitting.

7. The classification method for aero-engine rotor vibration data based on network structure search according to claim 6, characterized in that: The basic architecture of the convolutional neural network based on network structure search includes defining the number of convolutional layers, the pooling layer structure, and the structure and number of fully connected layers; the number of input and output channels of each convolutional layer, the kernel size, and the structure-related hyperparameters of the fully connected layers are obtained through search. Given the search space for hyperparameters, and restricting the hyperparameters to integers within the search space; through random search, a set of hyperparameters will be obtained within the specified search space; The hyperparameters are updated to update the convolutional neural network structure, and the network is trained for a specified number of epochs to obtain the network classification accuracy for the dataset under this structure. After training reaches a given number of rounds, new parameters are obtained through random search, and training continues. The accuracy of each search is recorded. When training is completed, the classification accuracy under different network structures can be obtained. The network structure parameters corresponding to the highest accuracy are taken as the desired result.

8. The classification method for aero-engine rotor vibration data based on network structure search according to claim 7, characterized in that: The convolutional neural network based on network structure search includes five convolutional layers (conv), a ReLU activation function layer, a pooling layer (MaxPool), and a fully connected layer (Fully Connection). Fallout means flattening the matrix into a vector, and dropout means randomly discarding data to prevent overfitting.

9. The classification method for aero-engine rotor vibration data based on network structure search according to claim 8, characterized in that: The fully connected layers of the convolutional neural network based on network architecture search are configured the same as those of a regular convolutional neural network, and the definitions of each convolutional layer are as follows: conv1: Input channel 1, output channel search space: [1, 100), kernel scale search space: [1, 10); ReLU activation function; maxPool: Pooling kernel size: 2×2; conv2: The number of input channels is equal to the number of output channels of conv1. The output channel search space is [1, 100), and the kernel scale search space is [1, 10). ReLU activation function; maxPool: Pooling kernel size: 2×2; conv3: The number of input channels is equal to the number of output channels of conv2. The output channel search space is [1, 100), and the kernel scale search space is [1, 10). ReLU activation function; maxPool: Pooling kernel size: 2×2; conv4: The number of input channels is equal to the number of output channels of conv3. The output channel search space is [1, 100), and the kernel scale search space is [1, 10). ReLU activation function; maxPool: Pooling kernel size: 2×2; conv5: The number of input channels is equal to the number of output channels of conv4. The output channel search space is [1, 100), and the kernel scale search space is [1, 10). ReLU activation function; maxPool: Pooling kernel size: 2×2.

Images