Fault prediction method and system for aerospace payload rotating equipment

Abstract

The invention provides a fault prediction method and system for aerospace payload rotating equipment, and relates to the technical field of space science experiments. According to the method provided by the invention, by obtaining in-orbit operation data of the payload rotating equipment, a rotating equipment fault prediction model taking GPT-2 as a trunk is provided, and data embedding and a feature mapping layer are redesigned. And finally, knowledge migration is carried out on a specific layer of the GPT-2 model through a migration learning method, and an embedded and feature mapping layer is retrained, so that the model has cross-modal fault prediction capability, and compared with a fault prediction method for the payload rotating equipment in the related technology, the accuracy and timeliness of a prediction result can be improved.

Description

Technical Field

[0001] The present invention relates to the field of space science experiment technology, and in particular to a method and system for predicting failures of aerospace payload rotating equipment. Background Art

[0002] At present, the space station has entered the application and development stage. As an important space science experimental platform for the next decade, ensuring the long-term safe and stable operation of the space science payload in orbit is an important application requirement in the application and development stage of the space station. Among them, the rotating mechanical components of the effective experimental payload, as the key measurement components of the experimental data, have a vital impact on the reliability of the entire payload and the smooth conduct of the experiment. They have a limited lifespan and are difficult to replace, so fault prediction is of great significance.

[0003] In related technologies, the fault detection method of space station scientific experiment payloads mainly relies on the setting of sensor parameter thresholds based on expert experience, which has certain limitations and makes it difficult to achieve early prediction of faults. Research on payload fault prediction in the scientific field based on deep learning still faces problems such as weak early fault signal characteristics, scarce fault samples, and difficulty in model migration under variable working conditions and cross-working conditions. This results in low accuracy and poor timeliness of fault prediction. Summary of the invention

[0004] The embodiment of the present invention provides a method and system for predicting faults of aerospace payload rotating equipment. Based on the vibration signal of the payload rotating equipment, a trained neural network model can accurately and timely predict the fault type of the space station payload rotating equipment.

[0005] To achieve the above object, the embodiments of the present invention adopt the following technical solutions:

[0006] In a first aspect, a method for predicting a fault of a space payload rotating device is provided, the method comprising: obtaining historical data of the payload rotating device, the historical data being a first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, the first parameter being acceleration or angular velocity, and l and n being positive integers; determining a plurality of first training samples and a fault category corresponding to each first training sample based on the historical data, wherein each first training sample comprises a first parameter corresponding to each of y moments of the payload rotating device obtained by x sensors, x and y being positive integers, and x being less than or equal to n, and y being less than or equal to l; training a neural network model based on the plurality of first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model; obtaining on-orbit operation data of the payload rotating device, the on-orbit operation data being a first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, and l and n being positive integers; inputting the on-orbit operation data into the trained neural network model, and outputting a fault prediction result corresponding to the on-orbit operation data, the fault prediction result comprising a fault category corresponding to the on-orbit operation data.

[0007] In a possible implementation of the first aspect, multiple first training samples and a fault category corresponding to each first training sample are determined based on historical data, including: normalizing and denoising the historical data to eliminate noise interference in the historical data, and making the value of the first parameter corresponding to each moment included in the historical data fall within a preset threshold range; sampling the historical data to obtain multiple first training samples and a fault category corresponding to each first training sample.

[0008] In a possible implementation of the first aspect, a neural network model is trained according to a plurality of first training samples and a fault category corresponding to each first training sample to obtain a trained neural network model, including: inputting a plurality of first training samples into the neural network model, and outputting a predicted fault category corresponding to each first training sample; establishing an objective function according to the predicted fault category and the fault category corresponding to each first training sample; iteratively training the neural network model according to the objective function to obtain a trained neural network model;

[0009] The objective function L is:

[0010]

[0011] Among them, L represents the cross entropy loss value, n represents the number of fault categories, and y i represents the fault category of the i-th type, Represents the output of the neural network model for the i-th fault category.

[0012] In a possible implementation of the first aspect, the neural network model includes an embedding unit and multiple stacked Transformer units, the output end of the embedding unit is connected to the input end of the Transformer unit; the embedding unit is used to convert the input first training sample into a feature vector of a preset dimension corresponding to each first training sample; each Transformer unit includes an encoder and a decoder; the output end of the encoder of each Transformer unit is connected to the output end of the decoder of each Transformer unit; each encoder includes an attention layer and a feedforward neural network layer, the attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to encode each feature vector to obtain a latent vector corresponding to each feature vector; each decoder includes a masked multi-head attention layer, an attention layer and a feedforward neural network layer, the multi-head attention layer is used to associate the input of each encoding unit with the output of each decoding unit; the attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to decode each latent vector to obtain a feature vector corresponding to each feature vector.

[0013] In a possible implementation of the first aspect, before training the neural network model according to multiple first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model, the method further includes: constructing a neural network model; training the neural network model to obtain a pre-trained neural network model; and adjusting model parameters of the pre-trained neural network model based on the dimension of the first training sample to obtain a neural network model.

[0014] In a possible implementation of the first aspect, a neural network model is trained to obtain a pre-trained neural network model, including: performing modal decomposition processing on multiple first training samples to obtain second training samples corresponding to each first training sample, where the second training sample is a spectrum image including multiple frequency domain features; training the neural network model according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

[0015] The beneficial effects of the present invention are as follows: the method provided by the present invention obtains historical data of the payload rotating device, determines multiple first training samples based on the historical data of the payload rotating device, and determines the second training samples corresponding to the multiple first training samples through modal decomposition processing, and then pre-trains the neural network model based on the multiple second training samples. After the training is completed, the model parameters of the neural network model are adjusted based on the dimension of the first training sample, and some parameters obtained from the previous training are retained. The pre-trained neural network model is trained according to the multiple first training samples to obtain a trained neural network model, and then based on the trained neural network model, the fault category of the payload rotating device can be accurately and timely predicted according to the on-orbit operation data of the payload rotating device. Compared with the method for predicting faults of payload rotating devices in related technologies, the accuracy and timeliness of the prediction results can be improved. In addition, training time can be saved and training efficiency can be improved.

[0016] In a second aspect, an embodiment of the present invention provides a fault prediction system for aerospace payload rotating equipment, the system comprising: a data acquisition unit, used to acquire historical data of the payload rotating equipment, the historical data being a first parameter corresponding to each of l moments of the payload rotating equipment acquired by n sensors, the first parameter being acceleration or angular velocity, l and n being positive integers; a sample generation unit, used to determine a plurality of first training samples and a fault category corresponding to each first training sample based on the historical data, wherein each first training sample comprises a first parameter corresponding to each of y moments of the payload rotating equipment acquired by x sensors, x and y being positive integers, and x is less than or equal to n, and y is less than or equal to l; a model training unit, used to train the neural network model according to multiple first training samples and the fault category corresponding to each first training sample, to obtain a trained neural network model; a data acquisition unit, further used to obtain on-orbit operation data of the payload rotating device, the on-orbit operation data being the first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, where l and n are positive integers; a fault prediction unit, used to input the on-orbit operation data into the trained neural network model, and output a fault prediction result corresponding to the on-orbit operation data, the fault prediction result including the fault category corresponding to the on-orbit operation data.

[0017] In a possible implementation of the second aspect, the model training unit is further used to: construct a neural network model; train the neural network model to obtain a pre-trained neural network model; and adjust model parameters of the pre-trained neural network model based on the dimension of the first training sample to obtain a neural network model.

[0018] In a possible implementation of the second aspect, the model training unit is specifically used to: perform modal decomposition processing on multiple first training samples to obtain second training samples corresponding to each first training sample, where the second training sample is a spectrum image including multiple frequency domain features; train the neural network model according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

[0019] According to a third aspect, an electronic device is provided, comprising a memory and one or more processors; the memory is coupled to the processor; wherein the memory stores computer program code, the computer program code comprises computer instructions, and when the computer instructions are executed by the processor, the electronic device executes a method as in any implementation of the first aspect.

[0020] According to a fourth aspect, a computer-readable storage medium is provided, comprising computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes the method in any implementation of the first aspect.

[0021] According to a fifth aspect, a computer program product is provided. When the computer program product is run on a computer, the computer is enabled to execute the method in any implementation of the first aspect.

[0022] It can be understood that the beneficial effects that can be achieved by the system of the second aspect, the electronic device of the third aspect, the computer-readable storage medium of the fourth aspect, and the computer program product of the fifth aspect provided above can be referred to the beneficial effects in the first aspect and any possible design method thereof, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of the hardware structure of an electronic device shown in an embodiment of the present invention;

[0024] Figure 2 A flow chart of a method for predicting a fault of a rotating device of a space payload shown in an embodiment of the present invention;

[0025] Figure 3 A flow chart of a method for predicting a fault of a rotating device of a space payload shown in an embodiment of the present invention;

[0026] Figure 4 A flow chart of a method for predicting a fault of a rotating device of a space payload shown in an embodiment of the present invention;

[0027] Figure 5 A flow chart of a method for predicting a fault of a rotating device of a space payload shown in an embodiment of the present invention;

[0028] Figure 6A schematic flow chart of a method for predicting a fault of a rotating device of an aerospace payload according to an embodiment of the present invention;

[0029] Figure 7 The figure is a schematic diagram of the hardware structure of a prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The technical solution in the embodiment of the present invention will be described below in conjunction with the accompanying drawings in the embodiment of the present invention. Among them, in the description of the present invention, unless otherwise specified, " / " indicates that the objects associated before and after are in an "or" relationship, for example, A / B can represent A or B; the "or" in the present invention is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. And, in the description of the present invention, unless otherwise specified, "multiple" means two or more than two. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items.

[0031] In addition, in order to clearly describe the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, the words "first", "second", etc. are used to distinguish the same items or similar items with substantially the same functions and effects. Those skilled in the art can understand that the words "first", "second", etc. do not limit the quantity and execution order, and the words "first", "second", etc. do not necessarily limit the difference.

[0032] Meanwhile, in the embodiments of the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way for easy understanding.

[0033] At present, the space station has entered the application and development stage. As an important space science experimental platform for the next decade, ensuring the long-term safe and stable operation of the space science payload in orbit is an important application requirement in the application and development stage of the space station. Among them, the rotating mechanical components of the effective experimental payload, as the key measurement components of the experimental data, have a vital impact on the reliability of the entire payload and the smooth conduct of the experiment. They have a limited lifespan and are difficult to replace, so fault prediction is of great significance.

[0034] In related technologies, the fault detection method of space station scientific experiment payloads mainly relies on the setting of sensor parameter thresholds based on expert experience, which has certain limitations and makes it difficult to achieve early prediction of faults. Research on payload fault prediction in the scientific field based on deep learning still faces problems such as weak early fault signal characteristics, scarce fault samples, and difficulty in model migration under variable working conditions and cross-working conditions. This results in low accuracy and poor timeliness of fault prediction.

[0035] In view of this, an embodiment of the present invention provides a method for predicting a fault of a space payload rotating device, the method comprising: obtaining historical data of the payload rotating device, the historical data being a first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, the first parameter being acceleration or angular velocity, and l and n being positive integers; determining a plurality of first training samples and a fault category corresponding to each first training sample according to the historical data, wherein each first training sample comprises a first parameter corresponding to each of y moments of the payload rotating device obtained by x sensors, x and y being positive integers, and x being less than or equal to n, and y being less than or equal to l; training a neural network model according to the plurality of first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model; obtaining on-orbit operation data of the payload rotating device, the on-orbit operation data being a first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, and l and n being positive integers; inputting the on-orbit operation data into the trained neural network model, and outputting a fault prediction result corresponding to the on-orbit operation data, the fault prediction result comprising a fault category corresponding to the on-orbit operation data.

[0036] The method provided by the embodiment of the present invention obtains historical data of the payload rotating device, determines multiple first training samples based on the historical data of the payload rotating device, and determines the second training samples corresponding to the multiple first training samples through modal decomposition processing, and then pre-trains the neural network model based on the multiple second training samples. After the training is completed, the model parameters of the neural network model are adjusted based on the dimension of the first training sample, and some parameters obtained by the previous training are retained. The pre-trained neural network model is trained according to the multiple first training samples to obtain a trained neural network model, and then based on the trained neural network model, the fault category of the payload rotating device can be accurately and timely predicted according to the on-orbit operation data of the payload rotating device. Compared with the method for predicting faults of payload rotating devices in related technologies, the accuracy and timeliness of the prediction results can be improved. In addition, training time can be saved and training efficiency can be improved.

[0037] In some embodiments, the fault prediction method for aerospace payload rotating equipment provided by the embodiments of the present invention can be performed by a fault prediction system 100 for aerospace payload rotating equipment (hereinafter referred to as prediction system 100). As an example, the prediction system 100 can be any electronic device 200 with data processing capabilities, such as a general-purpose computer, a personal computer, a laptop, a switch, or a tablet computer, etc. The specific implementation of the purchase list prediction system 100 is not limited here.

[0038] Figure 1 The hardware structure diagram of the electronic device provided by the embodiment of the present invention is shown. The electronic device 200 includes a processor 210, a memory 220 and a communication interface 230.

[0039] The processor 210 may include one or more processing cores. The processor 210 uses various interfaces and lines to connect various parts in the electronic device 200, and executes various functions and processes data of the electronic device 200 by running or executing instructions, programs, code sets or instruction sets stored in the memory 220, and calling data stored in the memory 220. Optionally, the processor 210 can be implemented in at least one hardware form of a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processing (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA).

[0040] The memory 220 may include a random access memory (RAl) or a read-only memory (ROL). Optionally, the memory 220 includes a non-transitory computer-readable storage medium (non-transitory colputer-readable storage lediul). The memory 220 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 220 may include a program storage area. Among them, the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as a data acquisition function, a model training function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.

[0041] The communication interface 230 is used to communicate with other devices, equipment or communication networks, such as data storage devices, image processing equipment or Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0042] In physical implementation, the above-mentioned components (such as processor 210, memory 220 and communication interface 230) can be components in the same device (such as a laptop). Alternatively, at least two of the components can be set in the same device, that is, as different components in one device, such as a deployment method similar to devices or components in a distributed system.

[0043] It is to be understood that the structure illustrated in this embodiment does not constitute a specific limitation on the electronic device 200. In other embodiments of the present invention, the electronic device 200 may include more or fewer components than those illustrated, or combine certain components, or separate certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0044] The following describes a method for predicting a fault of a space payload rotating device provided by an embodiment of the present invention in conjunction with the accompanying drawings.

[0045] Figure 2 A flowchart of a method for predicting a fault of a space payload rotating device provided by an embodiment of the present invention. Optionally, the method may be Figure 1 The electronic device 200 shown is executed, that is, executed by the prediction system 100. The method may include the following steps:

[0046] S1. Obtain historical data of payload rotation equipment.

[0047] Specifically, the historical data is the first parameter corresponding to each of l moments of the effective load rotating device obtained by n sensors, the first parameter is acceleration or angular velocity, and l and n are positive integers. It can also be understood that the historical data is the vibration signal of the effective load rotating device obtained by n sensors in the time period corresponding to l moments.

[0048] Specifically, the number of sensors is n, and each sensor obtains a one-dimensional time series data segment of length l, that is, including the first parameter corresponding to each of the l moments. The historical data is an input sequence set of size [l,n].

[0049] S2. Determine a plurality of first training samples and a fault category corresponding to each first training sample according to historical data.

[0050] Each first training sample includes a first parameter corresponding to each of y moments of the payload rotation device acquired by x sensors, x and y are positive integers, and x is less than or equal to n, and y is less than or equal to l;

[0051] In some embodiments, see Figure 3 The above S2 specifically includes the following steps:

[0052] S21. Perform normalization and denoising processing on the historical data to eliminate noise interference of the historical data, and make the value of the first parameter corresponding to each moment included in the historical data fall within a preset threshold range.

[0053] S22: Perform sampling operations on historical data to obtain a plurality of first training samples and a fault category corresponding to each first training sample.

[0054] Specifically, in the actual application process, the length of l of historical data is usually as high as hundreds of millions. If all historical data are input into the neural network model for training at one time, it will cause the computing device memory to overflow and fail to train effectively. Therefore, the method provided by the present invention performs sampling operations on historical data. In one example, first, a sliding sampling window is set, and the width of the sliding window is always consistent with the number of feature map channels v after data fusion. The length of the sliding window W is the length of the first training sample to be generated; secondly, the sliding window slides forward according to a certain step length B. After each sliding step, the original data points with a length of W and a width of v channels contained in the sliding window are the newly generated feature maps of the first training sample.

[0055] Optionally, the calculation formula for the number L of generated first training samples is:

[0056]

[0057] Among them, l represents the number of data points corresponding to each sensor in the historical data, W represents the length of the sliding window, that is, the length of the first training sample, B is the step size of the sliding window, λ represents the sampling overlap rate, and L is the number of newly generated training samples.

[0058] The method provided in the embodiment of the present invention can flexibly control the number of generated first training samples by adjusting the sampling overlap rate λ. When λ>0, it is incremental overlapping sampling for data expansion; when λ=0, it is equally divided and non-overlapping sampling; when λ<0, it is sparsely intercepted interval sampling; in addition, the width W of the sliding sampling window can be set according to the fault occurrence frequency, sampling cycle and operating speed of the rotating machinery. According to Shannon's sampling theorem, the recommended setting of W should be at least greater than or equal to two complete fault waveform cycles. The embodiment of the present invention does not impose any special restrictions on the specific implementation of the sampling overlap rate λ and the width W of the sampling window.

[0059] S3. Train the neural network model according to the multiple first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model.

[0060] In some embodiments, the neural network model includes an embedding unit and multiple stacked Transformer units, the output end of the embedding unit is connected to the input end of the Transformer unit; the embedding unit is used to convert the input first training sample into a feature vector of a preset dimension corresponding to each first training sample; each Transformer unit includes an encoder and a decoder; the output end of the encoder of each Transformer unit is connected to the output end of the decoder of each Transformer unit; each encoder includes an attention layer and a feedforward neural network layer, the attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to encode each feature vector to obtain a latent vector corresponding to each feature vector; each decoder includes a masked multi-head attention layer, an attention layer and a feedforward neural network layer, the multi-head attention layer is used to associate the input of each encoding unit with the output of each decoding unit; the attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to decode each latent vector to obtain a feature vector corresponding to each feature vector.

[0061] It can also be understood that the neural network model provided by the embodiment of the present invention is a neural network architecture with stacked Transformer units as the main trunk. The Transformer unit in the architecture consists of two parts: an encoder and a decoder. In the Transformer unit, the encoder is responsible for converting the input sequence data into a high-dimensional representation for subsequent processing. The encoder is stacked with multiple identical layers, each of which contains two main sublayers: multi-head attention and a feedforward neural network. Multi-head attention establishes an attention relationship between the glyphs at each position and the glyphs at all other positions, thereby capturing global contextual information; while the feedforward neural network maps the vectors at each position to a higher-dimensional space, which helps the model capture more complex semantic information. Each sublayer is followed by accumulation and layer normalization to stabilize training and accelerate convergence. The stacking of multiple layers of the encoder enables the model to extract and fuse the features of the input sequence layer by layer. The difference between the decoder and the encoder is that the decoder introduces masked multi-head attention to associate the input sequence with the output of the current decoding position to generate a meaningful output sequence. This structure can more effectively capture semantic information and maintain consistency when generating sequences. The Transformer-based network achieves modeling of global information and more efficient feature extraction through the self-attention mechanism, and alleviates the problems of gradient explosion and gradient vanishing during training, which greatly expands the number of parameters in the language model.

[0062] In one example, the first training sample is a two-dimensional tensor of shape (batch_size, sequence_length), where batch_size is the batch size and sequence_length is the sequence length; first, the input data is embedded through the embedding layer to convert the data into a format that the neural network model can process. Assuming the embedding dimension is embed_dim, the shape of the data becomes (batch_size, sequence_length, embed_dim); then, the data is input into the Transformer unit. A series of self-attention and feedforward network operations are performed on the input data. Assuming that n stacked Transformer units are retained, the shape of the data after passing through the Transformer unit is still (batch_size, sequence_length, embed_dim); then, the data is transformed through a linear layer. The task is to predict k time points in the future, so the output dimension of the linear layer should be set to k. Therefore, after the data passes through the linear layer, its shape becomes (batch_size, sequence_length, k); finally, the data is processed through a normalization layer. The normalization layer is usually used to normalize or otherwise transform the last dimension of the data to make it consistent with our prediction goal. For example, if our goal is to predict the value of k time points in the future, the normalization layer normalizes the last dimension of the data so that its value is between 0 and 1. Therefore, after the data passes through the normalization layer, its shape is still (batch_size, sequence_length, k), but its value has been appropriately transformed.

[0063] In one possible implementation, see Figure 4 The above S3 specifically includes the following steps:

[0064] S31. Input a plurality of first training samples into a neural network model, and output a predicted fault category corresponding to each first training sample.

[0065] S32: Establish an objective function according to the predicted fault category and the fault category corresponding to each first training sample.

[0066] S33. Iteratively train the neural network model according to the objective function to obtain a trained neural network model.

[0067] Among them, the objective function L is:

[0068]

[0069] Among them, L represents the cross entropy loss value, n represents the number of fault categories, and y irepresents the fault category of the i-th type, Represents the output of the neural network model for the i-th fault category.

[0070] S4. Acquire on-orbit operation data of the payload rotating device, where the on-orbit operation data is a first parameter corresponding to each of l moments of the payload rotating device acquired by n sensors, where l and n are positive integers.

[0071] S5. Input the on-orbit operation data into the trained neural network model, and output the fault prediction results corresponding to the on-orbit operation data, where the fault prediction results include the fault categories corresponding to the on-orbit operation data.

[0072] From the above S1-S5, it can be known that the method provided in the embodiment of the present invention obtains the historical data of the payload rotating device and trains the neural network model according to the historical data of the payload rotating device to obtain a trained neural network model, and then can accurately and timely predict the fault type of the payload rotating device based on the on-orbit operation data of the payload rotating device based on the trained neural network model. Compared with the method for predicting faults of payload rotating devices in related technologies, the accuracy and timeliness of the prediction results can be improved.

[0073] In some embodiments, see Figure 5 Before the prediction system executes the above S3, the method provided by the embodiment of the present invention further includes the following steps:

[0074] S61. Construct a neural network model;

[0075] Among them, the neural network model is the GPT-2 model.

[0076] S62, training the neural network model to obtain a pre-trained neural network model;

[0077] S63. Based on the dimension of the first training sample, adjust the model parameters of the pre-trained neural network model to obtain a neural network model.

[0078] In a possible implementation, the above S62 specifically includes the following steps:

[0079] Modal decomposition is performed on multiple first training samples to obtain second training samples corresponding to each first training sample, where the second training samples are spectrum images including multiple frequency domain features; a neural network model is trained according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

[0080] Specifically, Empirical Mode Decomposition (EMD) is an adaptive signal processing method, which is particularly suitable for the analysis of nonlinear and non-stationary signals. The EMD method does not require any basis function to be set in advance, but decomposes the vibration signal included in the first training sample according to the time scale characteristics of the data itself, and decomposes the vibration signal included in the first training sample into a finite number of intrinsic mode functions (IMFs), each of which contains local features of the original signal at different time scales. That is, it includes a spectrum image of multiple frequency domain features.

[0081] The method provided in the embodiment of the present invention trains the neural network model through the second training sample to obtain a pre-trained neural network model, thereby realizing cross-modal migration of the architecture of the neural network model, that is, the self-attention layer and feedforward neural network of the pre-trained neural network model contain most of the knowledge learned during the pre-training process, and then the model parameters of the position embedding and layer normalization layer of the pre-trained neural network model are fine-tuned for cross-modal migration based on the dimension of the first training sample. The original input embedding layer is for the second training sample, so the input and output layers are redesigned here according to the dimension of the first training sample so that the first training sample is mapped to the dimension required by the specific pre-trained model.

[0082] Furthermore, the architecture of the neural network model retains the multi-head attention and feedforward neural network in the pre-trained model. Since the multi-head attention and feedforward neural network contain most of the knowledge from the pre-trained neural network model, they are chosen to be retained during fine-tuning.

[0083] In one example, during the training of a neural network model, the neural network model learns the potential representation of sequence data through unsupervised learning. For a large language model, a potential representation of language is constructed by learning a large text corpus, so that the model learns the statistical laws and semantic information of the language. Given a sequence data S = {s1, s2, s3, ..., s n}, the model task goal is based on the previous text s1,s2,s3,...,s i-1 , by autoregressively predicting the target text s i The objective function L pre (s) is expressed as the maximum likelihood function, where

[0084] Where n is the number of shape symbols divided in the sequence data, and P(·) is the likelihood function.

[0085] Through the training of the neural network model with multiple second training samples, the neural network model has learned a lot of prior knowledge based on the corpus data. In order to enable the large prediction model to have the ability to reason about time series fault data of similar modes, it is necessary to learn the model in the cross-modal migration stage. That is, supervised learning is carried out on the mechanical fault diagnosis task, and the model parameters are fine-tuned and optimized. For the supervised diagnosis task, given the state monitoring sequence data X of the mechanical equipment = {x1, x2, ..., x n} and its corresponding fault label set Y = (y1, y2, ...y n}, the fine-tuning target of the fault diagnosis model is based on any sample x in the given sequence data i Determine the corresponding fault type i , the objective function L(x) of the process.

[0086]

[0087] The neural network model provided in the embodiment of the present invention adopts the GPT-2 model as its core component, and fine-tunes and configures the model parameters accordingly according to different task requirements (dimensions of training samples). In the initialization stage, the neural network model sets various properties such as sequence length, prediction length, patch size, step size, etc. according to the configuration parameters, and calculates the number of patches required. In addition, the model also initializes the embedding layer DataEmbedding, which is responsible for converting the input data into a format suitable for model processing (that is, converting the first training sample into a format suitable for model processing). At the same time, the model loads the pre-trained GPT-2 model (pre-trained neural network model) and selects the number of layers to be retained according to the configuration. In order to adapt to specific tasks, the model uses different linear layers and normalization layers before output. For example, in the classification task, the model uses a linear layer to convert the output of GPT-2 into the output of the number of categories, and applies activation functions and normalization. These task-specific layers enable the model to better adapt to different application scenarios. During the fine-tuning process, the model specifically retains some parameters of GPT-2, and only allows the parameters of some layers to be updated during the training process. The purpose of this is to retain the advantages of the pre-trained model while preventing overfitting during fine-tuning and speeding up the training process. If the configuration specifies the use of a GPU, the model will also transfer the GPT-2 components to the corresponding GPU device to utilize the computing power of the GPU. In general, the present invention demonstrates how to fine-tune and adapt the pre-trained GPT-2 model according to different task requirements (a first training sample of type time series data and a second training sample of type spectral image). By retaining some parameters and introducing task-specific layers, the model can adapt to specific application scenarios while maintaining the advantages of pre-training, thereby achieving effective performance in a variety of tasks.

[0088] For example, to facilitate understanding of this solution, see Figure 6 , Figure 6 The present invention provides a flow chart of a method for predicting a fault of a space payload rotating device. The prediction system 100 first obtains historical data of the payload rotating device; then the prediction system 100 determines a plurality of first training samples and a fault category corresponding to each first training sample based on the historical data; then the prediction system 100 determines a second training sample corresponding to each first training sample based on the plurality of first training samples; then the prediction system 100 trains the neural network model based on the second training sample to obtain a pre-trained neural network model. It should be understood that the first training sample is a vibration signal of a first parameter (acceleration or angular velocity) corresponding to each of a plurality of moments acquired by a plurality of sensors, and the second training sample is a spectrum image including a plurality of frequency domain features corresponding to each first training sample, and the first training sample and the second training sample have different modes.

[0089] The prediction system 100 adjusts the parameters of the pre-trained neural network model based on the first training sample (freezes the multi-head attention and feedforward neural network of the pre-trained neural network model, adjusts the embedding layer and feature processing layer of the pre-trained neural network model, so that the pre-trained neural network model can take multiple first training samples as input for further training), and obtains the adjusted neural network model. Then the prediction system 100 trains the adjusted neural network model based on multiple first training samples to obtain a trained neural network model. Finally, the prediction system 100 obtains the on-orbit operation data of the payload rotating equipment; inputs the on-orbit operation data into the trained neural network model, and outputs the fault prediction results corresponding to the on-orbit operation data, and the fault prediction results include the fault category corresponding to the on-orbit operation data. Among them, the on-orbit operation data includes the vibration signal of the first parameter (acceleration or angular velocity) corresponding to each of the multiple moments obtained by multiple sensors.

[0090] The above mainly introduces the scheme of the embodiment of the present invention from the perspective of the method. It can be understood that in order to realize the above functions, the prediction system includes at least one of the hardware structure and software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiment disclosed in this article, the embodiment of the present invention can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the embodiment of the present invention.

[0091] The embodiment of the present invention can divide the prediction system into functional units according to the above method example. For example, the prediction system can be divided into functional units corresponding to various functions, or two or more functions can be integrated into one processing unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional units. It should be noted that the division of units in the embodiment of the present invention is schematic and is only a logical functional division. There may be other division methods in actual implementation.

[0092] For example, Figure 7 A schematic diagram of the hardware structure of a prediction system provided by an embodiment of the present invention is shown. The prediction system 100 includes: a data acquisition unit 110, used to acquire historical data of a payload rotating device, the historical data being a first parameter corresponding to each of l moments of the payload rotating device acquired by n sensors, the first parameter being acceleration or angular velocity, l and n being positive integers; a sample generation unit 120, used to determine a plurality of first training samples and a fault category corresponding to each first training sample based on the historical data, wherein each first training sample includes a first parameter corresponding to each of y moments of the payload rotating device acquired by x sensors, x and y being positive integers, and x being less than or equal to n, and y being less than or equal to l; a model The training unit 130 is used to train the neural network model according to multiple first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model; the data acquisition unit 110 is also used to obtain the on-orbit operation data of the payload rotating device, the on-orbit operation data is the first parameter corresponding to each of l moments of the payload rotating device obtained by n sensors, l and n are positive integers; the fault prediction unit 140 is used to input the on-orbit operation data into the trained neural network model, and output the fault prediction result corresponding to the on-orbit operation data, the fault prediction result includes the fault category corresponding to the on-orbit operation data.

[0093] Optionally, the model training unit 130 is further used to: construct a neural network model; train the neural network model to obtain a pre-trained neural network model; and adjust model parameters of the pre-trained neural network model based on the dimension of the first training sample to obtain a neural network model.

[0094] Optionally, the model training unit 130 is specifically used to: perform modal decomposition processing on multiple first training samples to obtain second training samples corresponding to each first training sample, where the second training sample is a spectrum image including multiple frequency domain features; train the neural network model according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

[0095] Optionally, the sample generation unit 120 is specifically used to: normalize and denoise the historical data to eliminate noise interference in the historical data, and make the value of the first parameter corresponding to each moment included in the historical data fall within a preset threshold range; perform sampling operations on the historical data to obtain multiple first training samples and a fault category corresponding to each first training sample.

[0096] Optionally, the model training unit 130 is specifically used to: input a plurality of first training samples into the neural network model, and output a predicted fault category corresponding to each first training sample; establish an objective function according to the predicted fault category and the fault category corresponding to each first training sample; and iteratively train the neural network model according to the objective function to obtain a trained neural network model;

[0097] The objective function L is:

[0098]

[0099] Among them, L represents the cross entropy loss value, n represents the number of fault categories, and y i represents the fault category of the i-th type, Represents the output of the neural network model for the i-th fault category.

[0100] It should be understood that the specific description of the above optional methods can refer to the above method embodiments, which will not be repeated here. In addition, the explanation of any prediction system 100 provided above and the description of the beneficial effects can refer to the above corresponding method embodiments, which will not be repeated here.

[0101] The embodiment of the present invention further provides a computer-readable storage medium, in which at least one computer instruction is stored, and the at least one computer instruction is loaded and executed by a processor to implement the methods of the above embodiments. For the explanation of the relevant contents and the description of the beneficial effects in any of the above-mentioned computer-readable storage media, reference can be made to the above-mentioned corresponding embodiments, which will not be repeated here.

[0102] The embodiment of the present invention further provides a chip. The chip integrates a control circuit and one or more ports for implementing the functions of the above-mentioned prediction system 100. Optionally, the functions supported by the chip can be referred to above and will not be described in detail here.

[0103] Those skilled in the art will appreciate that all or part of the steps of the above embodiments can be implemented by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned storage medium can be a read-only memory, a random access memory, etc. The above-mentioned processing unit or processor can be a central processing unit, a general-purpose processor, a specific circuit structure (application specific integrated circuit, ASIC), a microprocessor (digital signal processor, DSP), a field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof.

[0104] The embodiment of the present invention also provides a computer program product including instructions, when the instructions are run on a computer, the computer executes any one of the methods in the above embodiments. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the process or function according to the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., an SSD), etc.

[0105] It should be noted that the above-mentioned devices for storing computer instructions or computer programs provided in the embodiments of the present invention, such as but not limited to the above-mentioned memory, computer-readable storage medium and communication chip, etc., all have non-transitory. Those skilled in the art should be aware that in one or more of the above-mentioned examples, the functions described in the embodiments of the present invention can be implemented with hardware, software, firmware or any combination thereof. When implemented using software, these functions can be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein the communication medium includes any medium that is convenient for transmitting a computer program from one place to another. The storage medium can be any available medium that a general or special-purpose computer can access.

[0106] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting failure of aerospace payload rotating equipment, characterized in that: The method comprises: Acquire historical data of a payload rotating device, wherein the historical data is a first parameter corresponding to each moment in l moments of the payload rotating device acquired by n sensors, wherein the first parameter is acceleration or angular velocity, and l and n are positive integers; Determine a plurality of first training samples and a fault category corresponding to each first training sample according to the historical data, wherein each first training sample includes a first parameter corresponding to each of y moments of the payload rotating device acquired by x sensors, where x and y are positive integers, and x is less than or equal to n, and y is less than or equal to 1; Training the neural network model according to the multiple first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model; Acquire on-orbit operation data of a payload rotating device, where the on-orbit operation data is a first parameter corresponding to each of l moments of the payload rotating device acquired by n sensors, where l and n are positive integers; The on-orbit operation data is input into the trained neural network model, and a fault prediction result corresponding to the on-orbit operation data is output, wherein the fault prediction result includes a fault category corresponding to the on-orbit operation data.

2. The method according to claim 1, characterized in that The determining, according to the historical data, a plurality of first training samples and a fault category corresponding to each first training sample comprises: Performing normalization and denoising processing on the historical data to eliminate noise interference of the historical data and ensure that the value of the first parameter corresponding to each moment included in the historical data is within a preset threshold range; A sampling operation is performed on the historical data to obtain a plurality of first training samples and a fault category corresponding to each first training sample.

3. The method according to claim 2, characterized in that The step of training the neural network model according to the plurality of first training samples and the fault category corresponding to each first training sample to obtain a trained neural network model includes: Inputting the plurality of first training samples into a neural network model, and outputting a predicted fault category corresponding to each first training sample; Establishing an objective function according to the predicted fault category and the fault category corresponding to each first training sample; Iteratively train the neural network model according to the objective function to obtain a trained neural network model; The objective function L is: Among them, L represents the cross entropy loss value, n represents the number of fault categories, and y i represents the fault category of the i-th type, Represents the output of the neural network model for the i-th fault category.

4. The method according to claim 3, characterized in that The neural network model includes an embedding unit and a plurality of stacked Transformer units, wherein the output end of the embedding unit is connected to the input end of the Transformer unit; the embedding unit is used to convert the input first training sample into a feature vector of a preset dimension corresponding to each first training sample; each Transformer unit includes an encoder and a decoder; the output end of the encoder of each Transformer unit is connected to the output end of the decoder of each Transformer unit; each encoder includes an attention layer and a feedforward neural network layer, the attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to encode each feature vector to obtain a latent vector corresponding to each feature vector; each decoder includes a masked multi-head attention layer, an attention layer and a feedforward neural network layer, and the multi-head attention layer is used to associate the input of each encoding unit with the output of each decoding unit; The attention layer is used to determine the context information included in each feature vector, and the feedforward neural network layer is used to decode each latent vector to obtain the feature vector corresponding to each feature vector.

5. The method according to claim 4, characterized in that Before training the neural network model according to the multiple first training samples and the fault category corresponding to each first training sample to obtain the trained neural network model, the method further includes: Build a neural network model; Train the neural network model to obtain a pre-trained neural network model; Based on the dimension of the first training sample, the model parameters of the pre-trained neural network model are adjusted to obtain a neural network model.

6. The method according to claim 5, characterized in that The training of the neural network model to obtain a pre-trained neural network model includes: Performing modal decomposition processing on the plurality of first training samples to obtain a second training sample corresponding to each first training sample, wherein the second training sample is a spectrum image including a plurality of frequency domain features; The neural network model is trained according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

7. A fault prediction system for aerospace payload rotating equipment, characterized in that: The system comprises: A data acquisition unit, used to acquire historical data of the payload rotating device, wherein the historical data is a first parameter corresponding to each moment in l moments of the payload rotating device acquired by n sensors, wherein the first parameter is acceleration or angular velocity, and l and n are positive integers; a sample generating unit, configured to determine a plurality of first training samples and a fault category corresponding to each first training sample according to the historical data, wherein each first training sample comprises a first parameter corresponding to each of y moments of the payload rotating device acquired by x sensors, where x and y are positive integers, and x is less than or equal to n, and y is less than or equal to 1; A model training unit, used to train the neural network model according to the multiple first training samples and the fault category corresponding to each first training sample, to obtain a trained neural network model; The data acquisition unit is further used to acquire on-orbit operation data of the payload rotating device, wherein the on-orbit operation data is a first parameter corresponding to each of l moments of the payload rotating device acquired by n sensors, where l and n are positive integers; A fault prediction unit is used to input the on-orbit operation data into the trained neural network model, and output a fault prediction result corresponding to the on-orbit operation data, wherein the fault prediction result includes a fault category corresponding to the on-orbit operation data.

8. The system according to claim 7, characterized in that The model training unit is further used for: Build a neural network model; Train the neural network model to obtain a pre-trained neural network model; Based on the dimension of the first training sample, the model parameters of the pre-trained neural network model are adjusted to obtain a neural network model.

9. The system according to claim 8, characterized in that The model training unit is specifically used for: Performing modal decomposition processing on the plurality of first training samples to obtain a second training sample corresponding to each first training sample, wherein the second training sample is a spectrum image including a plurality of frequency domain features; The neural network model is trained according to the second training samples corresponding to the multiple first training samples and the fault category corresponding to each second training sample to obtain a pre-trained neural network model.

10. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the fault prediction method for aerospace payload rotating equipment as described in any one of claims 1-6.

Images