System and method for single event effect simulation and fault diagnosis based on spaceborne computer

Abstract

The application discloses a single-particle effect simulation and fault diagnosis system and method based on a spaceborne computer, which comprises a simulation host computer, a simulation engine arranged in a hardware environment of the simulation host computer, a protocol monitor and a log module, a configuration and injection manager, and a peripheral model library connected with the simulation engine, wherein the simulation engine is connected with the protocol monitor and the log module and the configuration and injection manager, and the peripheral model library is mounted with the peripheral model library; the peripheral model library is connected with an interrupt controller through a bus and an interface, and the interrupt controller is connected with the simulation engine; the protocol monitor and the log module are connected with a data set generator, the data set generator is connected with a diagnosis server, and the diagnosis server is internally provided with a dynamic-attention-temporal (DAT) hybrid neural network model. The system is used for fault type identification by using the DAT hybrid neural network based on general peripheral abstraction and discrete event semantics, and is suitable for on-orbit working condition simulation and diagnosis of multiple protocols and multiple peripherals.

Description

Technical Field

[0001] This invention relates to the field of on-orbit electronic system reliability of spacecraft, and in particular to a single-event effect simulation and fault diagnosis system and method based on a spaceborne computer. Background Technology

[0002] Single-event effects in the space radiation environment often cause bit flips, transients, and latch-up in onboard computers (OBCs), leading to attitude and mission control anomalies, and in extreme cases, system resets or failures. Traditional radiation mitigation methods (redundancy, shielding, beaming, irradiation testing) are costly, time-consuming, and difficult to cover complex on-orbit coupling behaviors. Existing system-level simulation tools have shortcomings in protocol timing, peripheral modeling, and tag system consistency, making it difficult to obtain reproducible, scalable, and on-orbit consistent diagnostic data.

[0003] On the fault diagnosis side, rule / threshold-based and traditional machine learning methods have limited ability to represent temporal features across protocols and peripherals. Existing one-dimensional CNNs or CNN+attention models can extract local patterns, but they are insufficient in capturing long-term dependencies and rare faults with low signal-to-noise ratios, and the fixed convolutional kernels result in weak feature transfer capabilities across protocol domains. Pure RNN / BiLSTM methods suffer from gradient decay and low efficiency for long sequences. Transformer-like models have strong data dependencies, high computational costs, and unstable training under short sample and strong class imbalance conditions. In addition, many methods have not shown alignment with protocol-level temporal semantics and a unified label space, making it difficult to achieve stable and interpretable diagnostic performance in multi-protocol / multi-peripheral scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a single-event effect simulation and fault diagnosis system and method based on a spaceborne computer. On the simulation side, it can uniformly describe the state machine, registers, and timing relationships of control and communication peripherals using peripheral six-tuple abstraction and discrete event semantics (DES). On the labeling side, it achieves unified labeling through position-aware cross-protocol fault mapping. On the diagnosis side, it proposes a DAT (Dynamic-Attention-Temporal) hybrid neural network, which uses dynamic convolution, attention reweighting, and bidirectional temporal coding to collaboratively model cross-protocol long-term dependencies and low signal-noise rare faults. It is suitable for on-orbit simulation and diagnosis of multiple protocols and multiple peripherals, and has significant advantages in accuracy and overall balance, demonstrating stability and robustness.

[0005] To achieve the above objectives, the present invention provides a single-event effect simulation and fault diagnosis system based on a spaceborne computer, including a simulation host for loading a peripheral model library and running a system-level simulation based on discrete event semantics. The simulation host has a simulation engine deployed in its hardware environment. The simulation engine is connected to the protocol monitor and log module, the configuration and injection manager, and is mounted with the peripheral model library.

[0006] The peripheral model library is connected to the interrupt controller via a bus and interface. The interrupt controller interfaces with the simulation engine. The interrupt controller is used to receive interrupt requests generated by each peripheral in the peripheral model library, mask and prioritize the interrupt requests, and vectorize the arbitrated interrupts into discrete interrupt events and output them to the simulation engine to participate in the discrete event scheduling of the simulation engine.

[0007] The protocol monitor and log module are connected to a dataset generator, which is connected to a diagnostic server. The diagnostic server has a built-in dynamic-attention-temporal hybrid neural network model and is used to identify faults caused by single-event effects and output fault diagnosis results.

[0008] Preferably, the simulation engine uses discrete event semantics to define event types and assigns propagation delay, logic delay, and interruption constraints to each peripheral device, as shown in the formula:

[0009] ;

[0010] in, This represents the delay between writing combinational logic and registers to the observed output. This represents the link delay from the peripheral output to the sampling point at the other end. This indicates the timing verification reference period.

[0011] Preferably, the peripheral model library includes a communication peripheral model and a control peripheral model. The communication peripheral model and the control peripheral model respectively describe the interface, state, register, function mapping, and timing semantics through a peripheral abstract six-tuple, as shown in the formula:

[0012] ;

[0013] in, For peripheral devices, discrete signals and data are read from the bus and pins via ports. For discrete signals and data ports that drive peripheral devices externally, For each state of the internal finite state machine of the peripheral device, It is a read / write register. Within an event step, in ,enter Calculate new and The mapping, The timeline for sampling and submission is determined by the sequence of events.

[0014] The peripheral model library provides register read / write callbacks, timing event callbacks, and interrupt triggering interfaces, and exposes binding points with the system bus.

[0015] Preferably, the configuration file of the configuration and injection manager includes: target device, target interface, field type, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds, and random seed;

[0016] The configuration and injection manager is used to inject faults caused by single-event effects into specified bits or bit segments of selected peripherals within a set time window with a set probability, according to the configuration file, and to maintain random seeds and round numbers.

[0017] Preferably, the protocol monitor and log module are in protocol-level temporal semantics. The system captures discrete event data output by the simulation engine and outputs a minimum sufficient record of a quadruple as a log. The formula for the quadruple is:

[0018] ;

[0019] Among them, the protocol monitor and log module output To and Aligned discrete-time scale, For peripheral instance identification, The observed signal on the peripheral device, For mapping kernel The unified set of labels for judgment The source of the elements in the tag set includes single-event effect fault tags.

[0020] Preferably, the dataset generator converts logs from the protocol monitor and log module into a dataset under a unified label set, wherein the unified label set is set as follows:

[0021] ;

[0022] The dataset generator implements a cross-protocol mapping kernel, as shown in the formula:

[0023] ;

[0024] in, For protocol or device category, For bit or field semantics.

[0025] Preferably, the dynamic-attention-temporal hybrid neural network model includes a signal branch, a device embedding branch, and a temporal embedding branch;

[0026] The signal branch includes a dynamic convolutional layer I. and dynamic convolutional layers The dynamic convolutional layer and the dynamic convolutional layer Each of these is subsequently connected to a maximum pooling layer. and max pooling layer The maximum pooling layer and the maximum pooling layer Each of these is subsequently connected to a randomly deactivated layer. and random deactivation layer The random deactivation layer I is connected to the dynamic convolutional layer II; the random deactivation layer II is connected to a bidirectional temporal coding module, and the bidirectional temporal coding module is connected to an attention aggregation module;

[0027] The device embedding branch includes a device embedding mapping layer, which is used to map a device identifier vector to a device embedding vector of a preset dimension. The input of the device embedding mapping layer is connected to a device input layer. The device input layer is used to receive a device identifier vector representing the source of a peripheral device or protocol.

[0028] Temporal embedding branch: includes a temporal embedding mapping layer, used to map the time scalar into a temporal embedding vector of a preset dimension, the input of the temporal embedding mapping layer is connected to a temporal input layer; the temporal input layer is used to receive normalized time scalar features;

[0029] The attention aggregation module, the device embedding mapping layer, and the time embedding mapping layer are all connected to the fusion discrimination layer, and the fusion discrimination layer is connected to the output layer.

[0030] The dynamic-attention-temporal hybrid neural network model is used to identify weak amplitude and sparse fragmented fault characteristics caused by single-event effects, and the attention aggregation module enhances the weak signal response of single-event effect faults.

[0031] Preferably, both the dynamic convolutional layer I and the dynamic convolutional layer II are conditional convolutional kernels that adaptively generate weights based on samples or operating conditions, in order to mitigate distribution drift across protocols and peripherals caused by single-event failures.

[0032] Preferably, it also includes an error assessment component for quantifying systematic errors, using the following formula:

[0033] ;

[0034] in, For systematic error, For standard caliber error, To account for alignment and windowing errors, This refers to device mapping error.

[0035] A method for single-event effect simulation and fault diagnosis based on a spaceborne computer is also provided, the steps of which include:

[0036] S1. Set the parameters of the peripheral model library and the simulation engine on the simulation host, load the peripheral model library, and bind the bus and the interrupt controller;

[0037] S2. Set the target device, target interface, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds and random seed of the configuration and injection manager, and enable the cross-protocol mapping core;

[0038] S3. The simulation engine executes the simulation, defining event types according to discrete event semantics, and satisfying... ;

[0039] S4. Enable the protocol monitor and log module, export the minimum sufficient records and save the configuration summary and random seed, use the dataset generator to convert the logs into a dataset under a unified label set, and divide the dataset into training set, validation set and test set;

[0040] S5. Train the dynamic-attention-temporal hybrid neural network model in the diagnostic server and output the single-event effect fault diagnosis results.

[0041] Therefore, the present invention adopts the above-mentioned single-event effect simulation and fault diagnosis system and method based on spaceborne computer, which can automatically generate hundreds of thousands of protocol-level logs and extract tens of thousands of usable data samples for analysis and training within about 20 minutes, significantly shortening the R&D cycle, improving the efficiency of diagnosis and design iteration, and greatly reducing the cost of testing and data acquisition while ensuring the accuracy of timing.

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0043] Figure 1 This is a simulation flowchart of the present invention;

[0044] Figure 2 This is a schematic diagram illustrating communication peripheral fault types according to an embodiment of the present invention;

[0045] Figure 3 This is a comparison chart of ablation experiment results from an embodiment of the present invention;

[0046] Figure 4This is a structural diagram of the DAT neural network model according to an embodiment of the present invention;

[0047] Figure 5 This is a confusion matrix diagram of the DAT model and the CNN, CNN+ATT, and CNN+ATT+BiLSTM models in this embodiment of the invention.

[0048] Figure 6 The above are t-sne diagrams of the DAT model and the CNN, CNN+ATT, and CNN+ATT+BiLSTM models in this embodiment of the invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. In the description of the present invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0050] Example

[0051] This invention provides a single-event effect simulation and fault diagnosis system based on a spaceborne computer, such as... Figure 1 As shown, the system includes a simulation host for loading peripheral model libraries and running system-level simulations based on discrete event semantics. A simulation engine is deployed within the host's hardware environment. The simulation engine uses discrete event semantics to define event types and assigns propagation delay, logic delay, and interrupt constraints to each peripheral device, as shown in the formula:

[0052] ;

[0053] in, This represents the delay between writing combinational logic and registers to the observed output. This represents the link delay from the peripheral output to the sampling point at the other end. This indicates the timing verification reference period.

[0054] The simulation engine establishes connections with the protocol monitor and log module, as well as the configuration and injection manager. The configuration file of the configuration and injection manager includes: target device, target interface, field type, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds, and random seed.

[0055] The configuration and injection manager is used to inject faults caused by single-event effects into specified bits or bit segments of selected peripherals within a set time window with a set probability, based on the configuration file, and to maintain random seeds and round numbers.

[0056] The simulation engine is equipped with a peripheral model library, which includes communication peripheral models and control peripheral models. The communication and control peripheral models are described using a peripheral abstraction six-tuple to represent the interface, state, registers, function mapping, and timing semantics, respectively. The formula is as follows:

[0057] ;

[0058] in, For peripheral devices, discrete signals and data are read from the bus and pins via ports. For discrete signals and data ports that drive peripheral devices externally, For each state of the internal finite state machine of the peripheral device, It is a read / write register. Within an event step, in ,enter Calculate new and The mapping, The timeline for sampling and submission is determined by the sequence of events.

[0059] The peripheral model library provides register read / write callbacks, timing event callbacks, and interrupt triggering interfaces, and exposes binding points to the system bus.

[0060] The peripheral model library is connected to the interrupt controller via a bus and interface. The interrupt controller interfaces with the simulation engine. The interrupt controller is used to receive interrupt requests generated by each peripheral in the peripheral model library, mask and prioritize the interrupt requests, and vectorize the arbitrated interrupts into discrete interrupt events and output them to the simulation engine to participate in the discrete event scheduling of the simulation engine.

[0061] Protocol monitor and log module in protocol-level time semantics The system captures discrete event data output by the simulation engine and outputs a minimum sufficient record of a quaternion as a log. The formula for the quaternion is:

[0062] ;

[0063] Among them, the output of the protocol monitor and log module To and Aligned discrete-time scale, For peripheral instance identification, The observed signal on the peripheral device, For mapping kernel The unified set of labels for judgment The elements in the tag set are sourced from single-event effect fault tags.

[0064] The protocol monitor and log module are connected to a dataset generator. The dataset generator converts the logs from the protocol monitor and log module into a dataset under a unified label set, which is set as follows:

[0065] ;

[0066] The dataset generator implements a cross-protocol mapping kernel, as shown in the formula:

[0067] ;

[0068] in, For protocol or device category, For bit or field semantics.

[0069] The dataset generator connects to the diagnostic server, which has a built-in dynamic-attention-temporal hybrid neural network model. The diagnostic server identifies faults caused by single-event effects and outputs fault diagnosis results. The dynamic-attention-temporal hybrid neural network model includes a signal branch, a device embedding branch, and a time embedding branch.

[0070] The signal branch includes dynamic convolutional layers. and dynamic convolutional layers Dynamic convolutional layer I and dynamic convolutional layer All are conditional convolutional kernels, with weights adaptively generated based on samples or operating conditions, used to mitigate distribution drift across protocols and peripherals caused by single-event failures. Dynamic convolutional layers. and dynamic convolutional layers Then, they are connected to the max pooling layer I and the max pooling layer respectively. Max pooling layer and max pooling layer Subsequently connected to random deactivation layer I and random deactivation layer The random deactivation layer I is connected to the dynamic convolutional layer II; the random deactivation layer II is connected to a bidirectional temporal coding module, and the bidirectional temporal coding module is connected to an attention aggregation module.

[0071] The device embedding branch includes a device embedding mapping layer, which maps device identifier vectors to device embedding vectors of a preset dimension. The input of the device embedding mapping layer is connected to a device input layer. The device input layer is used to receive device identifier vectors representing the source of peripherals or protocols.

[0072] Temporal embedding branch: includes a temporal embedding mapping layer, which maps time scalars to time embedding vectors of a preset dimension; the input of the temporal embedding mapping layer is connected to a time input layer; the time input layer is used to receive normalized time scalar features.

[0073] The attention aggregation module, device embedding mapping layer, and time embedding mapping layer are all connected to the fusion discrimination layer, which in turn is connected to the output layer.

[0074] A dynamic-attention-temporal hybrid neural network model is used to identify weak-amplitude and sparse fragmented fault characteristics caused by single-event effects. The attention aggregation module enhances the weak signal response of single-event effect faults. The system also includes an error evaluation component to quantify system errors, using the following formula:

[0075] ;

[0076] in, For systematic error, For standard caliber error, To account for alignment and windowing errors, This refers to device mapping error.

[0077] The standard caliber error evaluation unit is used to compare the fault injection truth value recorded by the configuration and injection manager with the observation logs output by the protocol monitor and log module to obtain the standard caliber error. ;

[0078] The alignment and windowing error evaluation unit compares the fault injection timestamp with the start and end times of the sample window extracted by the dataset generator to obtain the alignment and windowing error. ;

[0079] The device mapping error evaluation unit is used to associate peripheral instance identifiers with cross-protocol mapping cores. The mapping results are subjected to a consistency check to obtain the device mapping error. ;

[0080] This invention also provides a method for single-event effect simulation and fault diagnosis based on a spaceborne computer, the specific steps of which include:

[0081] S1. Based on C#, complete the behavioral-level modeling of the target peripherals and build a peripheral model library adapted to the system: Write the behavioral logic and register interface of the communication peripheral model and the control peripheral model in C#. Both models are implemented through the unified peripheral abstract six-tuple. Describe the interface, state, registers, function mapping, and timing semantics (where, For discrete signals and data read from peripherals, For discrete signals and data ports that drive peripheral devices externally, These are the states of the internal finite state machine. It is a read / write register. This represents the input-state-output mapping relationship within an event step. (To establish the time base and event sequence for sampling and submission); Implement register read / write callbacks, timing event callbacks, and interrupt trigger interfaces according to a unified convention, expose the binding points between peripherals and the system bus, and complete the bus interface connection between the peripheral model library and the interrupt controller to ensure that the interrupt controller can interface with the subsequent simulation engine.

[0082] S2. System Component Deployment and Bus Binding: The peripheral model library built in S1 is connected to the Renode simulation platform as a component. The peripheral model library, interrupt controller, and simulation engine are bound through the system bus and interface. This ensures that the bus and interface are connected to the simulation host, simulation engine, configuration and injection manager, protocol monitor and log module, dataset generator, diagnostic server, and interrupt controller, respectively, providing support for component collaboration. The simulation engine is loaded into the simulation host's hardware environment, and its discrete event semantic parameters are configured. Event types (including clock, frame boundary, register read / write, DMA, interrupt, and fault injection) are defined, and propagation delay, logical delay, and interrupt constraints are assigned to each peripheral to ensure compliance with... (in, The delay for writing combinational logic and registers to the observed output, For link latency, (This is the timing verification reference period).

[0083] S3. Configure fault injection, cross-protocol mapping, and data acquisition parameters: In the configuration and injection manager's configuration file, specify the object to be injected according to the target device and target interface, and fill in the field type, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds, and random seed; generate the injection script and load it when the platform starts, so as to automatically attach the single-event effect-specific fault injection task when the simulation starts; enable the cross-protocol mapping core. (in For protocol or device category, (For bit or field semantics), synchronously enable the protocol monitor and log module, and configure them for protocol-level time semantics. Capture discrete event data and output quaternions. The minimum sufficient record (where, This is the aligned discrete-time scale. For peripheral instance identification, The observed signal on the peripheral device, For mapping kernel The unified set of labels for judgment The elements in the tag set include single-event effect fault tags.

[0084] S4. Simulation and Data Acquisition: Start the simulation engine, schedule various events according to discrete event semantics, and strictly follow... The simulation is subject to time constraints; single-event effect fault injection is automatically performed during the simulation process; the protocol monitor and log module collect and export minimum sufficient records in real time, and save the configuration summary and random seed simultaneously to ensure that the experiment can be replayed and reproduced; after the simulation, the target dataset is extracted from the log through the data extraction script.

[0085] Dataset standardization: Utilizing a dataset generator, through cross-protocol mapping kernels The collected heterogeneous protocol logs are converted into a unified tag set. The standardized dataset is divided into training, validation, and test sets, and the errors in the data transformation process (including standard caliber errors) are quantified. Alignment and windowing errors Device mapping error ).

[0086] S5. Train the dynamic-attention-temporal hybrid neural network model in the diagnostic server, such as... Figure 4 As shown, the training process includes:

[0087] Through the dynamic convolutional layer in the signal branch and dynamic convolutional layers Feature extraction is performed on the input signal sequence, and conditional convolutional kernels are used in all dynamic convolutional layers; max pooling layers are then applied sequentially after each dynamic convolutional layer. and max pooling layer The process involves downsampling and preserving local pattern features, followed by regularization through random deactivation layer I and random deactivation layer II. Subsequently, a bidirectional temporal coding module (bidirectional LSTM) is introduced to model the contextual temporal dependence of the downsampled sequence, and an attention aggregation module performs weighted summarization of the temporal features to obtain a global representation of the signal branch.

[0088] In the device embedding branch, the device input layer receives a device identifier vector representing the source of a peripheral device or protocol, which is then mapped by the device embedding mapping layer to a device embedding vector of a preset dimension.

[0089] In the temporal embedding branch, the temporal input layer receives normalized temporal scalar features, which are then mapped by the temporal embedding mapping layer into a temporal embedding vector of a preset dimension.

[0090] Finally, the signal branch representation, device embedding vector, and time embedding vector are concatenated and fused. The probability distribution of the fault category is output through the fused fully connected layer and the random deactivation layer, and the single-event effect fault diagnosis result is given by the output layer.

[0091] Example 1

[0092] To systematically quantify the contribution of each module in the dynamic-attention-temporal hybrid neural network model of this invention, ablation experiments were conducted, and the results are as follows: Figure 3 As shown in Table 1, the experiment clearly reveals the independent value of each module and the indispensable synergistic enhancement effect among them.

[0093] Table 1 Comparison data of DAT model with other models

[0094]

[0095] Accuracy was approximately 0.9194, an improvement of about 0.9 to 1.2 percentage points compared to the second-best model; Recall was approximately 0.8067, generally 4 to 8 percentage points higher than the second-tier model; and F1 score was approximately 0.8509, about 3 to 7 percentage points higher than the second-best results.

[0096] Example 2

[0097] To verify the classification accuracy and class balance of the DAT (Dynamic-Attention-Temporal) hybrid neural network of this invention in multi-protocol, multi-peripheral single-particle effect scenarios, a comparative experiment was conducted based on a unified label dataset generated by the simulation host, and the fault diagnosis confusion matrix of each model on the test set was plotted for evaluation.

[0098] Through discrete event semantics System-level simulation generates protocol-level logs, which are then mapped via a cross-protocol mapping core. Labeled to a unified tag set Communication peripherals are the primary targets of single-event effect faults, and their specific fault types include... Figure 2 As shown, it covers the core fault modes corresponding to the unified label set. The dataset is divided into training set / validation set / test set = 70% / 15% / 15%, and stratified sampling is used to keep the proportion of each category consistent. Under the same data splitting and training configuration, CNN, CNN+ATT, CNN+ATT+BiLSTM and the DAT model of this invention are trained and compared in sequence.

[0099] The results are as follows Figure 5As shown, the main diagonal elements of DAT are more concentrated overall, and off-diagonal confusion is significantly reduced, indicating that it has higher consistency in identifying various types of faults and better cross-protocol generalization ability. (Very few classes) (Only 72 real samples, accounting for approximately 0.74% of the entire test set) were almost all misclassified in the CNN. The recall rate was only about 2.78% (2 / 72); after adding attention, the recall rate increased to 84.72% (61 / 72), but there are still a large number of other categories. False alarms lead to low accuracy; DAT further... The recall rate improved to 94.44% (68 / 72), and false alarms were essentially eliminated, demonstrating a strong ability to detect weak-amplitude sparse faults. For medium-sized sample classes... , The recall rates of DAT increased to approximately 91.76% and 92.82%, respectively, which are further improvements compared to CNN (approximately 82.06% and 90.13%) and CNN+ATT (approximately 86.76% and 90.93%).

[0100] Example 3

[0101] To further verify the separability and cross-protocol stability of the DAT model of this invention for single-event effect fault features from the perspective of feature space, high-dimensional feature representations of the fusion layers of each model were extracted on the test set, and t-SNE was used to reduce the dimensionality to two dimensions for visualization and comparative evaluation. The feature extraction layer extracts feature vectors from the output of the fusion discriminant layer (fully connected feature layer) of each model. The t-SNE parameters are: n_components=2, perplexity=30, learning rate=200, random seed=42. All control models use the same set of t-SNE hyperparameters to ensure comparability.

[0102] The results are as follows Figure 6 As shown, the feature distributions of CNN and CNN+ATT exhibit significant inter-class overlap and intra-cluster discrepancies, especially... The significant overlap of fault classes in the left region indicates that the fixed convolutional kernel is insufficient for representing cross-protocol temporal differences. While the CNN+ATT+BiLSTM model enhances some temporal dependencies compared to the attention-only model, the overall distribution still exhibits a mixed and blurred boundary appearance, with limited improvement in clustering weak signal minority classes. In the two-dimensional projection of DAT, samples of the same class form more compact clusters, while the distance between different classes is greater and the boundaries are clearer; among these, the minority classes... With major fault classes The separation is most obvious, with Figure 5 middle The phenomenon of a significant decrease in false alarms is consistent. The results indicate that the condition-adaptive feature extraction of dynamic convolution, the long dependency modeling of bidirectional temporal coding, and the weak signal reweighting of attention can synergistically improve the feature separability and robustness under cross-protocol / cross-peripheral conditions.

[0103] The t-SNE distribution is entirely enclosed in the control model; across protocols / peripherals and weak signal subsets, the performance of DAT degrades the least with increasing difficulty, demonstrating higher stability and robustness.

[0104] Therefore, this invention employs the aforementioned single-event effect simulation and fault diagnosis system and method based on a spaceborne computer, targeting single-event effect scenarios on spaceborne computers, emphasizing practicality and high throughput: it can automatically generate hundreds of thousands of lines of protocol-level logs and extract tens of thousands of usable data samples for analysis and training within approximately 20 minutes. This capability significantly shortens the R&D cycle, improves the efficiency of diagnosis and design iteration, and substantially reduces the cost of experimentation and data acquisition while ensuring timing fidelity.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A single-event effect simulation and fault diagnosis system based on a spaceborne computer, characterized in that: It includes a simulation host for loading peripheral model libraries and running system-level simulations based on discrete event semantics. The simulation host has a simulation engine deployed in its hardware environment. The simulation engine is connected to the protocol monitor and log module, the configuration and injection manager, and is mounted with the peripheral model library. The peripheral model library is connected to the interrupt controller via a bus and interface. The interrupt controller interfaces with the simulation engine. The interrupt controller is used to receive interrupt requests generated by each peripheral in the peripheral model library, mask and prioritize the interrupt requests, and vectorize the arbitrated interrupts into discrete interrupt events and output them to the simulation engine to participate in the discrete event scheduling of the simulation engine. The protocol monitor and log module are connected to a dataset generator, which is connected to a diagnostic server. The diagnostic server has a built-in dynamic-attention-temporal hybrid neural network model and is used to identify faults caused by single-event effects and output fault diagnosis results. The peripheral model library includes communication peripheral models and control peripheral models. The communication peripheral model and the control peripheral model respectively describe the interface, state, register, function mapping, and timing semantics through a peripheral abstract six-tuple, as shown in the formula: ； in, For peripheral devices, discrete signals and data are read from the bus and pins via ports. For discrete signals and data ports that drive peripheral devices externally, For each state of the internal finite state machine of the peripheral device, It is a read / write register. Within an event step, in ,enter Calculate new and The mapping, Protocol-level time semantics; The peripheral model library provides register read / write callbacks, timing event callbacks, and interrupt triggering interfaces, and exposes binding points with the system bus. The dynamic-attention-temporal hybrid neural network model includes a signal branch, a device embedding branch, and a temporal embedding branch; The signal branch includes a dynamic convolutional layer I and a dynamic convolutional layer II. Each dynamic convolutional layer I and dynamic convolutional layer II is followed by a max pooling layer I and a max pooling layer II, respectively. Each max pooling layer I and max pooling layer II is followed by a random deactivation layer I and a random deactivation layer II, respectively. The random deactivation layer I is connected to the dynamic convolutional layer II. The random deactivation layer II is connected to a bidirectional temporal coding module, and the bidirectional temporal coding module is connected to an attention aggregation module. The device embedding branch includes a device embedding mapping layer, which is used to map a device identifier vector to a device embedding vector of a preset dimension. The input of the device embedding mapping layer is connected to a device input layer. The device input layer is used to receive a device identifier vector representing the source of a peripheral device or protocol. Temporal embedding branch: includes a temporal embedding mapping layer, used to map the time scalar into a temporal embedding vector of a preset dimension, the input of the temporal embedding mapping layer is connected to a temporal input layer; the temporal input layer is used to receive normalized time scalar features; The attention aggregation module, the device embedding mapping layer, and the time embedding mapping layer are all connected to the fusion discrimination layer, and the fusion discrimination layer is connected to the output layer. The dynamic-attention-temporal hybrid neural network model is used to identify weak amplitude and sparse fragmented fault characteristics caused by single-event effects, and the attention aggregation module enhances the weak signal response of single-event effect faults.

2. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: The simulation engine uses discrete event semantics to define event types and assigns propagation delay, logic delay, and interruption constraints to each peripheral device, as shown in the formula: ； in, This represents the delay between writing combinational logic and registers to the observed output. This represents the link delay from the peripheral output to the sampling point at the other end. This indicates the timing verification reference period.

3. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: The configuration file of the configuration and injection manager includes: target device, target interface, field type, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds, and random seed; The configuration and injection manager is used to inject faults caused by single-event effects into specified bits or bit segments of selected peripherals within a set time window with a set probability, according to the configuration file, and to maintain random seeds and round numbers.

4. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: The protocol monitor and log module are based on protocol-level temporal semantics. The system captures discrete event data output by the simulation engine and outputs a minimum sufficient record of a quadruple as a log. The formula for the quadruple is: ； Among them, the protocol monitor and log module output To and Aligned discrete-time scale, For peripheral instance identification, The observed signal on the peripheral device, For mapping kernel The unified set of labels for judgment The source of the elements in the tag set includes single-event effect fault tags.

5. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: The dataset generator converts logs from the protocol monitor and log module into a dataset under a unified label set, wherein the unified label set is set as follows: ； The dataset generator implements a cross-protocol mapping kernel, as shown in the formula: ； in, For protocol or device category, For bit or field semantics.

6. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: Both the dynamic convolutional layer I and the dynamic convolutional layer II are conditional convolutional kernels that adaptively generate weights based on samples or operating conditions, in order to mitigate distribution drift across protocols and peripherals caused by single-event failures.

7. The single-event effect simulation and fault diagnosis system based on a spaceborne computer according to claim 1, characterized in that: It also includes an error assessment component for quantifying systematic errors, using the following formula: ； in, For systematic error, For standard caliber error, To account for alignment and windowing errors, This refers to device mapping error.

8. A method for single-event effect simulation and fault diagnosis based on a spaceborne computer, based on the single-event effect simulation and fault diagnosis system based on a spaceborne computer as described in any one of claims 1-7, characterized in that the steps... include: S1. Set the parameters of the peripheral model library and the simulation engine on the simulation host, load the peripheral model library, and bind the bus and the interrupt controller; S2. Set the target device, target interface, bit or bit range, injection probability or occurrence rate interval, injection time window, injection level, number of rounds and random seed of the configuration and injection manager, and enable cross-protocol mapping core; S3. The simulation engine executes the simulation, defining event types according to discrete event semantics, and satisfying... ; S4. Enable the protocol monitor and log module, export the minimum sufficient records and save the configuration summary and random seed, use the dataset generator to convert the logs into a dataset under a unified label set, and divide the dataset into training set, validation set and test set; S5. Train the dynamic-attention-temporal hybrid neural network model in the diagnostic server and output the single-event effect fault diagnosis results.

Images