Incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data
By combining incremental learning models and long short-term memory networks, spacecraft telemetry data is dynamically updated, solving the problem of high time overhead in spacecraft anomaly detection and achieving efficient and flexible anomaly detection.
Patent Information
- Application Number
- CN202510917259.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-10-21
AI Technical Summary
Existing methods for detecting anomalies in spacecraft are time-consuming and difficult to adapt to dynamically changing telemetry data, resulting in low detection efficiency and insufficient accuracy.
An incremental learning model is adopted, which combines a long short-term memory network and multiple regularization methods to dynamically update the model weights, reduce the need for retraining, and maintain sensitivity and adaptability to new and old data.
It effectively reduces the training cost and time expenditure for spacecraft anomaly detection, enhances the adaptability and detection accuracy of the model, and can identify new anomaly patterns in real time.
Smart Images
Figure CN120822141A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of aerospace technology, and in particular to an incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data. Background Art
[0002] Spacecraft anomaly detection is crucial for ensuring the operational safety and reliability of aerospace equipment. It not only enables timely detection of potential faults and prevents major accidents, but also improves the mission success rate and service life of aerospace equipment, significantly impacting national security and modern warfare. Currently, spacecraft anomaly detection can be performed using telemetry data (also known as detection data). However, this process can be time-consuming. Summary of the Invention
[0003] The present application provides an incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data, which can solve the problem of large time overhead in spacecraft anomaly detection in the prior art.
[0004] To achieve the above objectives, this application adopts the following technical solutions: In a first aspect, an incremental learning anomaly detection method for dynamically updating spacecraft telemetry data is provided. The method is based on an incremental learning model and includes: training a long short-term memory network model based on a first portion of a first set of spacecraft detection data to obtain a label for the first portion of data, where the label indicates whether the data is normal or abnormal. Training an incremental learning model based on a second set of spacecraft detection data, the first portion of data, the label of the first portion of data, and a second portion of data in the first set of detection data other than the first portion of data. The first set of detection data is data that has not yet been identified by the incremental learning model, and the second set of detection data is data that has already been identified by the incremental learning model. Based on the training results of the incremental learning model, the incremental learning model is updated to identify abnormal data in the first set of detection data.
[0005] In this technical solution, when detection data changes dynamically, the incremental learning model can gradually update the model weights through learning, without having to retrain the entire model. This incremental update approach efficiently responds to changes in detection data, reducing the training cost and time overhead of spacecraft anomaly detection. It also enhances the model's adaptability, ensuring it can effectively identify new anomaly patterns and maintain high detection accuracy. The long-short-term memory network model preserves the characteristics of anomaly samples during the incremental learning process, ensuring that learning new data does not lead to the forgetting of historical anomaly patterns, thereby improving the learning ability of the incremental learning model.
[0006] In a possible implementation of the first aspect, an incremental learning model is trained based on a second set of detection data from a spacecraft, a first portion of data and labels of the first portion of data, and a second portion of data excluding the first portion of data in the first set of detection data. The method includes: obtaining a class-balanced batch of BCB data and labels of the BCB data from the second set of detection data and the labels of the second set of detection data. Obtaining a random sample batch of BRS data and labels of the BRS data from the first portion of data and the labels of the first portion of data. Obtaining an unlabeled batch of BUD data and pseudo-labels for the BUD data from the second portion of data. Training a main classifier of the incremental learning model based on the class-balanced batch of BCB data and labels of the BCB data yields a cross-entropy loss LCB and a first regularization term LLwF1 to prevent forgetting for the main classifier. Training an auxiliary classifier of the incremental learning model based on the random sample batch of BRS data, labels of the BRS data, the BUD data, and pseudo-labels of the BUD data yields a cross-entropy loss LRS and a second regularization term LLwF2 to prevent forgetting for the auxiliary classifier. Minimizing the loss based on LCB, LLwF1, LRS, and LLwF2. Based on the loss minimization result, the incremental learning model is trained.
[0007] In a possible implementation of the first aspect, an incremental learning model is trained based on the first part of data and the label of the first part of data, as well as the second part of data other than the first part of data in the first set of detection data, including: obtaining a random sample batch of BRS data and the label of the BRS data in the first part of data and the label of the first part of data. In the second part of data, an unlabeled data batch of BUD data is obtained, and a pseudo-label of the BUD data is obtained. Based on the random sample batch of BRS data, the label of the BRS data, the BUD data, and the pseudo-label of the BUD data, an auxiliary classifier of the incremental learning model is trained to obtain the cross entropy loss LRS of the auxiliary classifier and the second regularization term LLwF2 to prevent forgetting. Based on LRS and LLwF2, the loss function of EWC is obtained based on the elastic weight integration EWC method. The loss function of EWC is minimized. Based on the result of minimizing the loss function of EWC, the incremental learning model is trained.
[0008] In one possible implementation of the first aspect, an incremental learning model is trained based on a second set of detection data from a spacecraft, the first portion of data, and labels for the first portion of data. The training includes: using normal data in the second set of detection data and the first portion of data as positive samples for information-to-noise contrast estimation (InfoNCE) loss, and using abnormal data in the second set of detection data and the first portion of data as negative samples for InfoNCE loss, to obtain the InfoNCE loss. The InfoNCE loss function is minimized, and the incremental learning model is trained based on the result of minimizing the InfoNCE loss function.
[0009] In a possible implementation manner of the first aspect, abnormal data in the first set of detection data is used for the next update of the incremental learning model.
[0010] In a possible implementation of the first aspect, the first set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power. The second set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power.
[0011] In a second aspect, an incremental learning anomaly detection device for dynamic updating of spacecraft telemetry data is provided, which includes a module for executing the method provided by the first aspect or any possible implementation of the first aspect.
[0012] In a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the processor is configured to call programs and data in the memory to execute the method provided in the first aspect or any possible implementation of the first aspect.
[0013] In a fourth aspect, a computer-readable storage medium is provided, in which a computer program or instruction is stored. When the computer program or instruction is executed, the method provided by the first aspect or any possible implementation of the first aspect is implemented.
[0014] In a fifth aspect, a computer program product is provided, which includes a computer program or instructions. When the computer program or instructions are executed by a device, the device implements the method provided by the first aspect or any possible implementation of the first aspect.
[0015] It can be understood that the devices, electronic devices, computer-readable storage media and computer program products provided in the above aspects all include the contents described in the above methods. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects of the above methods and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A schematic diagram of an incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data provided in an embodiment of the present application; Figure 2 A schematic diagram of a long short-term memory network model provided in an embodiment of the present application; Figure 3 A schematic diagram of a category incremental learning model provided in an embodiment of the present application. DETAILED DESCRIPTION
[0017] It should be noted that the terms "first" and "second" in the embodiments of this application are used only to distinguish features of the same type and should not be understood as indicating relative importance, quantity, sequence, etc. The step numbers in the embodiments of this application are used only to distinguish different steps and should not be understood as indicating relative importance, quantity, sequence, etc.
[0018] The terms "exemplary" or "for example" in the embodiments of this application are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.
[0019] First, the application scenarios of the embodiments of this application are introduced. The embodiments of this application can be applied to spacecraft anomaly detection. A spacecraft can be aerospace equipment, such as aircraft, rockets, satellites, and deep space probes. A spacecraft can also refer to electromechanical products on aerospace equipment, such as instrument panels and sensors. Spacecraft anomaly detection refers to determining whether a spacecraft is abnormal based on its detection data.
[0020] When a spacecraft is in orbit, telemetry data is dynamic and continuous. As a spacecraft continues to operate in orbit, the patterns of telemetry data may change due to equipment aging, environmental changes, or varying mission requirements. During this process, relying on static datasets for training makes it difficult to adapt to changes in data distribution. Consequently, new abnormal patterns or equipment failures cannot be detected promptly, hindering timely health monitoring and fault diagnosis. Especially in the condition monitoring of critical spacecraft components, if the model cannot adapt to the new data distribution, anomaly detection failures or lags may occur, compromising spacecraft safety and mission completion.
[0021] Current anomaly detection methods for spacecraft are typically based on fixed training sets and lack the ability to adapt to new devices or new states. Existing models often need to be retrained whenever new device types or mission modes emerge, which not only incurs significant computational overhead but can also delay the updating of detection results. In practical applications, spacecraft detection data is generated incrementally during mission operations, and device types and states change frequently, meaning that spacecraft detection data is dynamic. Therefore, the key to solving these problems is how to update and adapt to new data patterns in real time without retraining the entire model.
[0022] To address this challenge, incremental learning (IL) methods have been proposed and applied to dynamically changing data environments. Incremental learning can dynamically incorporate new data and add new knowledge to existing models without retraining the entire model, effectively avoiding the computational cost of data retraining. For spacecraft anomaly detection, IL methods can update models in real time to adapt to new data patterns, maintaining high detection accuracy, especially when new devices, sensors, or system state changes are added.
[0023] Based on this, the embodiment of the present application provides an incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data. The method is based on the incremental learning model, such as Figure 1 As shown, the method may include at least one or more of the following steps: S100: Based on the first part of the first set of detection data of the spacecraft, a long short-term memory network model is used for training to obtain a label for the first part of the data, where the label is used to indicate whether the data is normal or abnormal.
[0024] Exemplarily, the incremental learning model may be a class-incremental learning model (CIL) or other incremental learning models.
[0025] As another example, spacecraft detection data can be collected in real time. For example, the detection data can be collected by sensors on the spacecraft, and the sources of the detection data include, but are not limited to, parameters such as temperature, pressure, vibration, current, or power. The detection data of each sensor type has different characteristics, and the detection data can be labeled and classified according to the specific sensor type. The raw detection data can also be preprocessed, such as denoising and normalization. Common preprocessing methods include filtering, smoothing, interpolation, and normalization, which can ensure data quality for subsequent model training and anomaly detection.
[0026] The first set of detection data is data that has not yet been identified by the incremental learning model. This first set of detection data can be referred to as new detection data. This new detection data refers to data that has not yet been labeled by the incremental learning model when the incremental learning model is updated to the current stage (or the current incremental stage). After training the long short-term memory network model, the first portion of the new detection data is new data with known labels. The first set of detection data includes at least one of the following spacecraft data: temperature, pressure, vibration, current, or power.
[0027] As another example, a portion of the new test data is used for preliminary training to generate a basic anomaly detection model. This basic anomaly detection model uses a long short-term memory (LSTM) network to perform labeled learning on the data and identify known normal and abnormal patterns.
[0028] like Figure 2 As shown, the LSTM network can include the following four parts: (1) Forget Gate: .
[0029] in, is the sigmoid activation function, and the hidden state at the previous moment is , is the input, and are the weight and bias of the forget gate.
[0030] (2) Input gate: ; .
[0031] in, is the activation value of the input gate, is the weight of the input gate, is the bias, is a candidate memory, is the weight of the input gate, is the bias of the input gate.
[0032] (3) Memory update: .in, is the output of the forget gate, is the memory state at time t-1, is the activation value of the input gate.
[0033] (4) Output gate: ; .in, is the output of the sigmoid function of the output gate, is the weight of the output gate, is the bias of the output gate, is the hidden state at the current moment.
[0034] In this implementation, the LSTM network model preserves the characteristics of anomaly samples during the incremental learning process, thereby expanding the existing anomaly knowledge base. The LSTM network model efficiently stores and replays these anomaly samples, ensuring that learning new data does not lead to forgetting historical anomaly patterns. This not only improves the incremental learning model's learning capabilities but also enables it to adapt to changing system data, updating anomaly patterns in real time and maintaining sensitivity to both new and old data.
[0035] S200: Training an incremental learning model based on a second set of detection data of the spacecraft, the first portion of data and a label of the first portion of data, and a second portion of data excluding the first portion of data in the first set of detection data.
[0036] Exemplarily, the second set of detection data is data that has already been identified by the incremental learning model. This second set of detection data can be referred to as learned detection data. Learned detection data can be data that has been identified and labeled by the incremental learning model when the incremental learning model is updated to the current stage (or the current incremental stage). The second set of detection data includes at least one of the following spacecraft data: temperature, pressure, vibration, current, or power.
[0037] Exemplarily, the first set of detection data is divided into two parts, the first part of the data is trained using a long short-term memory network to obtain labels, and the second part of the data is not trained using a long short-term memory network to obtain labels.
[0038] For example, in an incremental learning task, assume that the new detection data acquired in each incremental stage contains m new categories, and these new categories have no overlap with the learned categories. Specifically, in the tth incremental task, the training data set of the new category can be expressed as: .in, represents the i-th sample in the t-th task, is the label vector corresponding to the sample, Represents the total number of samples in the task. The total number of categories that may be included in the t-th task is: Where m represents the number of categories added in each incremental stage, so the label vector is defined as: At stage t, the parameters of the model are , in order to better adapt to the new data, only new datasets are used in the training process To update the model. The model update process can be expressed by the following formula: .in, represents the model parameters of the previous incremental stage, is the learning rate, Indicates the data of the previous stage model in the current stage In each iteration, the model of the previous stage updates the model of the current stage, and the updated model is used for the next round of incremental learning.
[0039] For example, a regularization-based incremental learning method can be used to add a regularization term during the incremental learning phase for new device types. Distillation loss is one of the commonly used regularization terms. Knowledge distillation ensures the retention of learned knowledge by transferring the knowledge of the old model to the new model. In incremental learning, the knowledge distillation loss is defined as: .in, and In the incremental learning phase and The prediction results of the model for the old task category, N is the number of samples that the old model has learned. is the i-th sample in The prediction results on is the i-th sample in When new task data is added, Force the model and ,The output on old task categories is equal, which helps retain old knowledge.,Category incremental learning is an extension of this method.
[0040] In some examples, such as Figure 3 As shown, S200 may include at least one or more of the following steps: S211: Obtain the balanced category batch (BCB) data and the labels of the BCB data from the second set of test data and the labels of the second set of test data. Exemplarily, the BCB data may be obtained by balanced selection from the second set of detection data.
[0041] S212: Obtain random sample batch (BRS) data and labels of the BRS data from the first part of data and the labels of the first part of data.
[0042] Exemplarily, the BRS data may be obtained by randomly sampling from the first portion of data.
[0043] S213: In the second part of the data, obtain unlabeled data batch (Unlabeled Data Batch, BUD) data and obtain pseudo labels for the BUD data.
[0044] For example, the K-Nearest-Neighbors (KNN) method can be used to obtain pseudo-labels for BUD data. The KNN method can be used to predict pseudo-labels for BUD data. Using the KNN method, auxiliary data can be searched within the feature embedding space of previously stored labeled data to query unlabeled data. A BUD is formed by randomly sampling from the query unlabeled data. BUD helps the model adapt to new data by injecting knowledge about similar past categories into the model.
[0045] S214: According to the category-balanced batch BCB data and the labels of the BCB data, the main classifier of the incremental learning model is trained to obtain the cross entropy loss LCB of the main classifier and the first regularization term LLwF1 to prevent forgetting.
[0046] S215: Based on the random sample batch BRS data, the label of the BRS data, the BUD data, and the pseudo label of the BUD data, the auxiliary classifier of the incremental learning model is trained to obtain the cross entropy loss LRS of the auxiliary classifier and the second regularization term LLwF2 to prevent forgetting.
[0047] S216: Minimize the loss based on LCB, LLwF1, LRS and LLwF2.
[0048] For example, the whole process is modeled as a regularized optimization problem, where the goal is to minimize the following loss: .in , from the main classifier and the auxiliary classifier respectively, is a hyperparameter that controls the contribution of the regularization term to the unlabeled data of the query.
[0049] S217: Based on the loss minimization results, train the incremental learning model.
[0050] In this implementation, constraints are imposed on the loss function of the new task to prevent old knowledge from being overwritten by new knowledge. Regularization terms are added during the incremental learning phase for new device types to constrain model training and prevent the forgetting of old knowledge.
[0051] In some examples, S200 may further include at least one or more of the following steps: S221: Obtain random sample batch BRS data and labels of the BRS data from the first part of data and the labels of the first part of data.
[0052] S222: In the second part of data, obtain an unlabeled data batch BUD data and obtain a pseudo label for the BUD data.
[0053] S223: Based on the random sample batch BRS data, the label of the BRS data, the BUD data, and the pseudo-label of the BUD data, the auxiliary classifier of the incremental learning model is trained to obtain the cross entropy loss LRS of the auxiliary classifier and the second regularization term LLwF2 to prevent forgetting.
[0054] The processes of S221 to S223 may refer to the processes of S212, S213 and S215, and will not be further described in detail in this embodiment of the present application.
[0055] S224: Based on LRS and LLwF2, the loss function of Elastic Weight Consolidation (EWC) is obtained based on the EWC method: .in, According to LRS and LLwF2, is a hyperparameter, is the i-th parameter, is the old task parameter, is the diagonal element of the Fisher information matrix. When processing new samples, the weights of historical samples and new samples are dynamically adjusted by introducing a flexible sample importance weight mechanism.
[0056] S225: Minimize the loss function of EWC.
[0057] S226: Based on the results of minimizing the loss function of EWC, train the incremental learning model.
[0058] In this implementation, an elastic sample importance weight (EWC) mechanism is introduced to dynamically update model weights to avoid forgetting old data distributions. This elastic weighting mechanism automatically adjusts the importance weight of each sample based on the learning progress of the new task, ensuring that new data does not excessively impact the model's memory of old task data. By weighting sample importance, the model can maintain an accurate response to historical data when processing new data, effectively avoiding the catastrophic forgetting problem in incremental learning. This dynamic weighting mechanism improves the model's adaptability to different data distributions, enabling continuous and effective anomaly detection.
[0059] In some examples, S200 may further include at least one or more of the following steps: S231: The normal data in the second set of detection data and the first part of data are used as positive samples of the information noise contrast estimation (InfoNCE) loss, and the abnormal data in the second set of detection data and the first part of data are used as negative samples of the InfoNCE loss to obtain the InfoNCE loss.
[0060] For example, during the incremental learning process, a contrastive learning mechanism may be used to reduce the distribution deviation between new data and learned data.
[0061] Contrastive learning uses the InfoNCE loss Lq, which optimizes the model by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. This allows the features of multiple normal data to be similar, while making abnormal data and normal data different.
[0062] The formula is as follows: . Among them, q is the query sample, is a positive sample related to the query sample, is the temperature coefficient, k is the number of samples, is a negative sample. During the back propagation process, The parameters are updated by differentiating the weights in q and k respectively, so that the loss is lower after the weight acts on k with high similarity to q.
[0063] S232: Minimize the InfoNCE loss function.
[0064] S233: Based on the results of minimizing the InfoNCE loss function, train the incremental learning model.
[0065] In this implementation, a contrastive learning mechanism is used to reduce the distribution bias between new and old data. By comparing the feature representations of new data with those of historical data, contrastive learning enables the model to retain a good memory of past tasks while learning new ones. This mechanism reduces the bias between data distributions, ensuring that the introduction of new data does not disrupt the structure and characteristics of the old data, thus avoiding performance degradation caused by distribution differences. This mechanism enables the model to accurately detect anomalies in historical tasks while simultaneously handling new ones, maintaining the accuracy and robustness of anomaly detection.
[0066] S300: updating the incremental learning model according to the training result of the incremental learning model, and identifying abnormal data in the first set of detection data.
[0067] For example, during the incremental learning model update process, the model can perform real-time anomaly detection on new test data. For each data point, the incremental learning model assesses whether it deviates from the normal state and labels it as abnormal or normal. If the model detects a new abnormal pattern, the abnormal sample is saved to the abnormal sample library and added to the model's memory module as new knowledge. This memory module stores the abnormal sample features for subsequent learning and model updates.
[0068] For example, after each incremental learning update, the model will be evaluated on new detection data to ensure that the updated model can effectively identify new abnormal patterns. At the same time, the evaluation process will use previously stored historical data for cross-validation to ensure the accuracy and stability of the model.
[0069] For example, once the model detects an abnormal event, the system will immediately notify the relevant operator through an alarm mechanism. The alarm information includes the time, type, severity and possible cause of the abnormality.
[0070] As another example, this method can also incorporate a feedback mechanism, whereby anomaly data from the first set of detection data can be used to further update the incremental learning model. This feedback mechanism ensures that detected anomaly information is used for subsequent model optimization. When the model detects a false positive or false negative anomaly, the feedback data is fed into the model as new training data for further optimization, forming a closed self-learning loop.
[0071] In this implementation, new detection data is continuously generated as the spacecraft operates. Whenever new detection data arrives, the incremental learning model updates the model weights through incremental learning, without requiring retraining the entire model. This incremental update approach efficiently adapts to changes in data streams. Within the incremental learning framework, by continuously learning new anomaly states, the spacecraft's anomaly detection model can gradually adapt to changes in data distribution brought about by different mission phases and equipment changes. This not only significantly reduces model training costs but also enhances the model's adaptability, ensuring it can effectively identify new anomaly patterns and maintain high detection accuracy. Furthermore, the application of incremental learning models makes spacecraft anomaly detection more flexible and efficient, enabling continuous optimization of model performance within dynamic and gradually changing data streams. With the diversification of spacecraft missions and the continuous increase in equipment types, anomaly detection methods based on incremental learning can provide real-time monitoring of equipment health and fault warnings during mission execution, making their application in spacecraft telemetry data processing crucial. This algorithm can adapt to changes in data patterns in real time and effectively integrate new and old data across different time periods to improve the accuracy and real-time performance of anomaly detection.
[0072] Furthermore, based on long-short-term memory neural networks and combined with a flexible sample importance weighting mechanism, an anomaly detection model based on incremental learning can be implemented. This method can dynamically learn new data distribution patterns and reduce the deviation between the old and new data distributions through comparative learning mechanisms. This helps to address the problem that anomaly detection based on static models cannot adapt to new data distribution changes. This method is suitable for real-time anomaly detection in spacecraft telemetry data. It can detect and identify potential failure risks in real time during spacecraft operation, providing strong support for spacecraft health management.
[0073] An embodiment of the present application also provides an incremental learning anomaly detection device for dynamic updating of spacecraft telemetry data, which includes a first training module, a second training module and an identification module.
[0074] The first training module is used to train a long short-term memory network model based on a first part of the first set of detection data of the spacecraft to obtain a label for the first part of the data, where the label is used to indicate whether the data is normal or abnormal.
[0075] The second training module is configured to train the incremental learning model based on a second set of spacecraft detection data, the first portion of data and its labels, and the second portion of data excluding the first portion of data in the first set of detection data. The first set of detection data is data that has not yet been identified by the incremental learning model, and the second set of detection data is data that has already been identified by the incremental learning model.
[0076] The identification module is used to update the incremental learning model according to the training results of the incremental learning model and identify abnormal data in the first set of detection data.
[0077] In some possible embodiments, the second training module is specifically used to: obtain the class-balanced batch BCB data and the label of BCB data from the second group of detection data and the label of the second group of detection data; obtain the label of the random sample batch BRS data and the BRS data from the first part of data and the label of the first part of data; obtain the unlabeled data batch BUD data from the second part of data, and obtain the pseudo-label of BUD data; train the main classifier of the incremental learning model according to the class-balanced batch BCB data and the label of BCB data, and obtain the cross-entropy loss LCB of the main classifier and the first regularization term LLwF1 to prevent forgetting; train the auxiliary classifier of the incremental learning model according to the random sample batch BRS data, the label of BRS data, BUD data and the pseudo-label of BUD data, and obtain the cross-entropy loss LRS of the auxiliary classifier and the second regularization term LLwF2 to prevent forgetting; minimize the loss according to LCB, LLwF1, LRS and LLwF2; and train the incremental learning model based on the result of loss minimization.
[0078] In some possible implementations, the second training module is specifically used to: obtain random sample batch BRS data and labels of BRS data from the first part of data and the labels of the first part of data; obtain unlabeled data batch BUD data from the second part of data, and obtain pseudo labels of BUD data; train the auxiliary classifier of the incremental learning model based on the random sample batch BRS data, the labels of BRS data, BUD data, and the pseudo labels of BUD data to obtain the cross-entropy loss LRS and the second regularization term LLwF2 to prevent forgetting of the auxiliary classifier; obtain the loss function of EWC based on the elastic weight integration EWC method based on LRS and LLwF2; minimize the loss function of EWC; and train the incremental learning model based on the result of minimizing the loss function of EWC.
[0079] In some possible implementations, the second training module is specifically used to: use the second set of detection data and the normal data in the first part of the data as positive samples for information-noise contrast estimation InfoNCE loss, and use the second set of detection data and the abnormal data in the first part of the data as negative samples for InfoNCE loss to obtain InfoNCE loss; minimize the InfoNCE loss function; and train the incremental learning model based on the result of minimizing the InfoNCE loss function.
[0080] In some possible implementations, abnormal data in the first set of detection data is used for the next update of the incremental learning model.
[0081] In some possible implementations, the first set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power. The second set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power.
[0082] An embodiment of the present application also provides an electronic device, which includes a memory and a processor, and the processor is used to call the program and data in the memory to execute the above-mentioned incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data.
[0083] An embodiment of the present application also provides a computer-readable storage medium, which stores program code. When the computer-readable storage medium is run on a device (for example, the device can be a single-chip microcomputer, a chip, a computer or a processor, etc.), the program code therein can be called by the processor to execute one or more steps in the above-mentioned embodiment of the incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data.
[0084] Based on this understanding, the embodiments of the present application also provide a computer program product containing instructions. The technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or the processor therein to execute all or part of the steps of the method described in each embodiment of the present application.
[0085] Finally, it should be noted that the above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
Claims
1. An incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data, characterized in that: The method is based on an incremental learning model and includes: training a long short-term memory network model based on a first portion of the first set of detection data of the spacecraft to obtain a label for the first portion of the data, the label being used to indicate whether the data is normal or abnormal; training the incremental learning model based on a second set of detection data of the spacecraft, the first portion of data and the label of the first portion of data, and a second portion of data in the first set of detection data excluding the first portion of data; wherein the first set of detection data is data that has not yet been recognized by the incremental learning model, and the second set of detection data is data that has already been recognized by the incremental learning model; According to the training results of the incremental learning model, the incremental learning model is updated to identify abnormal data in the first set of detection data.
2. The method according to claim 1, characterized in that The step of training the incremental learning model based on the second set of detection data of the spacecraft, the first portion of data and the label of the first portion of data, and the second portion of data in the first set of detection data excluding the first portion of data, comprises: Obtaining class-balanced batch BCB data and the labels of the BCB data from the second set of test data and the labels of the second set of test data; Obtaining, from the first portion of data and the label of the first portion of data, random sample batch BRS data and the label of the BRS data; In the second part of data, obtain an unlabeled data batch BUD data, and obtain a pseudo label for the BUD data; Training the main classifier of the incremental learning model according to the category-balanced batch BCB data and the label of the BCB data to obtain the cross entropy loss LCB of the main classifier and the first regularization term LLwF1 to prevent forgetting; Training an auxiliary classifier of the incremental learning model according to the random sample batch BRS data, the label of the BRS data, the BUD data, and the pseudo label of the BUD data to obtain a cross entropy loss LRS and a second regularization term LLwF2 for preventing forgetting of the auxiliary classifier; Minimizing losses according to the LCB, the LLwF1, the LRS, and the LLwF2; Based on the result of minimizing the loss, the incremental learning model is trained.
3. The method according to claim 1 or 2, characterized in that The step of training the incremental learning model based on the first portion of data and the label of the first portion of data, and the second portion of data in the first set of detection data excluding the first portion of data, includes: Obtaining, from the first portion of data and the label of the first portion of data, random sample batch BRS data and the label of the BRS data; In the second part of data, obtain an unlabeled data batch BUD data, and obtain a pseudo label for the BUD data; Training an auxiliary classifier of the incremental learning model according to the random sample batch BRS data, the label of the BRS data, the BUD data, and the pseudo label of the BUD data to obtain a cross entropy loss LRS and a second regularization term LLwF2 for preventing forgetting of the auxiliary classifier; According to the LRS and the LLwF2, an EWC method is integrated based on elastic weights to obtain a loss function of the EWC; Minimize the loss function of the EWC; The incremental learning model is trained based on the result of minimizing the loss function of the EWC.
4. The method according to any one of claims 1 to 3, characterized in that The step of training the incremental learning model according to the second set of detection data of the spacecraft, the first portion of data, and the label of the first portion of data includes: Using the normal data in the second set of detection data and the first part of data as positive samples for information-noise contrast estimation (InfoNCE) loss, and using the abnormal data in the second set of detection data and the first part of data as negative samples for the InfoNCE loss, to obtain the InfoNCE loss; Minimizing the InfoNCE loss function; The incremental learning model is trained based on the result of minimizing the InfoNCE loss function.
5. The method according to any one of claims 1 to 4, characterized in that The abnormal data in the first set of detection data is used for the next update of the incremental learning model.
6. The method according to any one of claims 1 to 5, characterized in that The first set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power; The second set of detection data includes at least one of the following data of the spacecraft: temperature, pressure, vibration, current, or power.
7. An incremental learning anomaly detection method for dynamic updating of spacecraft telemetry data, characterized in that: The apparatus comprises modules for executing the method according to any one of claims 1-6.
8. An electronic device, characterized in that: The electronic device includes a memory and a processor, and the processor is configured to call programs and data in the memory to execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program or instruction, and when the computer program or instruction is executed, the method according to any one of claims 1 to 6 is implemented.
10. A computer program product, characterized in that The computer program product includes a computer program or instructions, and when the computer program or instructions are executed by a device, the device is enabled to implement the method according to any one of claims 1 to 6.