Multi-task learning fault diagnosis method for intermediate bearing of aero-engine

Abstract

The invention discloses a multi-task learning fault diagnosis method for an aero-engine intermediate bearing, which relates to the technical field of aero-engine fault diagnosis, and comprises the following steps: vibration signal data acquisition and sample construction: acquiring an acceleration vibration signal of the aero-engine intermediate bearing, recording the rotating speed of an engine and the position information of a sensor, and constructing a sample; repeated sampling is carried out through a sliding window, and a fault sample set is constructed; data preprocessing: performing normalization processing on the sample data; designing a multi-task learning model: constructing a deep neural network model based on multi-task learning; optimizing a joint loss function: designing the joint loss function; and model training and testing: inputting the processed sample data into a multi-task learning model for training and testing. The multi-task learning fault diagnosis method for the aero-engine intermediate bearing has high diagnosis accuracy and robustness, and is especially suitable for a fault detection scene of the aero-engine bearing.

Description

Technical Field

[0001] The present invention relates to the technical field of aeroengine fault diagnosis, and more particularly to a multi-task learning fault diagnosis method for an aeroengine intermediate bearing. Background Art

[0002] An aeroengine is the core power device of an aircraft, and its safe operation is directly related to the reliable completion of flight missions and the safety of personnel. In an aeroengine, the intermediate bearing, as a key supporting component of the rotor system, its health state is crucial for the stability and reliability of the engine. However, due to the intermediate bearing being in a harsh environment such as high temperature, high pressure, and high noise for a long time, its fault signals often appear weak, non-linear, and are easily masked by the vibration signals of other components. Such complex signal characteristics make it difficult to extract fault features, posing a great challenge to the fault diagnosis of the intermediate bearing.

[0003] Existing fault diagnosis technologies mainly include methods based on signal processing and machine learning. Traditional signal processing methods, such as Fourier transform, wavelet analysis, and empirical mode decomposition, rely on artificially designed features to extract the fault feature frequencies in signals. These methods can achieve certain effects under simple working conditions, but in the face of high noise and complex vibration environments, it is often difficult to accurately extract key features, and they rely heavily on expert experience. With the development of deep learning technology, machine learning, especially deep models such as convolutional neural networks (CNNs), has gradually been applied to the field of vibration signal analysis. Such methods significantly improve the accuracy of fault diagnosis by automatically extracting high-level features. However, existing deep learning methods usually only focus on a single task (such as fault classification) and fail to fully utilize other relevant information (such as rotational speed or sensor position information), resulting in insufficient adaptability and generalization ability of the model under complex working conditions. In addition, deep learning models usually require a large amount of labeled data for training, and the cost of obtaining fault data for aeroengines is extremely high, and the actual available fault samples are very scarce, which further limits the application of traditional deep learning models in the aviation field.

[0004] The operating conditions of aeroengines are complex and variable, including frequent changes in rotational speed and the problem of multi-sensor signal fusion. For example, under different rotational speed states, the vibration signal characteristics of the intermediate bearing will show significant differences, and models that simply rely on vibration signals for analysis often have difficulty adapting to such changes. In addition, the installation position of the sensor will also significantly affect the characteristics of the collected vibration signals, and signals from different positions may contain different fault information. If the sensor position information is not considered, it may lead to insufficient learning of fault features by the model and even misdiagnosis. Therefore, how to make full use of multi-source information (such as rotational speed and sensor position) to improve the fault diagnosis ability of the model under limited sample conditions has become an urgent problem to be solved in the current field of aeroengine intermediate bearing fault diagnosis. Summary of the Invention

[0005] The object of the present invention is to provide a multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine, so as to solve the problems existing in the existing fault diagnosis technology of the intermediate bearing of the aero-engine, such as difficult signal feature extraction, scarce sample data, and poor adaptability to complex working conditions.

[0006] To achieve the above object, the present invention provides a multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine, including the following steps:

[0007] Step 1: Vibration signal data acquisition and sample construction: Collect the acceleration vibration signal of the intermediate bearing of the aero-engine through a data acquisition system, and at the same time record the corresponding engine speed and sensor position information. Repeat sampling of the original acceleration vibration signal through a sliding window to achieve data augmentation and construct a fault sample set;

[0008] Step 2: Data preprocessing: Normalize the sample data in the generated fault sample set to eliminate the influence of signal amplitude and dimension differences between different sample data on model training;

[0009] Step 3: Multi-task learning model design: Construct a deep neural network model based on multi-task learning;

[0010] Step 4: Joint loss function optimization: Design a joint loss function to balance the learning objectives of the main task and the auxiliary task during the optimization process;

[0011] Step 5: Model training and testing: Input the processed sample data into the multi-task learning model for training and testing.

[0012] Preferably, in step 1, the data acquisition system collects the acceleration vibration signals in the normal state and typical fault modes, and the engine speed ranges from low speed to high speed;

[0013] The type labels of the acceleration vibration signals include normal, inner ring fault, and outer ring fault.

[0014] Preferably, the sample set obtained by constructing the fault sample set in step 1 is:

[0015] Data=[(x1,y1,z1,w1),...,(x i ,y i ,z i ,w i ),...,(x n ,y n ,z n ,w n )];

[0016] Among them, x i is a single vibration signal sample, containing sampling points of a fixed length; y i is a fault type label; z i is the corresponding engine speed; w i is a sensor position label; n is the number of samples after overlapping sampling.

[0017] Preferably, the formula for normalization processing in step 2 is as follows:

[0018]

[0019] Among them, x max and x min are respectively the maximum and minimum values in the vibration signal sample, is the normalized sample, and x is the original vibration signal sample.

[0020] Preferably, the deep neural network model in step 3 includes the following modules:

[0021] Shared feature extraction network: Using a convolutional neural network CNN as a feature extraction module to extract the time-frequency features of the vibration signal. In the convolutional neural network CNN, a wide convolutional kernel and a residual connection mechanism are combined to improve the ability to capture the features of short-time signals and enhance the anti-noise performance of the model;

[0022] Main task classifier: On the basis of the shared features, the fault type classification result is output through a fully connected layer and a Softmax classifier;

[0023] Auxiliary task classifier: Including two independent branches for engine speed prediction and sensor position estimation, using the shared features for regression calculation, and outputting the speed prediction value and the sensor position estimation value.

[0024] Preferably, the specific steps of step 4 are as follows:

[0025] Step 41, Design of the main task loss: The fault classification task is used as the main task, and the cross-entropy loss function is used to measure the difference between the model prediction result and the true label;

[0026] Step 42, Design of the auxiliary task loss: The speed value and the position information are discretized into multiple categories, and each category is represented by one-hot encoding. The cross-entropy loss function is also used for the speed prediction and the sensor position prediction tasks;

[0027] Step 43, Design of the joint loss function: The losses of the main task and the auxiliary task are weighted and combined to form a joint loss function.

[0028] Preferably, the cross - entropy loss function for measuring the difference between the model prediction result and the true label in step 41 is as follows:

[0029]

[0030] where y i is the true fault label of the sample, expressed in one - hot encoding form; is the probability distribution predicted by the model, output after passing through the Softmax activation function.

[0031] Preferably, the cross - entropy loss function for rotational speed prediction in step 42 is as follows:

[0032]

[0033] where z i is the corresponding engine rotational speed, is the probability distribution predicted by the rotational speed model;

[0034] The cross - entropy loss function for the sensor position prediction task is as follows:

[0035]

[0036] where w i is the sensor position label, is the probability distribution predicted by the sensor position model.

[0037] Preferably, the joint loss function in step 43 is as follows:

[0038] Loss=L fault +λ1L speed +λ2L position ;

[0039] where L fault is the cross - entropy loss of the main task, used to optimize the fault classification performance; L speed and L position are the cross - entropy loss functions for rotational speed prediction and sensor position respectively, used to optimize the performance of the auxiliary tasks; λ1 and λ2 are the weight coefficients of the auxiliary tasks respectively, used to balance the contribution ratio of the main task and the auxiliary tasks to the total loss.

[0040] Preferably, step 5 specifically includes: in the training stage, using the training set to train the model parameters, minimizing the joint loss function based on the Adam optimizer; in the testing stage, inputting the test set sample data, outputting the fault classification result through the main task branch, and at the same time providing the prediction results of the rotational speed and the sensor position through the auxiliary task branches.

[0041] Therefore, the present invention adopts the above multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine, and has the following beneficial effects:

[0042] (1) Multi-task collaborative optimization: By introducing auxiliary tasks of rotational speed prediction and sensor position estimation, multi-source information is fully utilized, significantly improving the feature extraction ability and diagnostic accuracy of the fault classification task.

[0043] (2) Enhanced anti-noise performance: A shared feature extraction network with wide convolutional kernels and residual structures is adopted, enhancing the anti-interference ability of the model against high-noise signals under complex working conditions.

[0044] (3) Strong applicability under few-sample conditions: The dataset is expanded by the sliding window data augmentation method, and multi-dimensional information of the auxiliary tasks is utilized, significantly alleviating the limitation of sample scarcity on model training.

[0045] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0046] Figure 1 is a flowchart of a multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to the present invention;

[0047] Figure 2 is a confusion matrix showing the fault diagnosis results of the diagnostic model at 6 sensor positions in an embodiment of the present invention, where (a) is the confusion matrix at position 1, (b) is the confusion matrix at position 2, (c) is the confusion matrix at position 3, (d) is the confusion matrix at position 4, (e) is the confusion matrix at position 5, and (f) is the confusion matrix at position 6;

[0048] Figure 3 is a schematic diagram showing the distribution of 6 sensor positions in an embodiment of the present invention. Detailed Embodiments

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "comprising" or "including" mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0051] Embodiment

[0052] Please refer to Figures 1-3 , the present invention provides a multi-task learning fault diagnosis method for an aero-engine intermediate bearing, including the following steps:

[0053] Step 1: Vibration signal data acquisition and sample construction: During the operation of the aero-engine, the acceleration vibration signal of the aero-engine intermediate bearing is collected through a data acquisition system. In this embodiment, the data acquisition system collects data through an acceleration sensor, and at the same time records the corresponding engine speed and sensor position information. The collected data includes acceleration vibration signals in normal states and typical fault modes, and the engine speed is changed from low speed to high speed. The type labels of the acceleration vibration signals include normal, inner race fault and outer race fault, ensuring that the signals are representative of multi-condition conditions.

[0054] The original acceleration vibration signal is repeatedly sampled through a sliding window to achieve data augmentation and construct a fault sample set. Each sample consists of a vibration signal, a fault type label (such as normal, inner race fault, etc.), a speed label and a sensor position label.

[0055] The original vibration signal is repeatedly sampled through a sliding window. Set the sliding window length (L sample ), which is determined to be 2048 according to the signal sampling frequency and fault characteristics, to ensure that each sample contains complete fault characteristics; set the window step size (P) to 50% of the window length to generate adjacent but non-repeating samples to achieve data augmentation. Each segment of the vibration signal is sliced according to the sliding window to generate multiple sub-samples. Each sliced sample inherits the fault type, speed and sensor position information labels of the original signal to form an augmented sample set. The sample set is:

[0056] Data=[(x1,y1,z1,w1),...,(xi , y i , z i , w i ),...,(x n , y n , z n , w n )];

[0057] Among them, x i is a single vibration signal sample, containing sampling points of a fixed length; y i is the fault type label; z i is the corresponding engine speed; w i is the sensor position label; n is the number of samples after overlapping sampling.

[0058] Step 2: Data preprocessing: Normalize the sample data in the generated fault sample set to eliminate the influence of signal amplitude and dimension differences between different sample data on model training. The formula for normalization is as follows:

[0059]

[0060] Among them, x max and x min are the maximum and minimum values in the vibration signal sample respectively, is the normalized sample, and x is the original vibration signal sample. The normalized data samples are distributed within a fixed range, which is convenient for the neural network model to effectively learn the input features and avoid the problem of unstable gradients caused by signal amplitude differences. Take the vibration signal x i as the vibration signal input, and take the fault type y i , the engine speed z i and the sensor position w i as the supervision signal input.

[0061] Step 3: Multi-task learning model design: Build a deep neural network model based on multi-task learning, including the following modules:

[0062] Shared feature extraction network: Use the convolutional neural network CNN as the feature extraction module to extract the time-frequency features of the vibration signal. The wide convolutional kernel and residual connection mechanism are combined in the convolutional neural network CNN to improve the ability to capture short-time signal features and enhance the anti-noise performance of the model;

[0063] Main task classifier: Based on the shared features, output the fault type classification result through the fully connected layer and the Softmax classifier;

[0064] Auxiliary task classifier: It includes two independent branches for engine speed prediction and sensor position estimation, uses shared features for regression calculation, and outputs the speed prediction value and the sensor position estimation value.

[0065] Step 4: Joint loss function optimization: Design a joint loss function. The design of the joint loss function is the core of the multi-task learning model and directly determines the collaborative optimization effect between the main task (fault classification) and the auxiliary tasks (speed prediction, sensor position estimation). To balance the learning objectives of the main task and the auxiliary tasks during the optimization process, the joint loss function integrates multiple loss calculation methods, and the specific steps are as follows:

[0066] Step 41: Design of the main task loss: The fault classification task is the main task, and the cross-entropy loss function is used to measure the difference between the model prediction result and the true label;

[0067] The cross-entropy loss function for measuring the difference between the model prediction result and the true label in Step 41 is as follows:

[0068]

[0069] where, y i is the true fault label of the sample, represented in one-hot encoding form; is the probability distribution predicted by the model, output after passing through the Softmax activation function. Cross-entropy loss can effectively handle classification tasks and ensure that the model's prediction results for different fault categories have high confidence.

[0070] Step 42: Design of the auxiliary task loss: Discretize the speed value and position information into multiple categories, each category is represented by one-hot encoding, and the cross-entropy loss function is also used for the speed prediction and sensor position prediction tasks.

[0071] The cross-entropy loss function for speed prediction is as follows:

[0072]

[0073] where, z i is the corresponding engine speed, is the probability distribution predicted by the speed model;

[0074] The cross-entropy loss function for the sensor position prediction task is as follows:

[0075]

[0076] where, w i is the sensor position label, is the probability distribution predicted by the sensor position model.

[0077] Step 43: Design of the combined loss function: The losses of the main task and the auxiliary tasks are weighted and combined to form a combined loss function, which is as follows:

[0078] Loss = L fault + λ1L speed + λ2L position ;

[0079] where L fault is the cross-entropy loss of the main task, which is used to optimize the fault classification performance; L speed and L position are the cross-entropy loss functions of rotational speed prediction and sensor position estimation respectively, which are used to optimize the performance of the auxiliary tasks; λ1 and λ2 are the weight coefficients of the auxiliary tasks respectively, which are used to balance the contribution ratio of the main task and the auxiliary tasks to the total loss.

[0080] Through the combined optimization strategy, the information of the auxiliary tasks is used to enhance the feature extraction ability of fault classification and improve the overall performance of the model. The weight of the main task is usually set to 1 (by default, focusing on the fault classification task); the weights of the auxiliary tasks (λ1 and λ2) are adjusted according to the importance of the actual tasks. For example, they can be set to 0.5 or 0.3 to avoid excessive interference of the auxiliary tasks on the optimization of the main task; if it is found that the loss of a certain auxiliary task is large, the corresponding weight can be appropriately increased to balance the optimization direction of the model.

[0081] Step 5: Model training and testing: The processed sample data is input into the multi-task learning model for training and testing. The enhanced sample data set is divided into a training set and a testing set; the multi-task learning model is trained using the training set, and the Adam optimizer and the dynamic learning rate adjustment strategy are adopted to accelerate the convergence of the model; the combined loss function is optimized through multiple rounds of iteration to improve the learning effects of the main task and the auxiliary tasks. Subsequently, the test set samples are input into the trained model, and the fault classification results are output through the main task branch, and their classification accuracy is verified; the rotational speed prediction and sensor position estimation results of the auxiliary task branch are verified, and the performance of the auxiliary tasks is evaluated in combination with the mean square error index.

[0082] The aero-engine intermediate bearing fault diagnosis method of the present invention has a high fault recognition accuracy. The HIT aero-engine intermediate bearing data is used to verify this fault diagnosis method, and finally an accuracy of 92.80% is obtained.

[0083] In this embodiment, a comparative experiment is carried out for 6 sensors at different positions, as Figure 3 shown. The 1, 2, 3, 4, 5, and 6 in the figure represent the installation positions of the 6 sensors. To verify the applicability and robustness of this method under different position conditions.

[0084] To evaluate the diagnostic performance of the model, the precision of fault diagnosis was analyzed in detail using a confusion matrix, which intuitively reflects the performance of the diagnostic model by calculating the number of correct and incorrect classifications of the model for different fault categories. Figure 2 This is the confusion matrix of this method at 6 sensor positions. From Figure 2 It can be seen that the abscissa of the confusion matrix represents the fault diagnosis results of the diagnostic model, and the ordinate represents the actual fault category labels. At all sensor positions, this method can accurately identify various rolling bearing faults, and the diagnostic effect shows high accuracy.

[0085] Therefore, the present invention adopts the above-mentioned multi-task learning fault diagnosis method for an aero-engine intermediate bearing. By introducing auxiliary tasks of rotational speed prediction and sensor position estimation, it makes full use of multi-source information and significantly improves the feature extraction ability and diagnostic precision of the fault classification task; it uses a shared feature extraction network with wide convolution kernels and a residual structure to enhance the anti-interference ability of the model for high-noise signals under complex working conditions; it expands the dataset through the sliding window data augmentation method and uses the multi-dimensional information of the auxiliary tasks to significantly alleviate the limitation of sample scarcity on model training.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-task learning fault diagnosis method for intermediate bearings of aero-engines, characterized in that It includes the following steps: Step 1: Vibration signal data acquisition and sample construction: Collect the acceleration vibration signals of the intermediate bearing of the aero-engine through a data acquisition system, and simultaneously record the corresponding engine speed and sensor position information. Repeat the sampling of the original acceleration vibration signals through a sliding window to achieve data augmentation and construct a fault sample set; Step 2: Data preprocessing: Normalize the sample data in the generated fault sample set to eliminate the influence of signal amplitude and dimension differences between different sample data on model training; Step 3: Multi-task learning model design: Construct a deep neural network model based on multi-task learning; Step 4: Joint loss function optimization: Design a joint loss function to balance the learning objectives of the main task and the auxiliary task during the optimization process; Step 5: Model training and testing: Input the processed sample data into the multi-task learning model for training and testing.

2. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 1, characterized in that In Step 1, the data acquisition system collects the acceleration vibration signals in the normal state and typical fault modes, and the engine speed ranges from low speed to high speed; The type labels of the acceleration vibration signals include normal, inner race fault, and outer race fault.

3. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 2, wherein, The sample set obtained by constructing the fault sample set in Step 1 is as follows: Data = [(x1, y1, z1, w1),...,(x i , y i , z i , w i ),...,(x n , y n , z n , w n )]; Among them, x i is a single vibration signal sample, including sampling points of a fixed length; y i is the fault type label; z i is the corresponding engine speed; w i is the sensor position label; n is the number of samples after overlapping sampling.

4. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 3, wherein The formula for normalization processing in Step 2 is as follows: where x max and x min are the maximum and minimum values in the vibration signal sample respectively, is the normalized sample, and x is the original vibration signal sample.

5. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 4, characterized in that The deep neural network model in Step 3 includes the following modules: Shared feature extraction network: Use a convolutional neural network (CNN) as the feature extraction module to extract the time-frequency features of the vibration signals. The wide convolutional kernel and residual connection mechanism are combined in the CNN to improve the ability to capture the features of short-time signals and enhance the anti-noise performance of the model; Main task classifier: Based on the shared features, output the fault type classification results through a fully connected layer and a Softmax classifier; Auxiliary task classifier: It includes two independent branches for engine speed prediction and sensor position estimation, and uses the shared features for regression calculation to output the speed prediction value and the sensor position estimation value.

6. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 5, characterized in that The specific steps of Step 4 are as follows: Step 41: Design of the main task loss: The fault classification task is used as the main task, and the cross-entropy loss function is used to measure the difference between the model prediction result and the true label; Step 42: Design of the auxiliary task loss: Discretize the speed value and position information into multiple categories, and each category is represented by one-hot encoding. The cross-entropy loss function is also used for the speed prediction and sensor position prediction tasks; Step 43: Design of the joint loss function: Combine the losses of the main task and the auxiliary task by weighting to form a joint loss function.

7. A multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 6, characterized in that The cross-entropy loss function for measuring the difference between the model prediction result and the true label in Step 41 is as follows: Among them, y i is the true fault label of the sample, expressed in one-hot encoding form; is the probability distribution predicted by the model, which is output after passing through the Softmax activation function.

8. A multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 7, characterized in that: The cross-entropy loss function for speed prediction in Step 42 is as follows: where z i is the corresponding engine speed, is the probability distribution predicted by the speed model; The cross-entropy loss function for the sensor position prediction task is as follows: Among them, w i is the sensor position label, is the probability distribution predicted by the sensor position model.

9. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 8, characterized in that The joint loss function in Step 43 is as follows: Loss=L fault +λ1L speed +λ2L position ; Among them, L fault is the cross-entropy loss of the main task, which is used to optimize the fault classification performance; L speed and L position are the cross-entropy loss functions of rotational speed prediction and sensor position respectively, which are used to optimize the performance of the auxiliary tasks; λ1 and λ2 are the weight coefficients of the auxiliary tasks respectively, which are used to balance the contribution ratio of the main task and the auxiliary tasks to the total loss.

10. The multi-task learning fault diagnosis method for an intermediate bearing of an aero-engine according to claim 9, wherein Step 5 specifically includes: in the training phase, the model parameters are trained using the training set, and the joint loss function is minimized based on the Adam optimizer; in the testing phase, the test set sample data is input, the fault classification result is output through the main task branch, and at the same time, the prediction results of the rotational speed and the sensor position are provided through the auxiliary task branch.

Images