Aeroengine fault diagnosis method based on microphone array acoustic signal analysis

Abstract

The present invention discloses an aircraft engine fault diagnosis method based on microphone array acoustic signal analysis, including obtaining microphone array acoustic signals of the aircraft engine in multiple operating states; constructing an aircraft engine fault diagnosis model based on a two-dimensional time domain convolutional neural network and performing model training; obtaining microphone array acoustic signals in the real-time operating state of the aircraft engine; and performing aircraft engine fault diagnosis according to the microphone array acoustic signals in the real-time operating state of the aircraft engine using the trained aircraft engine fault diagnosis model based on a two-dimensional time domain convolutional neural network. The present invention realizes the diagnosis of weak fault aircraft engines under a strong fault noise background using multi-channel acoustic signals collected by a microphone array, and the fault diagnosis result is more accurate and has a stronger anti-noise capability. At the same time, the number of microphones can be reduced, thereby reducing the fault detection cost.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault detection, and in particular to an aircraft engine fault diagnosis method based on microphone array acoustic signal analysis. Background Art

[0002] Aircraft engines are key components of modern air transportation. As the core power components of spacecraft, aircraft engines are mostly in extreme working conditions such as high temperature, high pressure, and strong vibration. This means that any small fault in any part of the engine may have a serious impact on the entire power system of the aircraft, leading to huge economic losses or even casualties. Therefore, accurate and timely operational fault detection of aerospace engines is crucial to improving their reliability and safety. This technology has great practical significance and economic value. However, fault detection of aircraft engines has always been a challenging problem.

[0003] In the field of aircraft engine fault diagnosis, most non-destructive testing methods are based on vibration signals for fault detection. Vibration signals can usually be collected using non-contact or contact vibration sensors, such as accelerometers, vibration sensors, etc. Non-contact vibration sensors are suitable for long distances and can provide a wider range of vibration signals, but the accuracy is relatively low; contact vibration sensors are more accurate, but need to be installed on the engine surface and are not suitable for special conditions such as high temperature and high speed.

[0004] Acoustic detection technology, as a non-contact non-destructive detection technology that is gaining more and more attention, has the advantages of fast response speed, simple application, low cost, and wide application scenarios. However, unlike vibration signals, sound signals are more easily affected by the environment due to the different signal transmission media, such as environmental noise, wall reflection, and aliasing of different engine sound signals, which will have a significant impact on the sound signal.

[0005] In an aircraft engine maintenance workshop, there are usually multiple engines being repaired at the same time. If one of the engines fails, it needs to be diagnosed and repaired in time through fault detection. However, due to the influence of noise aliasing between engines, when there are multiple faulty engines, conventional acoustic detection methods are difficult to accurately detect the engine with weaker fault sound signals, resulting in diagnostic errors.

[0006] In traditional fault diagnosis methods, data filtering technology is usually used to pre-process the collected data to remove noise and interference. Subsequently, time domain analysis and frequency domain analysis are used to extract characteristic information of aircraft engines, and fault diagnosis is performed in combination with previous experience and knowledge. This method can achieve relatively accurate fault diagnosis results to a certain extent, but it also has problems such as insufficient real-time performance and poor accuracy.

[0007] In recent years, with the development of artificial intelligence technology, methods based on convolutional neural networks and multi-layer perceptrons have gradually become mainstream. These methods can automatically extract feature information from a large amount of data through training and learning, and have higher adaptability and generalization capabilities, which are superior to traditional methods. However, these methods still have poor performance in detecting weak faults under strong fault interference.

[0008] In the context of strong fault signals, weak fault detection becomes more difficult. How to extract useful information from extremely complex signals is an important challenge in the field of aircraft engine fault diagnosis. To address this problem, researchers have proposed a method based on deep learning, combining multi-task learning with acoustic fault detection to develop a new diagnostic technology that can detect and locate weak faults in a strong noise background, effectively solving the limitations of traditional methods. These technologies have attracted widespread attention due to their real-time and high accuracy.

[0009] Aircraft engines generate vibration signals when they are running, and when the equipment fails or is in an abnormal situation, the characteristics of the vibration signals will change. Therefore, by real-time monitoring and analysis of the vibration signals generated when the equipment is running, it is possible to determine whether the equipment has a fault or an abnormal situation. When the equipment generates vibration signals, it also generates sound signals. Therefore, the operating status of the equipment can also be determined by detecting the sound signals and their characteristics. However, since the transmission medium of the sound signal is air, it is very easy to be interfered by various forms, such as reverberation, environmental noise, etc.

[0010] At present, the noises of different engines in aircraft engine maintenance workshops overlap with each other, making it impossible to accurately diagnose and locate the engines with weaker faults through conventional acoustic fault detection methods. In addition, existing fault detection methods also have problems such as low efficiency and high labor costs. Summary of the invention

[0011] In view of the above-mentioned deficiencies in the prior art, the present invention provides an aircraft engine fault diagnosis method based on microphone array acoustic signal analysis, so as to realize the diagnosis of weak fault aircraft engines under the background of strong fault noise by utilizing the multi-channel acoustic signals collected by the microphone array, and at the same time, the directional angle of the weak fault engine relative to the microphone array can be determined.

[0012] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions to achieve it.

[0013] The present invention provides an aeroengine fault diagnosis method based on microphone array acoustic signal analysis, comprising the following steps:

[0014] Acquire microphone array acoustic signals of aircraft engines in multiple operating states;

[0015] An aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network is constructed, and the model is trained using microphone array acoustic signals of the aircraft engine under multiple operating states; the aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network specifically includes: an encoder, a mask estimation network based on a two-dimensional time-domain convolutional neural network, and a decoder connected in sequence; the encoder is used to convert the microphone array acoustic signal into a stacked feature vector of an intermediate feature space; the mask estimation network based on a two-dimensional time-domain convolutional neural network is used to generate a mask of a feature vector according to the stacked feature vector converted by the encoder, and extract feature relationships in different channels and time in the acoustic signal; the decoder is used to perform feature classification according to the feature relationships in different channels and time extracted by the mask estimation network based on the two-dimensional time-domain convolutional neural network;

[0016] Acquire microphone array acoustic signals of aircraft engines in real-time operation;

[0017] The trained aircraft engine fault diagnosis model based on two-dimensional time-domain convolutional neural network is used to perform aircraft engine fault diagnosis according to the microphone array acoustic signals under the real-time operating state of the aircraft engine.

[0018] Furthermore, the step of obtaining microphone array acoustic signals of the aircraft engine in multiple operating states specifically includes:

[0019] A microphone array is used to obtain microphone array acoustic signals of an aircraft engine under multiple operating conditions;

[0020] The microphone array comprises a detachable microphone, an involute microphone stand, a retractable base and a signal acquisition card.

[0021] Furthermore, the step of acquiring microphone array acoustic signals of the aircraft engine in multiple operating states further includes:

[0022] Performing low-frequency filtering on the microphone array acoustic signals acquired under multiple operating states of the aircraft engine to filter out low-frequency signals;

[0023] The microphone array acoustic signal after filtering is converted into a multi-channel segmented feature signal.

[0024] Furthermore, the encoder specifically includes: a first convolutional layer and a stacked layer connected in sequence;

[0025] The first convolution layer is used to perform a one-dimensional convolution operation on each channel signal in the microphone array acoustic signal to obtain a feature vector corresponding to each channel signal;

[0026] The stacking layer is used to perform a stacking operation on the feature vectors corresponding to the signals of each channel to obtain a stacked feature vector.

[0027] Furthermore, the mask estimation network based on the two-dimensional time-domain convolutional neural network specifically includes: a first normalization layer, a second convolutional layer, a third convolutional layer, a two-dimensional time-domain convolutional neural network, a first parameterized activation layer, a fourth convolutional layer and a fifth convolutional layer connected in sequence;

[0028] The first normalization layer is used to perform layer normalization processing on the stacked feature vector output by the encoder;

[0029] The second convolution layer is used to perform a convolution operation on the normalized feature vector output by the first normalization layer in a first dimension to obtain a first feature vector;

[0030] The third convolution layer is used to perform a convolution operation on the first feature vector output by the second convolution layer in a second dimension to obtain a second feature vector;

[0031] The two-dimensional time-domain convolutional neural network is used to extract feature relationships in different channels and time from the second feature vector output by the third convolutional layer;

[0032] The first parameterized activation layer is used to perform nonlinear transformation on the feature relationships between different channels and time extracted by the two-dimensional time-domain convolutional neural network;

[0033] The fourth convolutional layer is used to perform a convolution operation on the feature vector output by the first parameterized activation layer in a third dimension to obtain a third feature vector;

[0034] The fifth convolutional layer is used to perform a convolution operation on the third eigenvector output by the fourth convolutional layer in a fourth dimension to obtain a fourth eigenvector.

[0035] Furthermore, the two-dimensional time-domain convolutional neural network specifically includes: a plurality of two-dimensional time-domain convolutional modules connected in sequence; each two-dimensional time-domain convolutional module includes a plurality of two-dimensional convolutional blocks connected in sequence;

[0036] The two-dimensional time domain convolution module is used to extract feature relationships in different channels and time from the second feature vector output by the third convolution layer, and to perform feature fusion on the feature relationships extracted by multiple two-dimensional convolution blocks connected in sequence in each two-dimensional time domain convolution module.

[0037] Furthermore, the two-dimensional convolution block specifically includes: a sixth convolution layer, a second parameterized activation layer, a second normalization layer, a two-dimensional convolution layer, a third parameterized activation layer, a third normalization layer, a seventh convolution layer and an eighth convolution layer connected in sequence;

[0038] The sixth convolutional layer is used to increase the number of channels of the output feature vector to the number of hidden layers;

[0039] The second parameterized activation layer is used to perform nonlinear transformation on the feature vector output by the sixth convolutional layer;

[0040] The second normalization layer is used to perform layer normalization processing on the feature vector output by the second parameterized activation layer;

[0041] The two-dimensional convolution layer is used to perform two-dimensional deep convolution on the feature vector output by the two-dimensional convolution layer in the feature and time dimensions to extract features and time-related feature information;

[0042] The third parameterized activation layer is used to perform nonlinear transformation on the feature information extracted by the two-dimensional convolutional layer;

[0043] The third normalization layer is used to perform layer normalization processing on the feature vector output by the third parameterized activation layer;

[0044] The seventh convolutional layer is used to extract the feature relationship between channels from the feature vector output by the third normalization layer, and transmit the feature information to the next connected two-dimensional convolutional block through the residual path;

[0045] The eighth convolutional layer is used to extract the feature relationship between channels of the feature vector output by the third normalization layer, and perform feature fusion with the feature relationship extracted by other two-dimensional convolution blocks through jump connection.

[0046] Furthermore, the decoder specifically includes: a parallel fault classification channel and a fault location regression channel;

[0047] The fault classification channel is used to determine the predicted probability of each fault category according to the fourth eigenvector output by the fifth convolutional layer to obtain a fault classification result;

[0048] The fault position regression channel determines the prediction result of the regression target according to the fourth eigenvector output by the fifth convolutional layer to obtain the direction angle of the fault position.

[0049] Further, the fault classification channel includes a ninth convolutional layer, a first fully connected layer, an adaptive average pooling layer, and a softmax activation layer connected in sequence;

[0050] The fault location regression channel includes a tenth convolutional layer and a second fully connected layer connected in sequence.

[0051] Furthermore, the total loss function used by the aircraft engine fault diagnosis model is:

[0052] L MTL =w1L C +w2L MSE;

[0053]

[0054] Among them, L C The loss function used in the fault classification task; L MSE represents the loss function of the faulty engine direction task; w m (t) represents the weight of the mth task, m = 1 or 2; L m (t-1), r m (t-1) represents the loss and training speed of task m at step t-1 respectively; T represents a constant. When T=1, w m Equal to the result of softmax; when T is large enough, w m Tends to 1, and the loss weights of each task are the same.

[0055] Compared with the prior art, the aircraft engine fault diagnosis method based on microphone array acoustic signal analysis provided by the present invention has the following beneficial effects:

[0056] (1) In the existing fault diagnosis systems based on microphone arrays, most of them adopt the traditional sound source separation algorithm, and then perform fault diagnosis on the separated sound source signals after separating the sound sources; this type of method itself has the disadvantages of slow calculation speed, poor noise resistance, high requirements on the number of microphones, high requirements on detection distance, frequency band, etc.; while the method adopted by the present invention can directly input the microphone array signal and output the final fault diagnosis and positioning results, omitting the process of separating the sound sources and further simplifying the calculation process;

[0057] (2) Existing diagnostic systems that use deep learning methods have problems such as cumbersome calculations. At the same time, the deep learning algorithms they use cannot make good use of the relationship between multiple channel information, and thus cannot achieve good results. The deep learning network used in the present invention, compared to other deep learning networks used for fault diagnosis, can better utilize the relationship between 16 channel data, thereby making the results more accurate and more resistant to noise. At the same time, it can also reduce the number of microphones and reduce the cost of fault detection.

[0058] (3) Most of the current acoustic-based fault diagnosis methods use microphones or microphone arrays to directly detect a single aircraft engine. This method has very large limitations in actual maintenance workshops. Because there is often more than one aircraft engine that may have a fault in an actual maintenance workshop, the fault signals of each aircraft engine will interfere with each other, thereby affecting the final diagnosis result. The multi-task deep learning method used in the present invention can solve this problem. After collecting the fault signal through the microphone array, the weak fault type and its direction angle under the interference of the strong signal source are output, which can very intuitively and accurately point out the type and location of the faulty engine. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic flow chart of an aircraft engine fault diagnosis method based on microphone array acoustic signal analysis provided by an embodiment of the present invention;

[0060] Figure 2 is a schematic diagram of the structure of a microphone array in an embodiment of the present invention;

[0061] Figure 3 Schematic diagram of 16-channel acoustic data collected by a microphone array in an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of the structure of an aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network in an embodiment of the present invention;

[0063] Figure 5 Schematic diagram of the structure of a two-dimensional convolution block in an embodiment of the present invention;

[0064] Figure 6 1 is a T-SNE diagram of the aircraft engine fault classification result obtained in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0066] Example

[0067] like Figure 1 As shown, the embodiment of the present invention provides an aircraft engine fault diagnosis method based on microphone array acoustic signal analysis, comprising the following steps S1 to S4:

[0068] S1. Acquire microphone array acoustic signals of an aircraft engine under multiple operating states;

[0069] In an optional embodiment of the present invention, step S1 of this embodiment specifically uses a microphone array to obtain microphone array acoustic signals of the aircraft engine in multiple operating states; the microphone array used in this embodiment includes a detachable microphone, an involute microphone holder, a retractable base and a signal acquisition card, such as Figure 2 shown.

[0070] The detachable microphone in this embodiment can adopt the GRAS 46AE microphone set of the Danish GRAS company. The microphone chip is a high-quality IEC 61094WS2F standardized GRAS 40AE 1 / 2 inch polarization-ready free-field microphone, designed for long-term reliability in a variety of environments. High sensitivity and reliability make 40AE the preferred measurement microphone for sound level meters / analyzers equipped with the CCP option, and thus make IEC 61672 Class 1 measurements possible. Polarization is obtained by applying an electrostatic charge to the microphone backplate under highly controlled conditions. The electrostatic charge is then charged to match the specified microphone sensitivity. The preamplifier is a GRAS26CA1 / 2 inch CCP standard preamplifier with a BNC connector. In industrial applications, it has the characteristics of low self-noise, wide frequency and excellent scanning rate.

[0071] In this embodiment, the involute microphone holder is used to arrange the microphones in the form of an involute. The spacing between the microphones and the size of the microphone array will directly affect the positioning accuracy of the microphone array and the applicable frequency range. The specific design calculation method is as follows:

[0072] 1. Determine the target frequency band and the radius R of the target area;

[0073] 2. Calculate the maximum order of the sound field N = [ekR / 2];

[0074] 3. Determine the number of circular arrays to be implemented based on the maximum order N. For a first-order microphone configuration, there must be at least Nomni = [N / 2] omnidirectional sensor arrays and Nfirst = N-Nomni first-order arrays. For a differential microphone configuration, no less than Ndiff = [N / 2] pairs of microphone arrays are required;

[0075] 4. Determine the radius of each circular array; choose the radius so that the spherical Bessel zeros in the target frequency band are avoided. Ensure that the radius of the circular array has good diversity.

[0076] In this embodiment, the involute microphone holder includes three parts: a microphone clamp, a support plate, and a base. The microphone clamp is used to clamp the microphone, and the microphone is locked and fixed to the support plate by bolts. The support plate is used to position the microphone and keep the microphone in a specified position; in this embodiment, two microphones are fixed to one support plate, and the two microphones are facing opposite directions. The base is used to fix the support plate, and the relative position between the support plates is fixed by bolts, and the microphone holder and the retractable base are connected at the same time. The retractable base is directly connected to the microphone holder, and is used to clamp the microphone holder to the required height, while ensuring the stability of the microphone array.

[0077] The signal acquisition card in this embodiment is composed of the NI 9234 acquisition module and the NI 9189 chassis from National Instruments. The NI 9234 is a four-channel dynamic signal acquisition module for high-precision measurements from IEPE sensors. The NI 9234 provides a dynamic range of 102dB and uses integrated electronic piezoelectric signal conditioning to provide a 2mA constant current in the accelerometer and microphone. Four input channels can be acquired simultaneously at a rate of up to 51.2kS / s. In addition, the module also includes a built-in anti-aliasing filter that automatically adjusts to the required sampling rate. Compatible with single-module USB carriers and NI CompactDAQ and CompactRIO hardware, the NI 9234 is ideal for a variety of mobile or portable applications, such as industrial machine condition monitoring and in-vehicle noise, vibration and harshness testing. The NI 9189 chassis controls the timing, synchronization, and data transmission between C Series I / O modules and external hosts. These chassis provide connection options such as USB, Ethernet, or WiFi, and have multiple slots to provide sufficient I / O interfaces for a variety of applications.

[0078] This embodiment adopts Figure 2 The microphone array shown in the figure collects data from aircraft engines in multiple operating states in an actual working environment. When collecting data, it is necessary to ensure that there are enough aircraft engines with various types of faults as much as possible and to place the microphone array in different relative positions to ensure that the number of collected data sets is sufficient to complete the training of the model. Figure 3 As shown, 16 channels of acoustic data collected by the microphone array.

[0079] In an optional embodiment of the present invention, since in an actual application environment, low-frequency sound waves are more likely to produce reverberation and thus increase noise and reduce the signal-to-noise ratio, and at the same time the rotation speed of the aircraft engine is relatively high during normal operation, this embodiment pre-processes the collected microphone array acoustic signals, specifically including low-frequency filtering of the microphone array acoustic signals obtained from the aircraft engine in multiple operating states to filter out the low-frequency signals, thereby reducing noise and increasing the signal-to-noise ratio without losing the fault signal; before the original signal is input into the network, it is also necessary to convert it into a suitable format, that is, convert the filtered microphone array acoustic signal into a multi-channel segmented feature signal.

[0080] S2. Construct an aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network, and use the microphone array acoustic signals obtained from the aircraft engine under multiple operating conditions to train the model.

[0081] In an optional embodiment of the present invention, Figure 4 As shown, the aircraft engine fault diagnosis model based on the two-dimensional time-domain convolutional neural network constructed in step S2 of this embodiment specifically includes: an encoder, a mask estimation network based on the two-dimensional time-domain convolutional neural network, and a decoder connected in sequence.

[0082] The encoder is used to convert the microphone array acoustic signal into a stacked feature vector in an intermediate feature space;

[0083] The mask estimation network based on the two-dimensional time-domain convolutional neural network is used to generate a mask of the feature vector according to the stacked feature vector transformed by the encoder, and to extract the feature relationship of different channels and time in the acoustic signal;

[0084] The decoder is used to perform feature classification based on the feature relationships across different channels and time extracted by the mask estimation network based on a 2D time-domain convolutional neural network.

[0085] In this embodiment, the encoder specifically includes: a first convolutional layer and a stacking layer connected in sequence.

[0086] The first convolution layer is used to perform a one-dimensional convolution operation on each channel signal in the microphone array acoustic signal to obtain a feature vector corresponding to each channel signal;

[0087] The stacking layer is used to stack the feature vectors corresponding to each channel signal to obtain a stacked feature vector.

[0088] This embodiment uses the encoder to convert the input 16-channel acoustic signal into a feature representation, processes the acoustic signal of each channel through a one-dimensional convolution operation, and then stacks the convolution features as the encoder output features. The encoder output features will become a three-dimensional feature, which helps to retain information between channels. Compared with the traditional deep learning network that directly linearly adds the data of each channel and inputs it to the next layer, stacking the features after one-dimensional convolution in a way that adds one dimension can better retain the information between channels in the subsequent calculation process.

[0089] In this embodiment, the mask estimation network based on the two-dimensional time-domain convolutional neural network specifically includes: a first normalization layer, a second convolutional layer, a third convolutional layer, a two-dimensional time-domain convolutional neural network, a first parameterized activation layer, a fourth convolutional layer and a fifth convolutional layer connected in sequence.

[0090] The first normalization layer is used to perform layer normalization on the stacked feature vector output by the encoder;

[0091] The second convolution layer is used to perform a convolution operation on the normalized feature vector output by the first normalization layer in the first dimension to obtain a first feature vector;

[0092] The third convolutional layer is used to perform a convolution operation on the first feature vector output by the second convolutional layer in the second dimension to obtain a second feature vector;

[0093] The two-dimensional time-domain convolutional neural network is used to extract the feature relationships in different channels and time from the second feature vector output by the third convolutional layer;

[0094] The first parameterized activation layer is used to perform nonlinear transformation on the feature relationships between different channels and time extracted by the two-dimensional time-domain convolutional neural network;

[0095] The fourth convolutional layer is used to perform a convolution operation on the feature vector output by the first parameterized activation layer in a third dimension to obtain a third feature vector;

[0096] The fifth convolutional layer is used to perform a convolution operation on the third eigenvector output by the fourth convolutional layer in a fourth dimension to obtain a fourth eigenvector.

[0097] This embodiment uses a mask estimation network to generate a mask of a source signal. Specifically, a layer normalization layer is used to normalize the input stacked feature vector to help the network learn features better. Then, a 1×1 convolution layer (N) and a 1×1 convolution layer (C) are used to perform convolution operations in different dimensions to further extract features.

[0098] This embodiment uses a two-dimensional time-domain convolutional neural network in the mask estimation network to extract feature relationships between different channels and time in the audio signal.

[0099] In this embodiment, the two-dimensional time-domain convolutional neural network specifically includes: a plurality of two-dimensional time-domain convolutional modules connected in sequence; each two-dimensional time-domain convolutional module includes a plurality of two-dimensional convolutional blocks connected in sequence.

[0100] The two-dimensional time domain convolution module is used to extract the feature relationships in different channels and time for the second feature vector output by the third convolution layer, and to perform feature fusion on the feature relationships extracted by the multiple two-dimensional convolution blocks connected in sequence in each two-dimensional time domain convolution module.

[0101] like Figure 5 As shown, the two-dimensional convolution block specifically includes: a sixth convolution layer, a second parameterized activation layer, a second normalization layer, a two-dimensional convolution layer, a third parameterized activation layer, a third normalization layer, a seventh convolution layer and an eighth convolution layer connected in sequence.

[0102] The sixth convolutional layer is used to increase the number of channels of the output feature vector to the number of hidden layers.

[0103] The second parameterized activation layer is used to perform nonlinear transformation on the feature vector output by the sixth convolutional layer;

[0104] The second normalization layer is used to perform layer normalization processing on the feature vector output by the second parameterized activation layer;

[0105] The two-dimensional convolution layer is used to perform two-dimensional deep convolution on the feature vector output by the two-dimensional convolution layer in the feature and time dimensions to extract features and time-related feature information;

[0106] The third parameterized activation layer is used to perform nonlinear transformation on the feature information extracted by the two-dimensional convolutional layer;

[0107] The third normalization layer is used to perform layer normalization processing on the feature vector output by the third parameterized activation layer;

[0108] The seventh convolutional layer is used to extract the feature relationship between channels from the feature vector output by the third normalization layer, and transfer the feature information to the next connected two-dimensional convolutional block through the residual path;

[0109] The eighth convolutional layer is used to extract the feature relationship between channels of the feature vector output by the third normalization layer, and perform feature fusion with the feature relationship extracted by other two-dimensional convolution blocks through jump connections.

[0110] In this embodiment, a 2-D Conv block is used in a two-dimensional time-domain convolutional neural network to increase the number of input channels C to the number of hidden layers H (H = 4C) through a 1×1 convolution layer, keeping the feature and time dimensions unchanged to increase the diversity between channels; at the same time, a D-Conv (2D) block in the two-dimensional convolution block is used to perform two-dimensional deep convolution in the feature and time dimensions to extract feature information related to features and time; then a 2-D zero padding block is used in the D-Conv block to use two-dimensional zero padding to keep the size of the feature map unchanged; finally, two 1×1 convolution layers are used to extract the relationship between channels, and feature information is transmitted through skip connections and residual paths. Unlike the traditional deep learning network that uses a 1×1 convolution in the feature dimension, the 1×1 convolution layer used in this embodiment only extracts information in the channel dimension to extract feature information between channels.

[0111] In this embodiment, the decoder specifically includes: a parallel fault classification channel and a fault location regression channel.

[0112] The fault classification channel is used to determine the predicted probability of each fault category according to the fourth eigenvector output by the fifth convolutional layer to obtain the fault classification result;

[0113] The fault location regression channel determines the prediction result of the regression target according to the fourth eigenvector output by the fifth convolutional layer to obtain the direction angle of the fault location.

[0114] The fault classification channel includes the ninth convolutional layer, the first fully connected layer, the adaptive average pooling layer and the softmax activation layer connected in sequence. The fault location regression channel includes the tenth convolutional layer and the second fully connected layer connected in sequence.

[0115] After a one-dimensional convolution, the fault classification channel reduces the dimension of the output feature to 1 through global average pooling, and then passes through a fully connected layer and a softmax activation function to obtain the fault classification result; the role of the fully connected layer is to learn the complex mapping relationship between features and categories. By learning weights and biases, it can automatically extract the abstract features required for classification from the input features. It can map high-dimensional features to a low-dimensional space that matches the number of categories, and then use a suitable activation function (such as softmax) to normalize the output to obtain a probability prediction for each category. The softmax activation function makes the sum of the output probabilities on each category equal to 1, and the probability value of each category is between 0 and 1.

[0116] The fault location regression channel also undergoes a one-dimensional convolution, and then passes the result through a fully connected layer to obtain the direction angle of the sound source position. In the regression task, the function of the fully connected layer is to map the output of the convolution layer or other feature extraction layer to the dimension of the regression target. Its input can be the feature map or other feature vector output by the convolution layer, and the output is the prediction result of the regression target.

[0117] This embodiment uses a decoder to classify the input features. For the fault classification task, input 1 is sent to the first convolution layer for classification; for the location classification task, input 2 is sent to the second convolution layer for classification; thereby achieving the source separation and classification tasks of the audio signal.

[0118] In this embodiment, the 16-channel acoustic data processed in step S1 is input into the constructed aircraft engine fault diagnosis model based on the two-dimensional time-domain convolutional neural network for training. The loss function used in the training is as follows:

[0119] 1. Loss function for fault classification task:

[0120] The loss function used for the fault classification task is the cross entropy loss function:

[0121]

[0122] In the above formula, y i,k represents the true label value of the i-th sample, p i,k It represents the probability value of the actual output after softmax calculation. This formula can measure the distance between the true distribution and the actual output distribution. n represents the number of samples.

[0123] p i,k A normalized exponential function is used to compress the k-dimensional vector and map multiple outputs to the (0,1) interval.

[0124]

[0125] Among them, z represents z classification, k, k′ represent vector dimensions, k, k′=1, 2, …K, K represents the total vector dimension.

[0126] 2. Loss function for the weak fault engine direction task (i.e. fault location):

[0127] The mean square error loss function measures the average of the squares of the differences between the predicted and target values.

[0128]

[0129] Among them, f(x) represents the predicted value, y represents the target value, and n represents the number of samples.

[0130] 3. Total loss function-dynamic weighted average loss function:

[0131] L MTL =w1L C +w2L MSE ;

[0132]

[0133] Among them, w m (t) represents the weight of the mth task, m = 1 or 2; L m (t-1), r m (t-1) represents the loss and training speed of task m at step t-1 respectively; T represents a constant. When T=1, w m Equal to the result of softmax; when T is large enough, w m Tends to 1, and the loss weights of each task are the same.

[0134] The basic idea of ​​the training of the present invention is that different tasks have different levels of difficulty and different learning speeds. Different learning rates can be set for different tasks, but a better idea is to dynamically adjust so that each task learns at a similar speed. This is the core idea of ​​the Dynamic Weight Averaging algorithm.

[0135] For tasks with a fast loss decrease, the weight will become smaller; otherwise, the weight will become larger. Determine whether the training effect meets the requirements: After completing the training of the neural network, the training results need to be evaluated and optimized. This mainly includes the accuracy calculation of the training set and the test set, the judgment of overfitting and underfitting, the adjustment of the learning rate, cross-validation, and the adjustment of network updates, batch size and other settings. If the training results meet the requirements, proceed to the next step S3, otherwise repeat steps S1-S2.

[0136] S3. Acquire the microphone array acoustic signal of the aircraft engine in real-time operation state.

[0137] In an optional embodiment of the present invention, this embodiment uses a microphone array to collect and pre-process the acoustic signals of the aircraft engine during actual fault diagnosis. Similar to step S1, a suitable location should be selected under suitable environmental conditions to collect comprehensive engine sound signals. In the collection process, relevant parameters such as sound sampling frequency, sampling duration, sampling time interval, etc. need to be correctly set.

[0138] S4. Use the trained aircraft engine fault diagnosis model based on two-dimensional time-domain convolutional neural network to perform aircraft engine fault diagnosis according to the microphone array acoustic signal under the real-time operating state of the aircraft engine.

[0139] In an optional embodiment of the present invention, this embodiment inputs the microphone array acoustic signal of the aircraft engine in real-time operation after step S3 into the trained aircraft engine fault diagnosis model based on the two-dimensional time domain convolutional neural network for classification and positioning, so as to detect the fault category and location of the aircraft engine. Before the original signal is input into the network, it needs to be converted into a suitable format, such as converting the audio signal into a segmented feature form acceptable to the neural network. Afterwards, the trained model is used to perform relevant calculations of the neural network, and the input signal is processed to output the fault category and location. It should be noted that in actual applications, the output results need to be further processed and verified to improve the accuracy and reliability of the detection.

[0140] According to the above process, the sound signal of the aircraft engine can be input, and after the neural network analysis, the corresponding fault category and location information can be obtained. In addition, the output results also need to be analyzed and verified for validity to ensure the accuracy and reliability of the results. At the same time, the model and algorithm can be fed back and optimized to achieve the most optimized engine fault diagnosis system. Ultimately, it can be used for necessary engine maintenance, thereby improving engine performance efficiency and ensuring aviation safety.

[0141] The microphone array acoustic signal obtained in the real-time operation state of the aircraft engine is input into the trained aircraft engine fault diagnosis model based on the two-dimensional time domain convolutional neural network for classification (0-9 types of faults). The classification results are as follows: Figure 6 shown. Figure 6 The T-SNE diagram of the fault classification results is given. It can be seen from the figure that effective classification of aircraft engine faults can be achieved.

[0142] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0143] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0144] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0145] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0146] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. An aircraft engine fault diagnosis method based on microphone array acoustic signal analysis, characterized in that: The following steps are involved: Acquire microphone array acoustic signals of aircraft engines in multiple operating states; An aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network is constructed, and the model is trained using microphone array acoustic signals of the aircraft engine under multiple operating states; the aircraft engine fault diagnosis model based on a two-dimensional time-domain convolutional neural network specifically includes: an encoder, a mask estimation network based on a two-dimensional time-domain convolutional neural network, and a decoder connected in sequence; the encoder is used to convert the microphone array acoustic signal into a stacked feature vector of an intermediate feature space; the mask estimation network based on a two-dimensional time-domain convolutional neural network is used to generate a mask of a feature vector according to the stacked feature vector converted by the encoder, and extract feature relationships in different channels and time in the acoustic signal; the decoder is used to perform feature classification according to the feature relationships in different channels and time extracted by the mask estimation network based on the two-dimensional time-domain convolutional neural network; Acquire microphone array acoustic signals of aircraft engines in real-time operation; The trained aircraft engine fault diagnosis model based on two-dimensional time-domain convolutional neural network is used to perform aircraft engine fault diagnosis according to the microphone array acoustic signals under the real-time operating state of the aircraft engine.

2. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 1, characterized in that: The step of obtaining microphone array acoustic signals of the aircraft engine in multiple operating states specifically includes: A microphone array is used to obtain microphone array acoustic signals of an aircraft engine under multiple operating conditions.

3. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 1, characterized in that: The step of obtaining microphone array acoustic signals of the aircraft engine in multiple operating states further includes: Performing low-frequency filtering on the microphone array acoustic signals acquired under multiple operating states of the aircraft engine to filter out low-frequency signals; The microphone array acoustic signal after filtering is converted into a multi-channel segmented feature signal.

4. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 1, characterized in that: The encoder specifically comprises: a first convolutional layer and a stacking layer connected in sequence; The first convolution layer is used to perform a one-dimensional convolution operation on each channel signal in the microphone array acoustic signal to obtain a feature vector corresponding to each channel signal; The stacking layer is used to perform a stacking operation on the feature vectors corresponding to the signals of each channel to obtain a stacked feature vector.

5. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 4, characterized in that: The mask estimation network based on the two-dimensional time-domain convolutional neural network specifically includes: a first normalization layer, a second convolutional layer, a third convolutional layer, a two-dimensional time-domain convolutional neural network, a first parameterized activation layer, a fourth convolutional layer and a fifth convolutional layer connected in sequence; The first normalization layer is used to perform layer normalization processing on the stacked feature vector output by the encoder; The second convolution layer is used to perform a convolution operation on the normalized feature vector output by the first normalization layer in a first dimension to obtain a first feature vector; The third convolutional layer is used to perform a convolution operation on the first feature vector output by the second convolutional layer in a second dimension to obtain a second feature vector; The two-dimensional time-domain convolutional neural network is used to extract feature relationships in different channels and time from the second feature vector output by the third convolutional layer; The first parameterized activation layer is used to perform nonlinear transformation on the feature relationships between different channels and time extracted by the two-dimensional time-domain convolutional neural network; The fourth convolutional layer is used to perform a convolution operation on the feature vector output by the first parameterized activation layer in a third dimension to obtain a third feature vector; The fifth convolutional layer is used to perform a convolution operation on the third eigenvector output by the fourth convolutional layer in a fourth dimension to obtain a fourth eigenvector.

6. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 5, characterized in that: The two-dimensional time-domain convolutional neural network specifically includes: a plurality of two-dimensional time-domain convolutional modules connected in sequence; each two-dimensional time-domain convolutional module includes a plurality of two-dimensional convolutional blocks connected in sequence; The two-dimensional time domain convolution module is used to extract feature relationships in different channels and time from the second feature vector output by the third convolution layer, and to perform feature fusion on the feature relationships extracted by multiple two-dimensional convolution blocks connected in sequence in each two-dimensional time domain convolution module.

7. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 6, characterized in that: The two-dimensional convolution block specifically includes: a sixth convolution layer, a second parameterized activation layer, a second normalization layer, a two-dimensional convolution layer, a third parameterized activation layer, a third normalization layer, a seventh convolution layer and an eighth convolution layer connected in sequence; The sixth convolutional layer is used to increase the number of channels of the output feature vector to the number of hidden layers; The second parameterized activation layer is used to perform nonlinear transformation on the feature vector output by the sixth convolutional layer; The second normalization layer is used to perform layer normalization processing on the feature vector output by the second parameterized activation layer; The two-dimensional convolution layer is used to perform two-dimensional deep convolution on the feature vector output by the second normalization layer in the feature and time dimensions to extract features and time-related feature information; The third parameterized activation layer is used to perform nonlinear transformation on the feature information extracted by the two-dimensional convolutional layer; The third normalization layer is used to perform layer normalization processing on the feature vector output by the third parameterized activation layer; The seventh convolutional layer is used to extract the feature relationship between channels from the feature vector output by the third normalization layer, and transmit the feature information to the next connected two-dimensional convolutional block through the residual path; The eighth convolutional layer is used to extract the feature relationship between channels of the feature vector output by the third normalization layer, and perform feature fusion with the feature relationship extracted by other two-dimensional convolution blocks through jump connection.

8. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 5, characterized in that: The decoder specifically includes: a parallel fault classification channel and a fault location regression channel; The fault classification channel is used to determine the predicted probability of each fault category according to the fourth eigenvector output by the fifth convolutional layer to obtain a fault classification result; The fault position regression channel determines the prediction result of the regression target according to the fourth eigenvector output by the fifth convolutional layer to obtain the direction angle of the fault position.

9. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 8, characterized in that: The fault classification channel includes a ninth convolutional layer, a first fully connected layer, an adaptive average pooling layer and a softmax activation layer connected in sequence; the fault location regression channel includes a tenth convolutional layer and a second fully connected layer connected in sequence.

10. The method for aircraft engine fault diagnosis based on microphone array acoustic signal analysis according to claim 1, characterized in that: The total loss function used by the aircraft engine fault diagnosis model is: L MTL =w1L C +w2L MSE ; Among them, L C represents the loss function used in the fault classification task; L MSE represents the loss function of the faulty engine direction task; w m represents the weight of the mth task, m = 1 or 2; L m (t-1), r m (t-1) represents the Loss and training speed of task m at the t-1th step respectively; T represents a constant.

Images