A spacecraft fault diagnosis method based on recursive meta-learning

Abstract

The application provides a spacecraft fault diagnosis method based on recursive meta-learning, comprising the following steps: training an LSTM neural network model; dividing a data set of recursive meta-learning; training a fast-adaptive recursive meta-learning model; and predicting a spacecraft fault through the trained recursive meta-learning model. Compared with a traditional LSTM neural network, the spacecraft fault diagnosis method based on recursive meta-learning provided by the application can significantly reduce the mean square error at the beginning of adapting to a new fault diagnosis task, and can be trained to converge more quickly, thereby improving the generalization ability of an on-orbit spacecraft fault diagnosis model, improving the accuracy of fault diagnosis under the condition of scarce sample data, and significantly reducing the signal prediction mean square error compared with the traditional LSTM neural network, and has strong engineering application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spacecraft fault diagnosis, and in particular to a spacecraft fault diagnosis method based on recursive meta-learning. BACKGROUND

[0002] With the rapid development of aerospace technology and artificial intelligence, data-driven spacecraft telemetry signal fault diagnosis has become an important research direction in the field of spacecraft reliability and space safety. At present, deep neural networks, convolutional neural networks and recurrent neural network models have attracted great attention in the field of spacecraft fault diagnosis due to their powerful signal feature extraction capabilities. Although deep learning models have to some extent freed themselves from the dependence on professional knowledge and engineering experience, they still need a large number of labeled sample data to train the model. Since only a small part of the signals collected by the spacecraft during on-orbit operation have labels and the fault samples are relatively scarce, the application of the above fault diagnosis methods on spacecraft equipment is limited. Therefore, how to improve the self-recognition and self-learning ability of the model through artificial intelligence and machine learning, and develop a small-sample fault diagnosis algorithm with better real-time performance, higher precision and better autonomy is an urgent problem to be solved in the field of spacecraft safety.

[0003] In order to solve the problem of insufficient sample data in the field of spacecraft fault diagnosis, a new model and framework need to be designed for real-time and safe and reliable small-sample fault diagnosis method. Meta-learning, as a general paradigm for solving small-sample learning problems, can realize cross-task knowledge transfer and rapid adaptation to new tasks. Among them, the model-agnostic meta-learning (MAML) algorithm based on parameter optimization has a simple and effective learning strategy, and stands out among a large number of meta-learning algorithms. In the prior art, MAML is migrated to the field of fault diagnosis, and a fault classification learning framework is fused with a triplet neural network. This framework has good fault recognition ability under the condition of sparse samples. In order to further improve the performance of small-sample fault diagnosis, a long short-term memory (LSTM) neural network with cell state memory capability can be used as the internal network of the MAML framework, and the convergence, model generalization ability, fault prediction accuracy and other aspects of the LSTM neural network itself can still be further optimized. SUMMARY

[0004] The purpose of the present application is to provide an improved spacecraft signal fault diagnosis scheme that can further improve the performance of spacecraft fault diagnosis under the condition of scarce sample data, in view of the deficiencies in the above background art.

[0005] In order to achieve the above purpose, the present application provides a spacecraft fault diagnosis method based on recursive meta-learning, comprising the following steps:

[0006] S1, training an LSTM neural network model;

[0007] S2, dividing the dataset of recursive meta-learning;

[0008] S3, training a fast-adapting recursive meta-learning model;

[0009] S4, predicting spacecraft failures through the trained recursive meta-learning model.

[0010] Further, the LSTM neural network in S1 includes a forget gate, an input gate, and an output gate.

[0011] At each time, the spacecraft sample data is first transmitted to the forget gate to clean up the previous useless cell state, then the cell state is updated through the input gate, and finally the cell state is output through the output gate.

[0012] Further, the mathematical model of the forget gate is as follows:

[0013]

[0014] wherein, is the output of the forget gate, is the input of the previous hidden layer, is the input of the current time, W f and b f are the bias coefficient and weight matrix of the forget gate, respectively, and sigmoid is selected as the neural network activation function.

[0015] Further, the mathematical model of the input gate is as follows:

[0016]

[0017] wherein, is the output of the input gate, is the updated information due to the current input, is the cell state of the current time, W i , W C are the weight matrices of the output gate, b i , b C are the bias coefficients of the output gate.

[0018] Further, the mathematical model of the output gate is as follows:

[0019]

[0020] wherein, is the output of the output gate, is the state output of the current time, W o and bo These are the bias coefficients and weight matrix of the output gate, respectively.

[0021] Furthermore, the training set p(τ) used for the S2 meta-learning task includes K time series signal prediction tasks τ, each of which contains m sets of labeled time series data:

[0022] p(τ)={τ1,τ2,...,τ n}

[0023]

[0024] Where, x i For task τ i The input signal, y i The labels correspond to the signals in each group in the task, and the labels are the actual signal values ​​at the time to be predicted.

[0025] Furthermore, in S3, the initial weights θ of the LSTM neural network are used to form the neural network f. θ f θ The update process includes inner loop update and outer loop update.

[0026] Furthermore, during the inner loop update process, when adapting to the training set task τ i At that time, the neural network uses the mean squared error loss function. For each classification task, the computation is performed, and gradient descent is used to update the neural network weights θ to weights θ′:

[0027]

[0028]

[0029] Where α is the step size of the inner loop update process. For gradient operators;

[0030] The goal of the inner ring road renewal is:

[0031]

[0032] Furthermore, the outer loop is updated as part of the process of the recurrent meta-learning model optimizing the initial weights θ of the LSTM neural network across tasks:

[0033]

[0034] Where β is the step size of the outer loop update process.

[0035] The above-described solution of the present invention has the following beneficial effects:

[0036] The spacecraft fault diagnosis method based on recursive meta-learning provided by the application has the advantages that, compared with a traditional LSTM neural network, the mean square error is obviously reduced at the beginning time of adapting to a new fault diagnosis task, and the training can be faster to convergence, the generalization ability of the on-orbit spacecraft fault diagnosis model is improved, the accuracy of fault diagnosis under the condition of sample data scarcity is improved, the signal prediction mean square error is obviously reduced compared with the traditional LSTM neural network, and the method has strong engineering application value.

[0037] Other beneficial effects of the application will be described in detail in the subsequent specific embodiment part. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A method step flowchart of the application;

[0039] Figure 2 A loss function iteration diagram when the recursive meta-learning model and the random initial parameter LSTM neural network adapt to a small sample time series prediction task in the embodiment of the application;

[0040] Figure 3 A comparison diagram of small sample spacecraft bus current signal prediction results of the recursive meta-learning model and the random initial parameter LSTM neural network when K=3 and N=6 in the embodiment of the application;

[0041] Figure 4 A comparison diagram of small sample spacecraft bus current signal prediction results of the recursive meta-learning model and the random initial parameter LSTM neural network when K=3 and N=10 in the embodiment of the application. DETAILED DESCRIPTION

[0042] The embodiments of the present disclosure will be described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in the specification. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all. The present disclosure can also be implemented or applied by other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present disclosure.

[0043] It is to be understood that the embodiments described hereinbelow within the scope of the appended claims. It will be apparent to one of ordinary skill in the art that aspects described herein can be implemented in a wide variety of forms, and that any specific structure and / or function described herein is merely illustrative. Based on the teachings provided herein one skilled in the art will appreciate that one or more aspects described herein can be implemented independently of any other aspects described herein. For example, an apparatus can implement one but not all of the aspects described herein. An apparatus can implement two but not all of the aspects described herein. An apparatus can implement three or more, but not all of the aspects described herein. An apparatus can implement all of the aspects described herein. An apparatus can implement any combination of the aspects described herein. An apparatus can implement one or more aspects described herein in combination with one or more aspects of any other method described herein. An apparatus can implement one or more aspects described herein in combination with one or more aspects of any other method described herein.

[0044] It is also to be understood that the drawings provided herein are merely schematic and that actual implementations can differ from the schematic illustrations as described herein.

[0045] As shown in Figure 1 Embodiments of the present application provide a spacecraft fault diagnosis method based on recursive meta-learning, comprising the following steps:

[0046] S1, training an LSTM neural network model.

[0047] The LSTM neural network is a special recursive neural network that realizes the transmission of information of each memory cell by introducing a memory gate, effectively solving the long-term dependence problem in time series prediction. The LSTM neural network controls the influence of previous information on current information through three different gates, namely the forgetting gate, the input gate, and the output gate.

[0048] At each time, the spacecraft sample data is first transmitted to the forgetting gate to clean up the previous useless cell state, then the cell state is updated through the input gate, and finally the cell state is output through the output gate. After the above operation, the LSTM neural network can learn the short-term influence and long-term regularity in the time series information, and realize the prediction of the information at future time.

[0049] The forgetting gate is used to determine the unit state at the previous time, and at each time, the sample data is first transmitted to the forgetting gate. The function of the forgetting gate is to delete useless information, and the mathematical model of the forgetting gate is as follows:

[0050]

[0051] wherein, is the output of the forget gate, is the input of the previous time hidden layer, is the input of the current time, W f and b f are the bias coefficient and weight matrix of the forget gate respectively, and sigmoid is selected as the neural network activation function (which is prior art and will not be described here).

[0052] wherein the input gate is used to determine the current network input, and the mathematical model thereof is as follows:

[0053]

[0054] wherein, is the output of the input gate, is the updated information due to the current input, is the cell state of the current time, W i , W C is the output gate weight matrix, b i , b C is the bias coefficient of the output gate.

[0055] wherein the output gate is used to control the state of the cell, and the mathematical model thereof is as follows:

[0056]

[0057] wherein, is the output of the output gate, is the state output of the current time, W o and b o are the bias coefficient and weight matrix of the output gate respectively.

[0058] S2, divide the dataset of recursive meta-learning.

[0059] In the embodiment, the training set p(τ) used by the meta-learning task includes K time series signal prediction tasks τ, and each task contains multiple sets of labeled time series data:

[0060] p(τ)={τ1,τ2,...,τ n}

[0061]

[0062] wherein x i is the input signal in the task τ i , y iThe label corresponding to each group of signals in the task, m is the number of groups of signals, in this embodiment, the label is the true signal value at the time to be predicted. During the training process of the meta-learning model, N groups of signals are randomly extracted from the training set p(τ) for model training.

[0063] In this embodiment, the function of the recursive meta-learning model is to find a function F according to the data set p(τ) including multiple time series prediction tasks, which has the ability to learn the initial parameters θ widely applicable to all tasks in p(τ), as shown in the following formula.

[0064] F(τ)→θ

[0065] S3, training the fast-adapting recursive meta-learning model.

[0066] During the training process of the recursive meta-learning model, the initial weights θ of the LSTM neural network are composed into a neural network f θ , the update process of f θ is divided into inner loop update and outer loop update. The inner loop update process can be summarized as follows: when the task τ i in the adaptation training set, the neural network uses the mean square error loss function for each classification task, and the neural network weights θ are updated to weights θ' using gradient descent, and the inner loop update process is shown in the following two formulas:

[0067]

[0068]

[0069] where α is the step size of the inner loop update process, is the gradient operator.

[0070] The target of the inner loop update is defined as follows:

[0071]

[0072] The outer loop update is the process of optimizing the initial weights θ of the LSTM neural network across tasks in the recursive meta-learning model:

[0073]

[0074] where β is the step size of the outer loop update process.

[0075] Through the above training steps, a set of initial weights θ of the LSTM neural network with strong generalization ability can be obtained, and the network with this set of parameters can quickly adapt to new tasks.

[0076] S4, spacecraft fault diagnosis through the trained recursive meta-learning model.

[0077] For the on-orbit spacecraft fault diagnosis problem, in this embodiment, the telemetry signals of the on-orbit spacecraft provided by the important cooperative laboratory of the United States Aerospace Administration, the atmospheric and space physics laboratory are used as a simulation data set, the battery temperature, bus current and reaction flywheel speed of the on-orbit spacecraft are used as the training data set of the recursive meta-learning model, the on-orbit spacecraft bus current data with a small amount of artificial labels are used as the test set, and the telemetry signals of the four channels of the spacecraft are all single variable time series.

[0078] The specific simulation parameters of the simulation experiment provided by the embodiment are as follows:

[0079] The on-orbit spacecraft battery temperature data records 128288 spacecraft battery temperature readings in about 10 years. The sampling record frequency is once per hour, and the missing values are filled by using linear interpolation method. The battery temperature changes regularly with the change of seasons, and the temperature increases year by year with the increase of service life.

[0080] The on-orbit spacecraft bus current records 51704 bus voltage readings in about 10 years. The normal frequency of data measurement is every 3 hours. Because the original data change period is 5 minutes, the frequency is the result of resampling, and the missing values are filled by using linear interpolation.

[0081] The on-orbit spacecraft reaction flywheel speed records 984911 spacecraft reaction flywheel speeds in about 10 years. The normal frequency of data measurement is every 5 minutes, and the missing values are filled with zero because the reaction wheel is not always in motion.

[0082] Through numerical simulation demonstration, the on-orbit spacecraft fault diagnosis model of the scheme improves the generalization ability, as shown in Figure 2 , at the beginning of adapting to a new fault diagnosis task, the mean square error of the scheme is reduced by 3.58 compared with the LSTM neural network algorithm, and it can be trained to converge faster. At the same time, as shown in Figure 3 , Figure 4 , the scheme significantly improves the accuracy of fault diagnosis under the condition of sample data scarcity. In the small sample spacecraft bus current fault diagnosis task, the signal prediction mean square error is reduced by 0.291 compared with the LSTM neural network, which has strong engineering application value.

[0083] Based on the same inventive concept, the embodiment also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the spacecraft fault diagnosis method based on recursive meta-learning.

[0084] The computer readable medium includes, but is not limited to, any type of disk including floppy disks, optical disks, CD-ROMs, and magnetic-optical disks, ROMs, RAMs, EPROMs (Erasable Programmable Read-Only Memory), EEPROMs, flash cards, magnetic cards, or optical cards. That is, the computer readable medium includes any medium that can store or carry information that can be read by a device such as a computer.

[0085] The computer readable storage medium provided by the embodiment has the same inventive concept and the same beneficial effects as the method described above, and will not be described here.

[0086] The above is the preferred embodiment of the present application, it should be noted that for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A spacecraft fault diagnosis method based on recursive meta-learning, characterized in that, The method comprises the following steps: S1, training an LSTM neural network model; S2, dividing a data set of recursive meta-learning; Training set used for meta-learning tasks Comprising K sequential signal prediction tasks Each task in turn comprising m sets of labeled sequential data: ； ； wherein, is a task input signal, a label for each group of signals in the task, the label being a true signal value at a time instant to be predicted; S3, training a fast-adapting recursive meta-learning model; the initial weights of the LSTM neural network composing a neural network , The update process includes inner loop updating and outer loop updating; During the inner loop update process, when adapting to the tasks in the training set , the neural network uses a mean squared error loss function for each classification task and uses gradient descent to update the neural network weights to weights : ； ； wherein is the step size for the inner loop update procedure, is the gradient operator; The target of inner loop update is: ； The outer loop updates the initial weights of the recurrent meta-learning model across tasks to optimize the LSTM neural network The process: ； wherein is the step size for the outer loop update procedure; S4, predicting spacecraft failure through the trained recursive meta-learning model.

2. The spacecraft fault diagnosis method based on recursive meta-learning according to claim 1, characterized in that, The LSTM neural network in S1 comprises a forgetting gate, an input gate, and an output gate; At each time, the spacecraft sample data is first transmitted to the forgetting gate to clean up the previous useless cell state, then the cell state is updated through the input gate, and finally the cell state is output through the output gate.

3. The spacecraft fault diagnosis method based on recursive meta-learning according to claim 2, characterized in that, The mathematical model of the forgetting gate is as follows: ； wherein, is the output of the forget gate, is the input of the hidden layer at the previous time step, is the input at the current time step, and are the bias coefficient and the weight matrix of the forget gate, respectively, and are chosen as the neural network activation function.

4. The spacecraft fault diagnosis method based on recursive meta-learning according to claim 3, characterized in that, The mathematical model of the input gate is as follows: ； wherein, is the output of the input gate, is the information updated due to the current input, is the cell state at the current time instant, , is the output gate weight matrix, , is the output gate bias coefficient.

5. The spacecraft fault diagnosis method based on recursive meta-learning according to claim 4, characterized in that, The mathematical model of the output gate is as follows: ； wherein, is the output of the output gate, is the state output at the current time instant, and are the bias coefficient and the weight matrix of the output gate, respectively.

Images