An aircraft engine fault diagnosis method, device, storage medium and equipment

Abstract

The present invention discloses a method, device, storage medium and equipment for fault diagnosis of an aeroengine, belonging to the technical field of aeroengine fault diagnosis. The method includes obtaining vibration signals at a set position on the aeroengine, and obtaining the preprocessed vibration signals after preprocessing; inputting the preprocessed vibration signals into a trained improved binary neural network to obtain the fault diagnosis result at this position; the improvements of the improved binary neural network compared with the binary neural network include: binarizing the weights of all convolutional kernels and fully connected layers; improving the operations performed by the depthwise separable convolution module to: converting the preprocessed vibration signals to the frequency domain and then performing depth convolution, and then converting the output of the depth convolution back to the time domain for pointwise convolution; masking part of the channels of the input of the sparse residual block module through a set sparse mask in each residual block of the sparse residual block module. The present invention realizes efficient diagnosis of faults through the improved model.

Description

Technical Field

[0001] The present invention relates to a method, device, storage medium and equipment for fault diagnosis of an aeroengine, belonging to the technical field of aeroengine fault diagnosis. Background Art

[0002] The main power source of an aircraft is an aeroengine, and its performance and safety directly determine the reliability of flight missions. Due to being in extreme conditions of high temperature, high pressure and high vibration for a long time, key components of an aeroengine are prone to wear, fatigue or failure. In order to ensure flight safety and improve maintenance efficiency, establishing an efficient and accurate fault diagnosis system has become an important direction in aeroengine research.

[0003] Traditional aeroengine fault diagnosis mainly relies on the time-domain feature analysis of vibration signals, such as root mean square value, peak factor, etc., to classify fault types. However, the vibration signals of aeroengines often show strong non-linearity and non-stationarity, and it is difficult for traditional methods to capture subtle differences in fault patterns. And complex data analysis increases the computational burden, especially on resource-constrained edge devices, and the real-time performance is significantly affected.

[0004] In recent years, deep learning techniques, especially convolutional neural networks (CNNs, Convolutional Neural Networks), have made remarkable progress in the field of pattern recognition. CNNs have powerful feature extraction capabilities, providing a new perspective for the analysis of complex signals. And the residual network (ResNet, Residual Network) has become one of the important architectures in deep neural networks since its proposal. Its core feature is to introduce residual blocks, that is, to introduce skip connections in each layer of the network, enabling the network to better transmit gradients during training and solving the problem of gradient disappearance.

[0005] However, although deep neural networks perform very well in many tasks, their computational and storage requirements are extremely high, especially for complex models and large-scale datasets. Therefore, many studies have begun to try to significantly reduce these requirements by binarizing neural network weights and activation values. Binary neural networks have emerged. Binary neural networks limit weights and activation values to two states of -1 and +1 or 0 and +1, thus greatly reducing the parameter representation accuracy and reducing the computational overhead.

[0006] However, the advantages of binary neural networks are not obvious on small data sets. Binary neural networks usually require a large amount of training data to fully play their role. When the amount of data is small, the binary network cannot effectively learn enough features when the amount of data is insufficient, resulting in insignificant efficiency improvement during training. On the contrary, due to its special training method, the computational burden is increased, and the speed of model training and reasoning may even be slower than that of traditional networks.

[0007] In summary, the prior art still has the problems of low fault diagnosis efficiency and heavy computational burden when the amount of data is small. Summary of the invention

[0008] The purpose of the present invention is to provide an aircraft engine fault diagnosis method, device, storage medium and equipment to solve the problems of low diagnostic efficiency and heavy calculation burden in the prior art.

[0009] To achieve the above objectives, the present invention is implemented by adopting the following technical solutions:

[0010] In a first aspect, the present invention provides an aircraft engine fault diagnosis method, comprising:

[0011] Acquire a vibration signal of a set position on the aircraft engine, and obtain a preprocessed vibration signal after preprocessing;

[0012] Inputting the preprocessed vibration signal into the trained improved binary neural network to obtain the fault diagnosis result of the position;

[0013] Among them, the improvements of the improved binary neural network relative to the binary neural network include: binarizing the weights of all convolution kernels and fully connected layers; improving the operation performed by the depthwise separable convolution module to: converting the preprocessed vibration signal into the frequency domain and then performing deep convolution, and then converting the output of the deep convolution back into the time domain for point-by-point convolution; and masking part of the input channels of the sparse residual block module by a set sparse mask in each residual block of the sparse residual block module.

[0014] Further, the preprocessing includes spectral kurtosis data enhancement;

[0015] The spectral kurtosis data enhancement comprises the following steps:

[0016] The spectral kurtosis is calculated by the following formula:

[0017] ;

[0018] ;

[0019] ;

[0020] Among them, is the vibration signal, is the window function, τ represents the time integration variable, f is the frequency variable, t is the time variable, and j represents the imaginary unit. represents the power spectral density at time t and frequency f. is the mean value of the power spectral density, and E[ ] represents the calculation of the expected value. represents the spectral kurtosis.

[0021] According to the spectral kurtosis, the vibration signal is enhanced by the following formula:

[0022] ;

[0023] Among them, is the i-th sample, and each sample consists of multiple vibration signals. is the enhanced i-th sample. is the perturbation coefficient.

[0024] The vibration signal and the enhanced vibration signal are combined to obtain the preprocessed vibration signal.

[0025] Furthermore, the weights of all convolutional kernels and fully connected layers are binarized by the following formula:

[0026] ;

[0027] Among them, represents binarizing the elements in the weight to 1 or -1, and W b represents the binarized weight.

[0028] Furthermore, after the preprocessed vibration signal is transformed into the frequency domain, deep convolution is performed, and then the output of the deep convolution is transformed back into the time domain for pointwise convolution by the following formula:

[0029] ;

[0030] ;

[0031] ;

[0032] Among them, represents the element-wise product in the frequency domain. represents the output of the Fourier transform of the preprocessed vibration signal. represents the output of the Fourier transform of the binarized deep convolution kernel ; represents and The result of performing element-wise multiplication, y d represents the output of depthwise convolution in the time domain, represents the inverse Fourier transform, e represents the natural exponential, and j represents the imaginary unit, is the pointwise convolution kernel after binarization processing, is the output of pointwise convolution in the time domain.

[0033] Furthermore, the expression of the sparse residual block module is:

[0034] ;

[0035] where y m represents the output of the sparse residual block module, represents the depthwise separable convolution layer, and y sparse represents the input of the sparse residual block module with a sparse mask added.

[0036] Furthermore, the sparse mask is added to the input of the sparse residual block module through the following formula:

[0037] ;

[0038] where is a sparse mask, is the input of the sparse residual block module, and the sparse mask is a matrix containing 0s and 1s. When a certain element in is 0, the corresponding channel in will be masked, and the elements in the sparse mask correspond one-to-one with the channels of

[0039] Furthermore, after the improved binary neural network is trained, it is also verified by the accuracy rate. If the accuracy rate is greater than the preset value, it means that the improved binary neural network has been trained well; otherwise, the improved binary neural network is retrained until the accuracy rate is greater than the preset value;

[0040] The accuracy rate is calculated through the following formula:

[0041] ;

[0042] where is the probability of the set true label for the th class, is the probability predicted by the model for the th class, is the number of classes, is the indicator function, which takes the value of 1 when is equal to and 0 otherwise.

[0043] In a second aspect, the present invention provides an aeroengine fault diagnosis device, comprising:

[0044] A vibration signal acquisition module, configured to: acquire vibration signals at a set position on the aeroengine, and obtain preprocessed vibration signals after preprocessing;

[0045] A fault diagnosis module, configured to: input the preprocessed vibration signals into a trained improved binary neural network to obtain a fault diagnosis result at the said position;

[0046] Wherein, the improvements of the improved binary neural network compared with the binary neural network include: binarizing the weights of all convolutional kernels and fully connected layers; improving the operations performed by the depthwise separable convolution module to: convert the preprocessed vibration signals to the frequency domain and then perform depth convolution, and then convert the output of the depth convolution back to the time domain for pointwise convolution; masking some channels of the input of the sparse residual block module in each residual block of the sparse residual block module through a set sparse mask.

[0047] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program / instructions are stored, and when the computer program / instructions are executed by a processor, the steps of the aeroengine fault diagnosis method described in any item of the first aspect are implemented.

[0048] In a fourth aspect, the present invention provides a computer system, comprising:

[0049] A memory, for storing computer programs / instructions;

[0050] A processor, for executing the computer programs / instructions to implement the steps of the aeroengine fault diagnosis method described in any item of the first aspect.

[0051] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0052] An aeroengine fault diagnosis method, device, storage medium and equipment provided by the present invention, through the improvement of the binary neural network, that is, binarizing the weights of all convolutional kernels and fully connected layers, reduces the computational complexity and data storage, reduces the amount of calculation, optimizes the training time by improving the operations in the depthwise separable convolution module to first convert to the frequency domain and then perform convolution, improves the training efficiency and the accuracy of the model, and reduces the data storage by introducing the processing of the sparse mask in the residual block, realizing the efficient processing of vibration signals and the efficient diagnosis of whether there is a fault.

[0053] Meanwhile, for the environment of small datasets, the data augmentation strategy effectively solves the problems of heavy computational burden and insufficient performance of binary neural networks when the amount of data is small, and further enhances the generalization ability of the model.

[0054] Therefore, the present invention can not only quickly and accurately identify the types of aero-engine faults, but also achieve efficient and instant fault diagnosis in the edge computing environment with limited resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flowchart of a method for diagnosing aero-engine faults provided by an embodiment of the present invention;

[0056] Figure 2 is a schematic structural diagram of a sparse residual block provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the protection scope of the present invention.

[0058] Embodiment 1, as Figure 1 shown, this embodiment provides a method for diagnosing aero-engine faults, including:

[0059] Obtain the vibration signal at a set position on the aero-engine, and obtain the preprocessed vibration signal after preprocessing;

[0060] Input the preprocessed vibration signal into the trained improved binary neural network to obtain the fault diagnosis result of the position;

[0061] Among them, the improvements of the improved binary neural network compared with the binary neural network include: binarizing the weights of all convolutional kernels and fully connected layers; improving the operations performed by the depthwise separable convolution module to: converting the preprocessed vibration signal to the frequency domain and then performing depthwise convolution, and then converting the output of the depthwise convolution back to the time domain for pointwise convolution; masking some channels of the input of the sparse residual block module through a set sparse mask in each residual block of the sparse residual block module.

[0062] By improving the binary neural network of the present invention, that is, binarizing the weights of all convolutional kernels and fully connected layers, the computational complexity and data storage are reduced, and the amount of calculation is reduced. By improving the operations in the depthwise separable convolution module to first convert to the frequency domain and then perform convolution, the training time is optimized, and the training efficiency and the accuracy of the model are improved. By introducing the processing of the sparse mask in the residual block, the data storage is reduced, and the efficient processing of the vibration signal and the efficient diagnosis of whether there is a fault are realized.

[0063] Example 2 will further illustrate the implementation of the present invention in conjunction with the accompanying drawings. Figure 1 It is a network flowchart of a fault diagnosis method for an aeroengine provided by the present invention. As Figure 1 shown, this method includes the following seven steps. The improved binary neural network adopted is specifically a binary sparse residual network, and the specific process is as follows.

[0064] The first step: Vibration signal input and preprocessing.

[0065] Obtain the vibration signal of the aeroengine, denoted as X, , representing a space of dimension A×N, where is the number of samples of the obtained vibration signal, is the number of data points of each sample, and N represents the sum of the number of data points and the number of labels of each sample.

[0066] Construct N - 1 vibration signals into a sample, and put the true label of this sample at the last position of this sample to obtain a complete sample; the i-th sample is denoted as x i , , where x i1 represents the first vibration signal in the i-th sample, x i2 represents the second vibration signal in the i-th sample, x i3 represents the third vibration signal in the i-th sample, x iN-1 represents the (N - 1)-th vibration signal in the i-th sample, and the m-th vibration signal in the i-th sample can be represented by the general formula x im , m = 1, 2, 3, …, N - 1, , represents the true label of the i-th sample.

[0067] To solve the problem of the insufficient performance of the binary neural network when the amount of data is small, data augmentation is performed in the first step. For this, a spectral kurtosis data augmentation method based on the short-time Fourier transform (STFT) is used. Spectral kurtosis is a quantitative statistic of the sharpness of the signal spectrum, and the calculation method is as follows:

[0068] First, calculate the spectrum of the signal through the short-time Fourier transform STFT. Given the vibration signal , the spectrum can be expressed as:

[0069] ;

[0070] Among them, is the window function, which is used to control the duration of each Fourier transform, and τ represents the time integration variable. is a frequency variable, is a time variable, and j represents the imaginary unit.

[0071] Next, calculate the power spectral density (PSD), which is the square of the result of the STFT calculation (the spectrum of the signal):

[0072] ;

[0073] where represents the power spectral density at time t and frequency f.

[0074] Finally, calculate the spectral kurtosis , and the spectral kurtosis is calculated by the following formula:

[0075] ;

[0076] where represents the spectral kurtosis, is the mean value of the power spectral density, and E[ ] represents the calculation of the expected value.

[0077] For each sample, set multiple perturbation coefficients (such as 0.05, 0.1, 0.15, etc.) for enhancement processing. The specific method is:

[0078] ;

[0079] where is the i-th sample, is the i-th sample after enhancement, is the perturbation coefficient, is the spectral kurtosis of the signal.

[0080] Subsequently, merge the original signal and the enhanced signal to obtain the preprocessed vibration signal as , and this signal will be transmitted to the subsequent convolutional module.

[0081] A high spectral kurtosis value indicates that the spectrum is more concentrated and the signal is sharper; a low spectral kurtosis value indicates that the signal is relatively smoother. Through this method, data enhancement of the signal can be performed, increasing the diversity of training data and solving the problem of insufficient performance of binary neural networks when the amount of data is small.

[0082] Step 2: Binarization processing of the convolutional kernel and the fully connected layer.

[0083] To improve the efficiency of convolutional operations, the weight binarization technique is introduced. The weights of the convolutional kernel and the fully connected layer in the binary neural network will undergo binarization processing, restricting their values to +1 and -1, thereby reducing the computational complexity and data storage.

[0084] The preprocessed vibration signal is input into a binary neural network. Before entering the convolution operation, the binary neural network binarizes the weights of all convolutional kernels and fully connected layers, that is, converts the original weight W into a binary weight. , representing the dimension of the space, where is the number of input channels of the convolutional kernel, is the number of output channels of the convolutional kernel, is the size of the convolutional kernel.

[0085] The original weight W is converted into a binary weight through the following formula:

[0086] ;

[0087] where, represents binarizing the elements in the weight to ±1, which significantly reduces the computational amount, and W b represents the binary weight.

[0088] Step 3, the processing of the signal by the depthwise separable frequency-domain convolution module.

[0089] The preprocessed vibration signal is subjected to feature extraction by the depthwise separable frequency-domain convolution module. The depthwise separable convolution is achieved through two steps: depthwise convolution and pointwise convolution. The depthwise convolution independently performs convolution operations for each input channel, while the pointwise convolution uses a 1×1 convolutional kernel to integrate the output results of the depthwise convolution into new output channels. To further improve the efficiency of the convolution operation, frequency-domain convolution is added on the basis of the depthwise separable convolution to obtain the depthwise separable frequency-domain convolution module. In the depthwise separable frequency-domain convolution module, the input signal and the convolutional kernel are first transformed to the frequency domain through the Fourier transform, depthwise convolution is performed in the frequency domain, then restored to the time domain through the inverse Fourier transform, and finally pointwise convolution is performed.

[0090] Let the output of the preprocessed vibration signal after the Fourier transform be , and the output of the depthwise convolutional kernel after binarization and the Fourier transform be , , represents the dimension of the space, then the convolution operation can be represented in the frequency domain as:

[0091] ;

[0092] where, represents the element-wise product in the frequency domain, represents the output of the convolution operation in the frequency domain, and is also and the result of element-wise multiplication.

[0093] The final output of the depth convolution in the time domain is:

[0094] ;

[0095] where y d represents the output of the depth convolution in the time domain, represents the inverse Fourier transform, e represents the natural exponential, and j represents the imaginary unit;

[0096] The output of the pointwise convolution in the time domain is:

[0097] ;

[0098] where is the binarized pointwise convolution kernel, , represents the space of dimension , is the output of the pointwise convolution in the time domain.

[0099] The frequency domain convolution result is transformed back to the time domain signal through the inverse Fourier transform. Through frequency domain convolution, the frequency domain information of the signal can be more effectively extracted, thereby enhancing the fault diagnosis ability of the network.

[0100] Step 4. Processing of the signal by the sparse residual block module.

[0101] The output of the depthwise separable frequency domain convolution module passes through the activation function and the max pooling layer to obtain the input of the sparse residual block module , and is input into the sparse residual block module. In the design of this module, a static sparse mask is introduced to improve the calculation efficiency. Specifically, in each residual block, a part of the channels of is randomly masked through a custom sparse mask, thereby reducing the calculation amount. The structure of the obtained sparse residual block is as Figure 2 shown.

[0102] The operation process of each residual block in the sparse residual block module can be expressed as:

[0103] ;

[0104] where represents the depthwise separable convolution layer, is the input of the sparse residual block module, is the output of the residual block after residual connection.

[0105] To introduce sparsity, a sparse mask is applied to the input of the sparse residual block module and its formula is: ;

[0106] ;

[0107] where y sparse represents the input of the sparse residual block module with the sparse mask added, is a sparse mask, which is a matrix containing 0s and 1s used to mask a part of the channels. When a certain element in is 0, the corresponding channel in will be masked, thus reducing the computational complexity. The elements in the sparse mask correspond one-to-one with the channels of

[0108] Finally, the expression of the sparse residual block module is:

[0109] ;

[0110] where y m represents the output of the sparse residual block module, represents the depthwise separable convolutional layer.

[0111] Step 5, Fully Connected Layer and Classification.

[0112] The output of the sparse residual block module will be dimensionally reduced through the global average pooling layer to extract global features. The features after the global average pooling layer are input into a binary fully connected layer for final fault classification. Through the softmax function, the network outputs the probability of each fault category, and finally obtains the model prediction value .

[0113] The overall process of the network is as follows: The vibration signal after preprocessing, obtains the preprocessed vibration signal , and the preprocessed signal enters the depthwise separable frequency domain convolution module, and after depthwise frequency domain convolution and pointwise convolution processing, obtains , and after passing through the activation function and the max pooling layer, obtains . Input into the sparse residual block module, mask a part of the channels through the sparse mask, and obtain the final output through the residual connection. Finally, after passing through the fully connected layer and classification processing, the network outputs the fault category label to achieve fault diagnosis.

[0114] Step 6, Training the Network.

[0115] During the training phase, the AdamW (Adam with Weight Decay Fix) optimizer is adopted, and the CrossEntropyLoss function is used to optimize the network performance. The CrossEntropyLoss function is used for classification tasks and is defined as:

[0116] ;

[0117] where L is the value of the CrossEntropyLoss function, is the probability of the set true label for the th class, is the probability predicted by the model for the th class, is the number of classes.

[0118] The gradients are calculated through backpropagation and the network parameters are updated, ultimately minimizing the loss function L and improving the overall performance of the network.

[0119] Step 7: Fault diagnosis.

[0120] After the model training is completed, the model is evaluated using the test set. The test data is input into the network to obtain the prediction results, which are compared with the true labels, and the accuracy is calculated accordingly.

[0121] The formula for calculating the accuracy is:

[0122] ;

[0123] where is the indicator function, which takes the value of 1 when the predicted value is equal to the true label , and 0 otherwise.

[0124] Through this process, the model can quickly and accurately diagnose faults in different situations and provide a basis for equipment maintenance.

[0125] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows 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 the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0127] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0128] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operating steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or multiple flows and / or blocks Figure 1 or means for implementing the functions specified in one block or multiple blocks.

[0129] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for diagnosing aircraft engine faults, characterized in that: include: Acquire a vibration signal of a set position on the aircraft engine, and obtain a preprocessed vibration signal after preprocessing; Inputting the preprocessed vibration signal into the trained improved binary neural network to obtain the fault diagnosis result of the position; The improvements of the improved binary neural network over the binary neural network include: binarizing the weights of all convolution kernels and fully connected layers; improving the operation performed by the depthwise separable convolution module to: converting the preprocessed vibration signal into the frequency domain and then performing a depthwise convolution, and then converting the output of the depthwise convolution back into the time domain for pointwise convolution; masking part of the input channels of the sparse residual block module by a set sparse mask in each residual block of the sparse residual block module; The preprocessing includes spectral kurtosis data enhancement; The spectral kurtosis data enhancement comprises the following steps: The spectral kurtosis is calculated by the following formula: ； ； ； in, It is a vibration signal. is the window function, τ represents the time integral variable, f is the frequency variable, t is the time variable, j represents the imaginary unit, represents the power spectral density at time t and frequency f, is the mean of the power spectral density, E[ ] represents the expected value calculation, represents the spectral kurtosis; According to the spectral kurtosis, the vibration signal is enhanced by the following formula: ； in, is the i-th sample, each sample consists of multiple vibration signals, is the i-th sample after enhancement, is the disturbance coefficient; Combining the vibration signal and the enhanced vibration signal to obtain a preprocessed vibration signal; The preprocessed vibration signal is converted into the frequency domain and then deep convolution is performed, and then the output of the deep convolution is converted back to the time domain for point-by-point convolution, which is performed by the following formula: ； ； ； in, represents the element-wise product in the frequency domain, represents the output of the preprocessed vibration signal after Fourier transform, Represents the binary deep convolution kernel The output of the Fourier transform is express and The result of element-wise multiplication, y d represents the output of the deep convolution in the time domain, represents the inverse Fourier transform, e represents the natural exponential, j represents the imaginary unit, is the point-by-point convolution kernel after binarization. is the output of point-wise convolution in the time domain.

2. The method for diagnosing aircraft engine faults according to claim 1, characterized in that: The weights of all convolution kernels and fully connected layers are binarized using the following formula: ； in, Indicates that the weight The elements in are binarized to 1 or -1, W b Represents the binarization weight.

3. The method for diagnosing aircraft engine faults according to claim 1, characterized in that: The expression of the sparse residual block module is: ; Among them, y m represents the output of the sparse residual block module, represents the depthwise separable convolutional layer, y sparse Represents the input of the sparse residual block module with the sparse mask added.

4. The method for diagnosing aircraft engine faults according to claim 3, characterized in that: The sparse mask is added to the input of the sparse residual block module by the following formula: ; in, is a sparse mask, is the input of the sparse residual block module. The sparse mask is a matrix containing 0 and 1. When an element in is 0, The corresponding channels in will be masked, and the elements in the sparse mask and The channels correspond one to one.

5. The method for diagnosing aircraft engine faults according to claim 1, characterized in that: After the training is completed, the improved binary neural network is also verified by the accuracy rate. If the accuracy rate is greater than a preset value, it means that the improved binary neural network has been trained. Otherwise, the improved binary neural network is retrained until the accuracy rate is greater than the preset value. The accuracy is calculated by the following formula: ； in, is the true label set The probability of the class, The model predicts The probability of the class, is the number of categories, is the indicator function, when and The value is 1 if they are equal, otherwise 0.

6. An aircraft engine fault diagnosis device, characterized in that: include: The vibration signal acquisition module is configured to: acquire a vibration signal at a set position on the aircraft engine, and obtain a preprocessed vibration signal after preprocessing; The fault diagnosis module is configured to: input the preprocessed vibration signal into the trained improved binary neural network to obtain a fault diagnosis result of the position; The improvements of the improved binary neural network over the binary neural network include: binarizing the weights of all convolution kernels and fully connected layers; improving the operation performed by the depthwise separable convolution module to: converting the preprocessed vibration signal into the frequency domain and then performing a depthwise convolution, and then converting the output of the depthwise convolution back into the time domain for pointwise convolution; masking part of the input channels of the sparse residual block module by a set sparse mask in each residual block of the sparse residual block module; The preprocessing includes spectral kurtosis data enhancement; The spectral kurtosis data enhancement comprises the following steps: The spectral kurtosis is calculated by the following formula: ； ； ； in, It is a vibration signal. is the window function, τ represents the time integral variable, f is the frequency variable, t is the time variable, j represents the imaginary unit, represents the power spectral density at time t and frequency f, is the mean of the power spectral density, E[ ] represents the expected value calculation, represents the spectral kurtosis; According to the spectral kurtosis, the vibration signal is enhanced by the following formula: ； in, is the i-th sample, each sample consists of multiple vibration signals, is the i-th sample after enhancement, is the disturbance coefficient; Combining the vibration signal and the enhanced vibration signal to obtain a preprocessed vibration signal; The preprocessed vibration signal is converted into the frequency domain and then deep convolution is performed, and then the output of the deep convolution is converted back to the time domain for point-by-point convolution, which is performed by the following formula: ； ； ； in, represents the element-wise product in the frequency domain, represents the output of the preprocessed vibration signal after Fourier transform, Represents the binary deep convolution kernel The output of the Fourier transform is express and The result of element-wise multiplication, y d represents the output of the deep convolution in the time domain, represents the inverse Fourier transform, e represents the natural exponential, j represents the imaginary unit, is the point-by-point convolution kernel after binarization. is the output of point-wise convolution in the time domain.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the aircraft engine fault diagnosis method described in any one of claims 1 to 5 are implemented.

8. A computer system, characterized in that: include: Memory, for storing computer programs / instructions; A processor is used to execute the computer program / instructions to implement the steps of the aircraft engine fault diagnosis method according to any one of claims 1 to 5.

Images