A method for transfer learning of fault features in gear pumps of online adjustable pitch propeller hydraulic systems

Abstract

This invention discloses a fault feature transfer learning method for gear pumps in an online controllable pitch propeller hydraulic system, relating to the field of intelligent fault diagnosis. The method includes the following steps: establishing a database of gear pumps in a controllable pitch propeller hydraulic system, including source domain data and target domain data; using Spearman correlation analysis and empirical mode decomposition to analyze the correlation and fault features of the data, obtaining preprocessed measurement point data sequences; using a convolutional network to fuse the measurement point data sequences and extract gear pump fault features; transforming the fault features into feature vectors and feeding them into a cross-entropy classifier for classification to obtain fault diagnosis results; saving the measurement point data sequences to an empirical replay pool, randomly sampling data from the empirical replay pool, and using stochastic gradient descent to optimize the parameters of the convolutional network and the cross-entropy classifier, improving learning efficiency and diagnostic performance. This invention updates fault features in real time, improving fault diagnosis accuracy and increasing the convergence speed and accuracy of online fault feature transfer learning.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis, and in particular to a method for transferring fault features of a gear pump in an online pitch control propeller hydraulic system. Background Technology

[0002] The effectiveness of data-driven intelligent fault diagnosis in machinery often relies on sufficient available monitoring data, which requires the data used to train the intelligent diagnostic model to have rich fault information and health labeling information. However, this is difficult to meet in engineering practice, mainly for the following two reasons:

[0003] 1) Hydraulic equipment remains in a normal operating state during long-term operation, with uncertainties in failure occurrence and short duration. Therefore, the number of failure samples obtained is far less than the number of normal samples, resulting in incomplete failure information and an imbalance in failure categories in the monitoring data;

[0004] 2) Although a large amount of data is available in engineering practice, only a small portion corresponds to the known health status of hydraulic equipment and can be used to train intelligent diagnostic models. Most data requires labeling with the corresponding health status, and labeling data is costly, for example, due to frequent shutdowns for self-checking faults, laborious manual labeling of data, and infrequent fault types, resulting in a lack of health labeling information for monitoring large datasets.

[0005] In new scenarios, the trained fault diagnosis model can be transferred to another device or different operating conditions with limited fault data to solve the above problems.

[0006] Currently, most fault diagnosis systems utilize transfer learning to address the shortage of real-world fault samples. Transfer learning can learn feature representations that are similar between the source and target domains while maintaining good classification performance on the source data. The differences between the source and target domains mainly stem from the differences between the source and target domain machines, such as different speeds or loads, different operating environments, and different physical structures. This difference leads to a shift in distribution towards cross-domain data. The more severe the cross-domain data transfer, the worse the diagnostic accuracy in the transfer diagnosis task. With the long-term operation of equipment, the system collects a large amount of data, and manually updating the model increases labor costs. Existing online fault diagnosis transfer solutions have slow convergence speeds and low accuracy, making them difficult to meet practical needs.

[0007] Furthermore, unlike common mechanical systems such as rotating machinery like bearings, gears, and windmills, where vibration sensors are often used to collect operational data and fault modes exhibit a relatively obvious periodicity, hydraulic systems, in addition to control signals, often have data such as pressure and temperature, and their fault modes do not exhibit obvious periodicity.

[0008] Therefore, those skilled in the art are dedicated to developing a transfer learning method for fault features of gear pumps in online controllable pitch propeller hydraulic systems. Summary of the Invention

[0009] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to achieve high-quality online fault feature transfer and improve the fault state recognition capability of the classifier.

[0010] To achieve the above objectives, the present invention provides a method for fault feature transfer learning of a gear pump in an online adjustable pitch propeller hydraulic system, characterized in that the method includes the following steps:

[0011] S101: Establish a database for the gear pump of the controllable pitch propeller hydraulic system. The database includes source domain data and target domain data. The source domain data includes mechanism model simulation data or historical data. The target domain data includes real-time data acquired online that needs to be diagnosed.

[0012] S103: Using Spearman correlation analysis and empirical mode decomposition, analyze the correlation and fault characteristics of the data in the gear pump database of the pitch control propeller hydraulic system, and obtain the measurement point data sequence in which the correlation and fault mode diversity both meet the preset conditions.

[0013] S105: Use a convolutional network to fuse the measurement point data sequence and extract the fault features of the gear pump;

[0014] S107: Transform the fault features into feature vectors, and feed the feature vectors into a cross-entropy classifier to classify the fault types, thereby obtaining the fault diagnosis results of the gear pump;

[0015] S109: The measurement point data sequence is stored in the experience replay pool, and the data in the experience replay pool is randomly sampled. The parameters of the convolutional network and the cross-entropy classifier are optimized using the stochastic gradient descent method to improve learning efficiency and diagnostic performance.

[0016] Furthermore, in S101, the historical data includes data on the gear pump of the controllable pitch propeller hydraulic system to be diagnosed within a historical period. The historical data and the real-time data include flow data, pressure data, temperature data, pitch commands, and feedback data.

[0017] Furthermore, the historical data and the real-time data are set to a predetermined length of data sequence according to different sampling frequencies, and multiple data points are collected and written into the database.

[0018] Further, in S103, the Spearman correlation coefficient is calculated in the following manner:

[0019]

[0020] Where ρ is the Spearman correlation coefficient, and d i =rank1-rank2, representing the difference between the rank of the first variable and the rank of the second variable, where i is the observation sample number and n is the total number of observation samples.

[0021] Further, in S103, the empirical mode decomposition includes the following steps:

[0022] S1031: Determine all local extreme points of the data signal to be decomposed, wherein the data signal is one of the flow data, the pressure data, the temperature data, or the pitch command and feedback data;

[0023] S1032: Connect all the local maxima points with cubic splines to form the upper envelope, and connect all the local minima points with cubic splines to form the lower envelope.

[0024] S1033: Calculate the intrinsic mode function (IMF) component of the data signal using the upper envelope and the lower envelope.

[0025] Furthermore, in S1033, the intrinsic mode function (IMF) components are calculated in the following manner:

[0026] S10331: Calculate the average value m1 of the upper envelope and the lower envelope;

[0027] S10332: Calculate the difference h1 between the data signal x(t) and the average value m1: x(t) - m1 = h1;

[0028] S10333: Determine whether h1 satisfies the IMF condition. If it does, then h1 is the first IMF component of x(t); otherwise, treat h1 as the original signal x(t) and repeat S10331 to S10333 until the IMF condition is satisfied.

[0029] S10334: Set c1 = h1, calculate x(t) - c1 = r1, and then let x(t) = r1. Repeat S10332 to S10334 for the next iteration to obtain the next IMF component.

[0030] Furthermore, in S105, the convolutional network is a 4-layer one-dimensional convolutional network, with each layer having a kernel size of 7 and output feature channel dimensions of 32, 64, 128, and 256, respectively.

[0031] Furthermore, in S107, the cross-entropy classifier consists of a linear transformation layer and three gear pump fault states, and the cross-entropy classifier includes a softmax function:

[0032]

[0033] Where h represents the features extracted by the convolutional network, w and b are the weights of the linear transformation layer, and i and j represent the categories in the output vector.

[0034] Further, in S109, the stochastic gradient descent method optimizes the parameters of the convolutional network and the cross-entropy classifier by solving a multi-objective loss function, wherein the multi-objective loss function is:

[0035]

[0036] Among them, Loss 总 For the total loss, Loss 分类 For classification loss, Loss 评价差异 The maximum average difference loss is given by: t = target domain sample, s = source domain sample, T = total target domain sample, S = total source domain sample, i = fault category number, and y′ = 0. ti For the true fault label of the i-th type of fault in the target domain, y ti Let h be the type i fault in the target domain output by the softmax function, ||·|| be the 2-norm, and h be the value of h. s h is the feature extracted from the source domain data. t Features extracted from the target domain data.

[0037] Furthermore, the multi-objective loss function includes a cross-entropy classification loss function and a maximum average difference loss function. The cross-entropy classification loss function is used to learn the differences between different faults in the source domain data, and the maximum average difference loss function is used to map the source domain data and the target domain data to the same distribution.

[0038] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. This invention uses online fault transfer learning to update model parameters in real time, reducing the distance between the source and target domains. It can update fault features in real time based on the operating status of the gear pump in the hydraulic system, thereby improving the accuracy of fault diagnosis.

[0040] 2. This invention provides a method to improve the learning convergence speed by using experience pool replay, saving the data collected by the current device in real time, and sampling part of the data when updating parameters, thereby improving the convergence speed and accuracy of online fault feature transfer learning.

[0041] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating the fault feature transfer learning method according to an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the framework of the fault feature transfer learning method according to an embodiment of the present invention. Detailed Implementation

[0044] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0045] like Figure 1 , Figure 2 As shown, addressing the issues of slow convergence speed and low accuracy in fault diagnosis transfer learning for controllable pitch gear pumps in existing technologies, this invention combines convolutional networks and an experience replay mechanism to efficiently learn fault features from small samples, improving fault diagnosis and identification accuracy, accelerating the learning speed of gear pump fault features, and thus enhancing the classifier's fault state identification capability. The online fault feature transfer learning method for gear pumps in controllable pitch hydraulic systems provided in this embodiment first performs signal processing and analysis on historical flow, pressure, and temperature data from an experimental platform, and then uses a convolutional network to learn and fuse features from multiple sensor signals. To meet the requirements of real-time performance and high diagnostic accuracy, this invention proposes online fault feature fusion and pattern learning based on experience replay convolutional networks. Furthermore, addressing the data distribution changes caused by system degradation characteristics, this invention employs gradient descent for online parameter updates of the model, resulting in stronger generalization ability and better adaptability of the recursive convolutional network, significantly improving diagnostic accuracy and convergence speed. In engineering applications, the online fault transfer scheme also reduces the workload of workers periodically updating the model.

[0046] Specifically, the present invention provides a method for transferring fault features of a gear pump in an online adjustable pitch propeller hydraulic system, comprising the following steps:

[0047] Step 1: Establish a database for gear pumps in the controllable pitch propeller hydraulic system.

[0048] The database includes source domain data and target domain data. Source domain data includes mechanistic model simulation data or historical data, while target domain data includes real-time data acquired online that is currently needed for diagnosis.

[0049] When establishing a database for the gear pump of a controllable pitch propeller hydraulic system, different databases can be created for the source domain data and the target domain data, such as a source domain data database and a target domain data database.

[0050] In the established source and target domain databases, the input data includes: a historical sequence of flow rate data, pressure data, temperature data, and pitch command and feedback data for a specific period of time for the gear pump in the controllable pitch propeller hydraulic system to be diagnosed. A fixed-length data sequence can be set according to different sampling frequencies, collecting multiple data points and writing them into the database. For example, when the sampling frequency is 10Hz, sampling for 120 seconds results in 1200 data points being collected and written to the data center table.

[0051] Step 2: Using Spearman correlation analysis and empirical mode decomposition, analyze the correlation and fault characteristics of the data in the gear pump database of the controllable pitch propeller hydraulic system, and obtain the measurement point data sequence that meets the preset conditions in terms of both correlation and fault mode diversity.

[0052] Specifically, Spearman correlation coefficient and empirical mode decomposition are used to analyze the correlation of flow data series, pressure data series, temperature data series, pitch command and feedback data series, as well as the fault characteristics of measuring points, and select data series with low correlation and high fault mode diversity.

[0053] Spearman's correlation coefficient is calculated as follows:

[0054]

[0055] Where p is the Spearman correlation coefficient, d i =rank1-rank2, representing the difference between the rank of the first variable and the rank of the second variable, where i is the observation sample number and n is the total number of observation samples.

[0056] Empirical Mode Decomposition (EMD) assumes that any complex signal is composed of a number of different intrinsic mode functions (IMFs). Each IMF, regardless of whether it is linear or nonlinear or nonstationary, must satisfy the following condition:

[0057] 1) For the same number of extreme points and zero crossings, there is only one extreme point between two adjacent zero crossings;

[0058] 2) The upper and lower envelopes are locally symmetrical about the time axis.

[0059] In this embodiment, any signal can be decomposed through the following steps:

[0060] 1) Determine all local extrema of the signal x(t). The signal x(t) can be one of the following: flow rate data, pressure data, temperature data, or pitch command and feedback data.

[0061] 2) Connect all the local maxima points with cubic splines to form the upper envelope.

[0062] 3) Connect all the local minimum points with cubic splines to form the lower envelope.

[0063] In steps 2) and 3) above, it is required that the upper and lower envelopes should encompass all data points.

[0064] 4) Let m1 be the average of the upper and lower envelopes, and calculate...

[0065] x(t)-m1=h1

[0066] If h1 satisfies the IMF condition, then h1 is the first IMF component of x(t).

[0067] If the conditions are not met, h1 is used as the original signal and the calculation is repeated until the conditions of IMF are met.

[0068] 5) Repeat the steps to calculate the next IMF component.

[0069] Let c1 = h1, and separate c1 from x(t):

[0070] x(t)-c1=r1,

[0071] Using r1 as the original data for the next iteration, repeat the above calculation steps n times to obtain n IMF components.

[0072] The above steps complete the Empirical Mode Decomposition (IMF). Based on the structure of the IMF, the trend of fault characteristics and information such as signal and noise can be determined. Furthermore, the IMF components after IMF decomposition can be fed into a convolutional network as a feature to automatically learn fault modes.

[0073] Step 3: Use a convolutional network to fuse the measurement point data sequence and extract the fault characteristics of the gear pump.

[0074] In this embodiment, a four-layer one-dimensional convolutional network is selected. The four-layer one-dimensional convolutional network is used to fuse the measurement point data sequence and extract the fault characteristics of the gear pump of the pitch control propeller hydraulic system from the IMF and measurement point signals.

[0075] The convolutional network consists of convolutional layers, pooling layers, and batch normalization layers. The network structure is as follows: the kernel size of each layer in this convolutional network is 7, and the output feature channel dimensions are 32, 64, 128, and 256, respectively.

[0076] Step 4: Transform the fault features into feature vectors, and then feed the feature vectors into a cross-entropy classifier to classify the fault types, thereby obtaining the fault diagnosis results of the gear pump.

[0077] Select the gear pump fault features extracted by the one-dimensional convolutional network in step 3, expand them into a one-dimensional feature vector, and then feed the one-dimensional feature vector into the cross-entropy classifier to classify the gear pump fault types.

[0078] The cross-entropy classifier consists of a linear transformation layer and three gear pump fault states. It also includes a softmax function, which is specifically defined as follows:

[0079]

[0080] Where h represents the features extracted by the convolutional network, w and b are the weights of the linear transformation layer, and i and j represent the categories in the output vector.

[0081] Step 5: Save the measurement point data sequence in the experience replay pool, and randomly sample the data in the experience replay pool. Use the stochastic gradient descent method to optimize the parameters of the convolutional network and cross-entropy classifier to improve learning efficiency and diagnostic performance.

[0082] To improve the system's learning efficiency and diagnostic performance, this embodiment also provides an experience replay pool, in which the data sequence from step 2 is concatenated into a matrix and stored. Figure 2 As shown.

[0083] After accumulating data over a certain period of time, and storing a certain amount of data in the experience replay pool, stochastic gradient descent is used to randomly sample the experience replay pool. Then, the parameters of the convolutional network and cross-entropy classifier are adjusted to meet the changes in data distribution, thereby improving learning efficiency and diagnostic performance.

[0084] The experience replay pool can be set up as a list to store N=1000 of the latest data. When the number of data exceeds 1000, the latest data replaces the oldest data to ensure the timeliness of the data.

[0085] Specifically, after m steps, n samples are randomly sampled from the experience replay pool, and then two loss functions are solved: the cross-entropy classification loss function and the maximum mean difference loss function. The cross-entropy classification loss function is used to learn the differences between different faults in the source domain; the maximum mean difference loss function is used to map the source domain data and the target domain data to the same distribution.

[0086] The stochastic gradient descent method optimizes the parameters of convolutional networks and cross-entropy classifiers by solving multi-objective loss functions. The multi-objective loss functions include the cross-entropy classification loss function and the maximum average difference loss function. The cross-entropy classification loss function is used to learn the differences between different faults in the source domain data, while the maximum average difference loss function is used to map the source domain data and the target domain data to the same distribution.

[0087] The multi-objective loss function is:

[0088]

[0089] Among them, Loss 总 For the total loss, Loss 分类 For classification loss, Loss 评价差异 The maximum average difference loss is given by: t = target domain sample, s = source domain sample, T = total target domain sample, S = total source domain sample, i = fault category number, and y′ = 0. ti For the true fault label of the i-th type of fault in the target domain, y ti Let h be the type i fault in the target domain output by the softmax function, ||·|| be the 2-norm, and h be the value of h. s h represents the features extracted from the source domain data by the convolutional network. t These are the features extracted from the target domain data by the convolutional network.

[0090] Compared with existing technologies, the online fault feature learning method for gear pumps in adjustable-pitch propeller hydraulic systems based on convolutional networks and empirical replay provided in this invention analyzes historical flow, pressure, and temperature data from an experimental bench, and uses convolutional networks to learn and fuse features from multiple sensor signals. To meet the requirements of real-time performance and high diagnostic accuracy, this invention proposes online fault feature fusion and pattern learning based on convolutional networks using empirical replay. Furthermore, to address data distribution changes caused by system degradation, this invention employs gradient descent for online parameter updates of the model, resulting in stronger generalization ability and better adaptability of the recursive convolutional network. This significantly improves diagnostic accuracy and convergence speed, and in engineering applications, the online fault migration scheme reduces the workload of workers periodically updating the model. Therefore, this invention has the following advantages:

[0091] 1. To address the problem that existing technologies cannot learn small sample fault modes in real time, this invention proposes online fault transfer learning, which updates model parameters in real time, reduces the distance between the source domain and the target domain, and updates fault features in real time according to the operating status of the gear pump in the hydraulic system, thereby improving the accuracy of fault diagnosis.

[0092] 2. To address the problem of slow online fault feature transfer learning speed in existing technologies, this invention uses experience pool replay to improve learning convergence speed, saves the data collected by the current device in real time, and samples part of the data when updating parameters, thereby improving the convergence speed and accuracy of online fault feature transfer learning.

[0093] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for transferring fault features of a gear pump in an online adjustable pitch propeller hydraulic system, characterized in that, The method includes the following steps: S101: Establish a database for the gear pump of the controllable pitch propeller hydraulic system. The database includes source domain data and target domain data. The source domain data includes mechanism model simulation data or historical data. The target domain data includes real-time data acquired online that needs to be diagnosed. S103: Using Spearman correlation analysis and empirical mode decomposition, analyze the correlation and fault characteristics of the data in the gear pump database of the pitch control propeller hydraulic system, and obtain the measurement point data sequence in which the correlation and fault mode diversity both meet the preset conditions. S105: Use a convolutional network to fuse the measurement point data sequence and extract the fault features of the gear pump; S107: Transform the fault features into feature vectors, and feed the feature vectors into a cross-entropy classifier to classify the fault types, thereby obtaining the fault diagnosis results of the gear pump; S109: The measurement point data sequence is stored in the experience replay pool, and the data in the experience replay pool is randomly sampled. The parameters of the convolutional network and the cross-entropy classifier are optimized using the stochastic gradient descent method to improve learning efficiency and diagnostic performance.

2. The method as described in claim 1, characterized in that, In S101, the historical data includes data on the gear pump of the controllable pitch propeller hydraulic system to be diagnosed within a historical period. The historical data and the real-time data include flow data, pressure data, temperature data, pitch commands, and feedback data.

3. The method as described in claim 2, characterized in that, The historical data and the real-time data are set to a predetermined length of data sequence according to different sampling frequencies, and multiple data points are collected and written into the database.

4. The method as described in claim 3, characterized in that, In step S103, the Spearman correlation coefficient is calculated as follows: Where ρ is the Spearman correlation coefficient, and d i =rank1-rank2, representing the difference between the rank of the first variable and the rank of the second variable, where i is the observation sample number and n is the total number of observation samples.

5. The method as described in claim 4, characterized in that, In step S103, the empirical mode decomposition includes the following steps: S1031: Determine all local extreme points of the data signal to be decomposed, wherein the data signal is one of the flow data, the pressure data, the temperature data, or the pitch command and feedback data; S1032: Connect all the local maxima points with cubic splines to form the upper envelope, and connect all the local minima points with cubic splines to form the lower envelope. S1033: Calculate the intrinsic mode function (IMF) component of the data signal using the upper envelope and the lower envelope.

6. The method as described in claim 5, characterized in that, In S1033, the intrinsic mode function (IMF) components are calculated in the following manner: S10331: Calculate the average value m1 of the upper envelope and the lower envelope; S10332: Calculate the difference h1 between the data signal x(y) and the average value m1: x(t) - m1 = h1; S10333: Determine whether h1 satisfies the IMF condition. If it does, then h1 is the first IMF component of x(t); otherwise, treat h1 as the original signal x(t) and repeat S10331 to S10333 until the IMF condition is satisfied. S10334: Set c1 = h1, calculate x(t) - c1 = r1, and then let x(t) = r1. Repeat S10332 to S10334 for the next iteration to obtain the next IMF component.

7. The method as described in claim 6, characterized in that, In S105, the convolutional network is a 4-layer one-dimensional convolutional network, with each layer having a kernel size of 7 and output feature channel dimensions of 32, 64, 128, and 256, respectively.

8. The method as described in claim 7, characterized in that, In S107, the cross-entropy classifier consists of a linear transformation layer and three gear pump fault states, and includes a softmax function: Where h represents the features extracted by the convolutional network, w and b are the weights of the linear transformation layer, and i and j represent the categories in the output vector.

9. The method as described in claim 8, characterized in that, In step S109, the stochastic gradient descent method optimizes the parameters of the convolutional network and the cross-entropy classifier by solving a multi-objective loss function, which is: Among them, Loss 总 For the total loss, Loss 分类 For classification loss, Loss 评价差异 The maximum average difference loss is given by: t = target domain sample, s = source domain sample, T = total target domain sample, S = total source domain sample, i = fault category number, and y′ = 0. ti For the true fault label of the i-th type of fault in the target domain, y ti Let h be the type i fault in the target domain output by the softmax function, ‖·‖ be the 2-norm, and h be the value of h. s h is the feature extracted from the source domain data. t Features extracted from the target domain data.

10. The method as described in claim 9, characterized in that, The multi-objective loss function includes a cross-entropy classification loss function and a maximum average difference loss function. The cross-entropy classification loss function is used to learn the differences between different faults in the source domain data, and the maximum average difference loss function is used to map the source domain data and the target domain data to the same distribution.

Images