Helicopter transmission part damage intelligent detection method

Abstract

The invention relates to a helicopter transmission part damage intelligent detection method. The method comprises the following steps that stress wave data of a helicopter transmission part are collected and preprocessed; decomposing the preprocessed stress wave data into a plurality of intrinsic mode decomposition signals by adopting an empirical mode decomposition method; constructing a feature selection framework based on a normalized mutual information method, and screening out a set number of optimal intrinsic mode decomposition signals; and constructing an intelligent detection model based on the convolutional deep belief network, and inputting the screened optimal intrinsic mode decomposition signal into the intelligent detection model to obtain a part damage diagnosis result. According to the method, the helicopter transmission part damage classification can be quickly and accurately diagnosed.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent detection of helicopter transmission component damage, and particularly to an intelligent detection method for helicopter transmission component damage. Background Art

[0002] The unique structure of a helicopter determines its characteristics of high maneuverability and high flexibility, making it play a wide range of roles in the military and civilian fields, and its flight reliability is particularly important. The helicopter transmission component is one of the core components and an important factor affecting the healthy operation of the helicopter. Therefore, it is of great significance to perform efficient and accurate intelligent detection on it.

[0003] Among them, the stress wave signal is a new type of signal source for state analysis. When a helicopter is operating, there are relative movements in the gear meshing, bearing rotation, and transmission shaft rotation of the transmission components. During the relative movement of the two components, friction and impact signals will be generated and propagated outward in the form of a high-frequency signal, which is the stress wave. Research shows that the stress wave is a very effective means for the intelligent detection of helicopter transmission component damage. Therefore, researching the intelligent detection method of the damage stress of helicopter transmission components, enabling the helicopter to operate smoothly and safely, and preventing losses caused by accidental failures, has important application prospects.

[0004] For the intelligent detection of the damage stress of helicopter transmission components, there are still certain problems at present. The stress wave is not only affected by the stress wave source, but also affected by the transmission path, and even the two aspects of influence are coupled with each other. When using the stress wave source for intelligent detection and analysis of helicopter transmission component damage stress, there are difficult problems such as multiple coupled interference signals and low detection and diagnosis classification accuracy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent detection method for helicopter transmission component damage stress, which can quickly and accurately diagnose the damage classification of helicopter transmission components.

[0006] The technical solution adopted by the present invention to solve its technical problems is: to provide an intelligent detection method for helicopter transmission component damage, including the following steps:

[0007] Collect stress wave data of helicopter transmission components and perform preprocessing;

[0008] Use the empirical mode decomposition method to decompose the preprocessed stress wave data into multiple intrinsic mode decomposition signals;

[0009] Based on the normalized mutual information method, construct a feature selection framework and screen out a set number of optimal intrinsic mode decomposition signals;

[0010] Construct an intelligent detection model based on a convolutional deep belief network, input the selected optimal intrinsic mode decomposition signals into the intelligent detection model, and obtain the component damage diagnosis results.

[0011] Further, the optimal intrinsic mode decomposition signal is the IMF component with the highest correlation with the preprocessed stress wave data.

[0012] Further, construct a feature selection framework based on the normalized mutual information method to screen out a set number of optimal intrinsic mode decomposition signals, including:

[0013] Calculate the information entropy of the preprocessed stress wave data and the information entropy of each intrinsic mode decomposition signal;

[0014] For any intrinsic mode decomposition signal, calculate its mutual information with the preprocessed stress wave data, and then calculate the NMI value corresponding to this intrinsic mode decomposition signal according to the information entropy of the stress wave data, the information entropy of this intrinsic mode decomposition signal, and their mutual information;

[0015] Screen out the intrinsic mode decomposition signals corresponding to a set number of the highest NMI values as the optimal intrinsic mode decomposition signals.

[0016] Further, the intelligent detection model is stacked by multiple convolutional restricted Boltzmann machines.

[0017] Further, the convolutional restricted Boltzmann machine consists of a visible layer and a hidden layer, and the weights are shared between the visible layer and the hidden layer.

[0018] Further, the intelligent detection model is trained by the following method:

[0019] Collect a stress wave data sample set containing different damage classifications of helicopter transmission components;

[0020] Use the stress wave data sample set to train the bottom convolutional restricted Boltzmann machine;

[0021] Take the output of the lower convolutional restricted Boltzmann machine as the input of this layer's convolutional restricted Boltzmann machine and train this layer's convolutional restricted Boltzmann machine;

[0022] Repeat the above step until all convolutional restricted Boltzmann machines are trained;

[0023] Use the stochastic gradient descent algorithm and backpropagation to fine-tune the weights of each layer of convolutional restricted Boltzmann machine;

[0024] Add a Softmax function at the top layer as the activation function to output the classification recognition label.

[0025] Further, the method of using empirical mode decomposition to decompose the preprocessed stress wave data into multiple intrinsic mode decomposition signals includes:

[0026] S300 uses the processed stress wave data as the original signal;

[0027] S301 fits the upper envelope line and the lower envelope line based on all the local maximum points and local minimum points of the current original signal, and calculates the local average values of the upper envelope line and the lower envelope line respectively;

[0028] S302 calculates the difference between the current original signal and the local average value to obtain the original data information;

[0029] S303 If the original data information does not meet the conditions for intrinsic mode decomposition, update the value of the current original signal to the original data information and return to step S301, otherwise output the original data information as the current order intrinsic mode decomposition signal;

[0030] S304 removes the current order intrinsic mode decomposition signal from the current original signal, and determines whether the obtained residual component is a monotonic function. If so, stop the decomposition to obtain multiple different order intrinsic mode decomposition signals, otherwise update the value of the current original signal to the residual component and return to step S301.

[0031] Further, the preprocessing is realized by using the Min - Max normalization method to linearly map the collected stress wave data to the range of [0, 1].

[0032] Further, the stress wave data is collected by stress wave sensors arranged on the surface of or near the surface of helicopter transmission components.

[0033] Beneficial Effects

[0034] Due to the adoption of the above - mentioned technical solutions, compared with the prior art, the present invention has the following advantages and positive effects: The present invention uses the empirical mode decomposition method and the normalized mutual information feature selection method to decompose and screen the original stress wave signal, select the best decomposition signal, reduce the data dimension, improve the model calculation efficiency, and solve the problem of the existence of various coupled interference signals in the original stress wave signal; The present invention constructs an intelligent detection model based on the convolutional deep belief network, which can fully extract the feature information in the signal, thus effectively solving the problem of low detection and diagnosis classification accuracy, and has characteristics such as translational invariance, weight sharing, and simple model training. Description of the Drawings

[0035] Figure 1 is the flowchart of the embodiment of the present invention;

[0036] Figure 2 It is a schematic diagram of the CRBM structure of an embodiment of the present invention. Specific embodiments

[0037] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0038] An embodiment of the present invention relates to an intelligent detection method for the damage stress of helicopter transmission components based on a convolutional deep belief network, as Figure 1 shown, including the following steps:

[0039] S1, collect the original stress wave signals of helicopter transmission components;

[0040] S2, preprocess the collected stress wave data;

[0041] S3, decompose the preprocessed stress wave data based on empirical mode decomposition (EMD) to obtain multiple intrinsic mode functions (IMFs);

[0042] S4, construct a feature selection framework in combination with normalized mutual information (NMI) to perform correlation screening on the obtained IMFs, select the best IMF signal, reduce the data dimension, and improve the model calculation efficiency;

[0043] S5, build a deep learning model based on a convolutional deep belief network (CDBN), utilize its strong feature extraction ability, input the selected best decomposition signal, output the intelligent detection fault diagnosis result, and realize the intelligent detection fault diagnosis of helicopter transmission components.

[0044] Among them, the stress wave data can be collected by a stress wave sensor. The stress wave sensor is installed on the surface near the helicopter transmission component. The piezoelectric crystal in the sensor converts the stress wave amplitude into an electrical signal, and then it is amplified and filtered by a high-frequency band-pass filter in the analog signal modulator, and a stress wave pulse train (SWPT) is output. The stress wave data can effectively represent the impact and mechanical friction conditions of the helicopter transmission component.

[0045] The original stress wave data of the helicopter transmission components collected need to be normalized, and the data is scaled to a specific range to eliminate the dimensional differences between features. In this embodiment, the Min-Max normalization method is adopted to linearly map the data to the range of [0, 1].

[0046] EMD can divide a certain complex information into the sum of several intrinsic mode functions (IMFs), thus completing the transformation from non-smooth information to smooth information. The EMD method mainly analyzes the signal using the information characteristics of the information itself. The number of IMF components obtained is very small, and all IMF components show the actual physical data characteristics contained in the information. In this embodiment, the following EMD decomposition process is adopted:

[0047] S301, Fit the upper and lower envelope lines according to all the local maximum and minimum points of the signal x(t). According to the upper and lower envelope lines, the average value of the upper and lower envelope lines of x(t) is m 1 , and the formula is as follows:

[0048]

[0049] S302, Taking h 1 as the original data information, the difference between the original signal and the local average value is:

[0050] x(t) - m 1 = h 1

[0051] S303, If h 1 does not meet the conditions of IMF, then use h 1 to replace x(t), and repeat the processes of step S301 and step S302; repeat the above steps k times until h 1(k-1) - m 1k = h 1k is obtained, so that for h 1k , it meets the conditions of IMF. At this time, let c 1 = h 1k , then c 1 is the first-order IMF.

[0052] S304, Replace the signal x(t) with the new signal r n obtained after removing the high-frequency component c n from the signal x(t), and repeat the processes of step S31, step S32, and step S33 to obtain a series of IMFs until the residual component r n is a monotonic function. At this time, the EMD decomposition ends, and the original signal is expressed as:

[0053]

[0054] where r n represents the average trend of the signal, also known as the residual function.

[0055] The NMI algorithm is a method that uses information theory to quantify the degree of correlation and dependence between variables, can effectively evaluate the similarity between information, and also has a certain robustness to noise and outliers, and has good applicability to the stress wave of helicopter transmission components with strong interference signals.

[0056] In this embodiment, by comparing the correlation between each decomposed IMF and the original data, the optimal IMF signal is selected as the one with the highest NMI value with the original data, that is, the optimal IMF component with the highest correlation. The specific method is as follows:

[0057] The NMI value of random variables X and Y is denoted as I NM (X; Y), and the operation expression is:

[0058]

[0059] where: H(.) is the information entropy; H(X) = I M (X; X); H(Y) = I M (Y; Y);

[0060] where, I M (X; Y) is the mutual information of variables X and Y, and its operation expression is:

[0061]

[0062] P MA (X) and P MA (Y) are the marginal probability distributions of variables X and Y respectively; P JO (X; Y) is the joint probability distribution of variables X and Y;

[0063] If I M (X; Y) is 0, it means that the random variables X and Y are independent; the larger I NM (X; Y) is, the more common information there is between the random variables X and Y, and the higher the degree of mutual dependence.

[0064] CDBN is composed of stacked Convolution Restricted Boltzmann Machines (CRBMs). A CRBM consists of a visible layer V and a hidden layer H, and the weights are shared between the visible layer V and the hidden layer H.

[0065] As Figure 2 shown, in the CRBM, V consists of 1 matrix of size N V ×N VIt consists of a binary unit matrix. H is composed of U matrices H of size N H ×N H (u = 1, 2, …, U), and is simultaneously connected to a filter matrix W of size u (u = 1, 2, …, U), and is simultaneously connected to a filter matrix W of size . The energy function of the CRBM can be expressed as: u . The energy function of the CRBM can be expressed as:

[0066]

[0067] Among them, (a, t = 1, 2, …, N H ) are neurons in H u ; v at are neurons in V; (q, r = 1, 2, …, N W ) are neurons in W u ; b u is the bias of H u , and c is the shared bias of V.

[0068] The conditional probability of the CRBM can be calculated by one-step Gibbs sampling and can be expressed as:

[0069]

[0070] Among them, σ(x) = 1 / (1 + e -x ) represents the sigmoid function; * represents the convolution operation.

[0071] The CDBN model can be trained by the following method:

[0072] S501, The output of the first-layer CRBM will be used as the input of the second-layer CRBM. At the same time, "freeze" the weights of the first-layer CRBM and train the second-layer CRBM, and so on to complete the subsequent training.

[0073] S502, After pre-training the weights of each layer of CRBM, fine-tune the weights through the stochastic gradient descent algorithm and backpropagation to achieve a better fitting effect.

[0074] S503, Finally, the top layer uses the Softmax function as the activation function to output the label of classification recognition.

[0075] After building the CDBN model based on the above method, input the training set data for model training. After the model training is completed, input the test set into the trained model and output the final fault diagnosis to realize the intelligent detection of the damage stress of the helicopter transmission components.

[0076] The following will further illustrate this embodiment in combination with specific embodiments.

[0077] In this embodiment, the sampling frequency of the gear stress wave data is 800,000 Hz, and the acquisition time is 6 s. The collected stress wave data set contains a total of 4 types of faults and 1 type of normal signal, a total of 5 types. Taking 1200 data points as a sample, there are 4000 experimental samples for each type.

[0078] First, for the original gear stress wave data collected, the Min-Max normalization method is used to linearly map the data to the range of [0, 1].

[0079] Then, the EMD decomposition method is used to decompose the gear stress wave data, and all types of stress wave signals are decomposed by EMD. Taking the data of the broken tooth fault part as an example, EMD decomposes this signal into 5 IMFs and 1 residual component; the same EMD is used for signal decomposition of other types of data such as missing teeth, wear, pitting, and normal.

[0080] After EMD decomposition, each IMF and the participating components are transformed into the frequency domain. Compared with the original signal, EMD decomposes the signal into different frequency scales under a certain bandwidth, extracts more detailed useful information and features in the signal, and reduces the complexity and non-stationarity of the signal.

[0081] Secondly, the NMI method is used to screen each decomposed IMF. Taking the broken tooth fault data as an example, the correlation with the original data is calculated for the 5 decomposed IMFs, so as to judge and obtain the best correlated IMF signal, and redundant data is removed to obtain the best state. The specific correlation calculation results are as follows:

[0082]

[0083] As can be seen from the above table, for the broken tooth fault data, IMF2 is the best modal component selected, so this data is selected as the input of the subsequent model. The above method is used to screen the best modal components for other types of data such as missing teeth, wear, pitting, and normal.

[0084] Finally, according to the above decomposition and screening results, the input size of the CDBN model is designed to be 1x1200. To obtain a rich receptive field, two hidden layers are designed to extract the features in the input IMF. The number of neurons in the output layer of the network is set to 5, and the ONE-HOT method is used to set the value corresponding to the fault category serial number to 1, and the rest are set to 0.

[0085] The constructed CDBN deep learning model has good feature extraction and learning capabilities. Its neurons can effectively learn the differences between various data, effectively improving the classification accuracy. Through testing, the classification accuracy of gear stress wave data is 95.82%, which can effectively classify various types of faults and has a good diagnostic effect.

Claims

1. An intelligent detection method for damage to helicopter transmission components, characterized in that: The following steps are involved: Collect stress wave data of helicopter transmission components and perform preprocessing; The preprocessed stress wave data is decomposed into multiple intrinsic mode decomposition signals using the empirical mode decomposition method; A feature selection framework is constructed based on the normalized mutual information method to screen out a set number of optimal intrinsic mode decomposition signals. An intelligent detection model based on a convolutional deep belief network is constructed, and the screened optimal intrinsic mode decomposition signals are input into the intelligent detection model to obtain component damage diagnosis results.

2. The method according to claim 1, characterized in that: The optimal intrinsic mode decomposition signal is the IMF component with the highest correlation with the preprocessed stress wave data.

3. The method according to claim 2, characterized in that The feature selection framework is constructed based on the normalized mutual information method to screen out a set number of optimal intrinsic mode decomposition signals, including: Calculate the information entropy of the preprocessed stress wave data and the information entropy of each natural mode decomposition signal; For any inherent modal decomposition signal, the mutual information between it and the preprocessed stress wave data is calculated, and then the NMI value corresponding to the inherent modal decomposition signal is calculated according to the information entropy of the stress wave data, the information entropy of the inherent modal decomposition signal and the mutual information between it and the preprocessed stress wave data; The intrinsic mode decomposition signals corresponding to a set number of highest NMI values ​​are screened out as the optimal intrinsic mode decomposition signals.

4. The method according to claim 1, characterized in that The intelligent detection model is composed of a stack of multiple layers of convolutional restricted Boltzmann machines.

5. The method according to claim 4, characterized in that The convolution restricted Boltzmann machine consists of a visible layer and a hidden layer, and weights are shared between the visible layer and the hidden layer.

6. The method according to claim 4, characterized in that The intelligent detection model is trained by the following method: Collect a sample set of stress wave data containing damage classifications of different helicopter transmission components; The underlying convolutional restricted Boltzmann machine is trained using the stress wave data sample set; The output of the convolution restricted Boltzmann machine in the lower layer is used as the input of the convolution restricted Boltzmann machine in this layer, and the convolution restricted Boltzmann machine in this layer is trained; Repeat the previous step until all convolutional restricted Boltzmann machines are trained; The weights of each layer of convolutional restricted Boltzmann machine are fine-tuned using stochastic gradient descent algorithm and back propagation; Add a Softmax function as an activation function at the top layer to output the classification recognition label.

7. The method according to claim 1, characterized in that The empirical mode decomposition method is used to decompose the preprocessed stress wave data into multiple intrinsic mode decomposition signals, including: S300 uses the processed stress wave data as the original signal; S301: fitting an upper envelope and a lower envelope based on all local maximum points and local minimum points of the current original signal, and calculating local average values ​​of the upper envelope and the lower envelope respectively; S302 calculates the difference between the current original signal and the local average value to obtain original data information; S303: if the original data information does not meet the condition of the intrinsic mode decomposition, then the value of the current original signal is updated to the original data information, and the process returns to step S301; otherwise, the original data information is output as the current-order intrinsic mode decomposition signal; S304 removes the current order intrinsic modal decomposition signal from the current original signal, determines whether the obtained residual component is a monotonic function, and if so, stops decomposition to obtain multiple intrinsic modal decomposition signals of different orders; otherwise, updates the value of the current original signal to the residual component and returns to step S301.

8. The method according to claim 1, characterized in that The preprocessing is achieved by linearly mapping the collected stress wave data into the range of [0, 1] using the Min-Max normalization method.

9. The method according to claim 1, characterized in that: The stress wave data is collected by stress wave sensors arranged on the surface of helicopter transmission components or on the surface near the components.

Images