Radiation hardening method for target detection network of space spacecraft ai chip

Abstract

The application discloses a space spacecraft AI chip-oriented target detection network anti-radiation reinforcement method and belongs to the technical field of neural network anti-radiation. The method comprises the following steps: obtaining a target detection network model; respectively adopting Dropout enhancement and exponential bit quantization optimization models; and adopting the optimized model to perform target detection in a space environment. The application can not only significantly improve the robustness of a neural network in a space radiation environment, but also maintain the reasoning accuracy and speed of the model, solves the reliable operation problem of a deep learning model in a space environment, and optimizes the stability and task success rate of a spacecraft intelligent system. The application proposes a training-reasoning collaborative optimization strategy for the single event upset problem caused by high-energy particle radiation in a space environment, and provides reliable protection for AI application on a spacecraft.

Description

Technical Field

[0001] This invention relates to the field of neural network radiation hardening technology, and more specifically to a radiation hardening method for target detection networks in AI chips for spacecraft. Background Technology

[0002] Single-event effects (SEEs) caused by high-energy particle radiation in the space environment are one of the major challenges faced by spacecraft electronic systems. Among these, single-event upsets (SEUs) are the most common and widespread type of effect, causing random bit flips in the data stored in the SRAM memory of chips. With the widespread application of artificial intelligence technology in the aerospace field, especially the application of deep neural network models in tasks such as target detection and image recognition, the reliability issues of artificial intelligence in the space environment are becoming increasingly prominent.

[0003] High-energy particles in the space environment mainly originate from solar radiation, galactic cosmic rays, and charged particles trapped by the Earth's magnetic field. When these high-energy particles strike semiconductor devices, they generate electrical charges inside the devices, thus interfering with the normal operation of electronic equipment. Historical cases show that single-event effects have a very serious impact on spacecraft. For example, the US UOSAT-2 satellite and China's Fengyun-1 (B) meteorological satellite both experienced system malfunctions due to SEU events.

[0004] Current methods for protecting against single-event effects mainly include hardware hardening (such as triple-modulus redundancy design, periodic scrubbing methods, and radiation hardening devices) and software hardening (such as error-correcting coding, redundant execution, and algorithm-level fault tolerance). However, these traditional methods have significant limitations when applied to complex neural networks such as YOLOv8: hardware hardening methods are resource-intensive and costly; while software hardening methods often significantly increase processing latency or storage requirements, making them unsuitable for resource-constrained spacecraft systems, especially for real-time applications such as target detection.

[0005] With the development of deep learning technology, object detection networks such as the YOLO series have been widely used in critical missions such as autonomous navigation and target recognition in spacecraft. When these neural networks are deployed in the space environment, single-event flips can have a serious impact on them. When high-energy particles hit the SRAM cells that store the weights of the neural network, it can cause bit flips in the weight values, thus affecting the inference accuracy of the neural network. In particular, when the single-event flip rate is high, the network performance may drop sharply or even fail completely.

[0006] Therefore, how to provide a neural network anti-single-event flip method, device, and storage medium that can solve the above problems for the space environment is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, this invention provides a radiation hardening method for target detection networks in spacecraft AI chips, effectively solving the single-event flip problem faced by neural networks in the space environment. This invention combines Dropout enhancement and exponential bit quantization techniques to provide a radiation hardening method for target detection networks in spacecraft AI chips. By analyzing the impact of the single-event flip effect on neural networks in the space radiation environment, a training-inference co-optimization strategy is constructed. Based on this strategy, a radiation hardening method is designed to improve the radiation resistance of the neural network, solving the reliability problem of target detection networks in the space radiation environment. Thus, software hardening technology can effectively protect AI chips. This invention improves the robustness of target detection neural networks in the space radiation environment, ensuring the stable operation of spacecraft intelligent systems.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A radiation hardening method for target detection networks in spacecraft AI chips, based on a software hardening technology platform for spacecraft target detection AI chips, includes the following steps:

[0010] (1) Obtain the target detection neural network model; collect the structural information of the target detection neural network model and determine the key layers that need to be reinforced;

[0011] This invention takes the YOLOv8 object detection network as an example. By analyzing the YOLOv8 network structure, it identifies the detection head as the most critical and susceptible to single-event upsets (SEORs). The detection head includes a bounding box prediction branch (cv2) and a class prediction branch (cv3), which directly affect the accuracy and reliability of object detection. Experimental analysis shows that when the weights of these critical layers are affected by SEORs, it leads to a significant decrease in detection accuracy, and even detection failure. Therefore, this invention identifies these critical layers as key areas requiring reinforcement.

[0012] Specifically, the bounding box prediction branch includes convolutional layers cv2[0], cv2[1], and cv2[2], while the category prediction branch includes convolutional layers cv3[0], cv3[1], and cv3[2]. The weight parameters of these convolutional layers have a direct impact on the detection results and are key components for reinforcement.

[0013] (2) Construct a training-inference co-optimization strategy; determine to use a combination of Dropout enhancement and exponential bit quantization to harden the neural network model against radiation;

[0014] This invention constructs a training-inference co-optimization strategy, which combines Dropout enhancement technology in the model training stage with exponential bit quantization technology in the model deployment stage to form a complete radiation hardening solution.

[0015] During the training phase, Dropout enhancement techniques are used to enable the model to learn more robust feature representations, increasing the model's tolerance to weight perturbations. Dropout forces the network to learn more dispersed feature representations by randomly "dropping" a portion of neurons during training, thereby reducing dependence on individual weights and improving resistance to single-event flips.

[0016] During the deployment phase, exponent quantization technology is used to reduce the impact of single-event flips on model weights. A 32-bit floating-point number in the IEEE 754 standard consists of 1 sign bit, 8 exponent bits, and 23 mantissa bits, with the exponent bit having the most significant impact on the numerical value. By compressing the 8 exponent bits into a 4-bit representation, the drastic changes in weight values ​​caused by single-event flips can be effectively reduced, while maintaining the model's inference accuracy.

[0017] This training-inference co-optimization strategy fully leverages the complementarity of the two technologies, forming a comprehensive radiation hardening solution.

[0018] (3) Dropout is used to enhance and optimize the target detection neural network model;

[0019] A. Identify the key convolutional layers responsible for bounding box prediction in the object detection neural network model, and add a Dropout layer after them;

[0020] B. Identify the key convolutional layers responsible for class prediction in the object detection neural network model, and add Dropout layers after them;

[0021] C. Set the inactivation probability p of the Dropout layer to a preset value;

[0022] This invention enhances the radiation resistance of the YOLOv8 model by adding a Dropout layer after the key convolutional layers. The specific implementation steps are as follows:

[0023] First, a single-particle flip was simulated to identify the sensitive layers, revealing the key convolutional layers in the YOLOv8 detector head responsible for bounding box prediction, namely the first convolutional layers of cv2[0], cv2[1], and cv2[2]. These convolutional layers are responsible for predicting the position and size of the target, directly affecting the accuracy of the detection results. Adding a Dropout layer after these convolutional layers can reduce the model's dependence on specific weights and improve its robustness to weight perturbations.

[0024] Secondly, the key convolutional layers responsible for class prediction in the YOLOv8 detection head were identified, namely the first convolutional layer of cv3[0], cv3[1], and cv3[2]. These convolutional layers are responsible for predicting the class of the target and have an important impact on the accuracy of the detection results. Adding a Dropout layer after these convolutional layers can enable the model to learn more distributed feature representations and reduce the dependence on a single neuron.

[0025] Finally, the inactivation probability p of the Dropout layer was set to 0.1. Extensive experiments verified that when p=0.1, the model's radiation resistance can be maximized while maintaining its detection accuracy. An excessively high inactivation probability leads to a decrease in model performance, while an excessively low probability fails to effectively improve radiation resistance.

[0026] The core idea of ​​Dropout enhancement is to randomly "drop" a portion of neurons during training, forcing the network to learn more dispersed and redundant feature representations. This reduces reliance on individual weights and improves resistance to single-particle flips. When high-energy particles in the space environment cause some weights to flip, the overall performance does not significantly decrease because the model does not overly rely on any single weight.

[0027] (4) The target detection neural network model is optimized using exponential bit quantization;

[0028] A. Compress the exponent bits in the 32-bit floating-point numbers of neural network model parameters into a smaller representation;

[0029] B. The following formula is used for exponent quantization calculation:

[0030]

[0031] in, For the original exponent bits, and These are the minimum and maximum values ​​of the exponent, respectively. This refers to the quantized bit width.

[0032] C. Reconstruct the floating-point number after exponent quantization using the following formula:

[0033]

[0034]

[0035] in, For the sign bit, For the last digit, This is the exponent after dequantization;

[0036] This invention employs exponential bit quantization technology to optimize neural network model parameters, effectively reducing the impact of single-particle flips on model weights. The specific implementation steps are as follows:

[0037] First, the 8-bit exponent in the 32-bit floating-point number of the neural network model parameters is compressed into a 4-bit representation. The IEEE 754 standard 32-bit floating-point number consists of 1 sign bit, 8 exponent bits, and 23 mantissa bits, with the exponent bit having the most significant impact on the numerical value. When the exponent bit flips, the weight values ​​may change drastically, leading to a sharp decline in model performance.

[0038] The choice to compress the exponent to 4 bits instead of other bit widths was based on a trade-off of extensive experimental results. Experiments showed that while compressing the exponent to 2 bits provided strong radiation resistance, it also resulted in a significant decrease in model accuracy. Compressing to 6 bits maintained good model accuracy, but the improvement in radiation resistance was limited. Compressing to 4 bits, however, maintained high model accuracy (with an accuracy decrease of no more than 3%) while significantly improving radiation resistance (with an approximately 65% ​​increase in resistance to single-event upsets).

[0039] The specific quantification process is as follows: First, statistically analyze the distribution of the exponents of all weights in the model and determine the minimum value of the exponent. (Usually around 115) and maximum value (Typically around 140). Then, a linear quantization formula is used to compress and map the 8-bit exponent (range 0-255) to a 4-bit representation (range 0-15). For example, if the original exponent E of a weight is 127 (meaning 2^0 = 1), the minimum value... (=115, maximum value) =140, then the quantized value is:

[0040]

[0041] In this way, the exponent 127, which originally required 8 bits to represent, is compressed into a value of 7 that only requires 4 bits. When the exponent of this weight is affected by a single-event upset in the chip, even if an upset occurs, its value range is limited to 0-15, instead of the original 0-255 range, greatly reducing the possibility of drastic changes in the weight value.

[0042] During model inference, the quantized exponent bits need to be converted back to floating-point numbers. Continuing the example above, if the quantized exponent bits... =7, then the exponent after dequantization is:

[0043]

[0044] Since the exponent must be an integer, in actual implementation, it will be... The result is 127. Then, according to the IEEE 754 floating-point representation, the complete floating-point weight value is reconstructed by combining the sign bit S and the mantissa bit M.

[0045] The core idea of ​​exponential bit quantization is to reduce the drastic changes in weight values ​​caused by single-particle flips by limiting the representation range of the exponent bits. When high-energy particles in the space environment cause the exponent bits of certain weights to flip, the change in weight values ​​is also limited to a small range because the representation range of the exponent bits is restricted to a small interval (such as the 4-bit representation in this invention, with a value range of only 0-15), thereby maintaining the overall performance of the model.

[0046] Experimental results show that the YOLOv8 model with 4-bit exponential quantization performs well in simulated space radiation environments (single-particle flip rate is 100%). The detection accuracy decreased by only 5% under the same conditions (days), while the detection accuracy of the model without quantization technology decreased by as much as 35%, which fully demonstrates the effectiveness of the technology.

[0047] (5) Deploy the optimized target detection neural network model onto the AI ​​chip of the spacecraft to perform target detection in the space radiation environment.

[0048] The YOLOv8 model, enhanced with Dropout and optimized with exponential bit quantization, will be deployed to an AI chip on a spacecraft for target detection in the radiation environment of space. During deployment, the model needs to be converted to a format suitable for the target AI chip, and the Dropout layer must be correctly handled during the inference phase (typically by removal or inactivation). Simultaneously, it must be ensured that the parameters of the exponential bit quantization are correctly applied to maintain the model's radiation resistance.

[0049] In the radiation environment of space, the optimized model can effectively resist the effects of single-event upsets and maintain stable target detection performance. Even under high radiation doses, the model's detection accuracy remains within an acceptable range, providing reliable support for spacecraft autonomous navigation, target recognition, and other missions.

[0050] Experimental results show that the YOLOv8 model optimized using the method of this invention improves the detection accuracy by more than 15% and reduces the false detection rate by more than 20% in a simulated space radiation environment compared to the unoptimized model, significantly improving the reliability and stability of the model in the space environment.

[0051] A radiation hardening device for a target detection network for AI chips in spacecraft includes: a memory, a processor, and a target detection network radiation hardening program stored in the memory and executable on the processor. The target detection network radiation hardening program is configured to implement the aforementioned target detection network radiation hardening method for AI chips in spacecraft.

[0052] A computer-readable storage medium stores a target detection network radiation hardening program, which, when executed by a processor, implements the aforementioned target detection network radiation hardening method for AI chips in spacecraft.

[0053] Compared with the prior art, the present invention has the following obvious advantages and beneficial effects:

[0054] (1) This invention addresses the single-particle flip problem caused by high-energy particle radiation in the space environment. Based on the characteristics of neural network structure, it adopts a method combining Dropout enhancement and exponential bit quantization to harden the target detection network against radiation, which has the characteristics of good real-time performance, good stability and high accuracy. This saves spacecraft from the current process of adopting expensive hardware hardening schemes or complex triple redundancy design, making it more economical and efficient.

[0055] (2) This invention employs a training-inference co-optimization strategy, combining Dropout enhancement technology in the training phase with exponential bit quantization technology in the deployment phase to form a comprehensive software-level radiation hardening solution. This co-optimization strategy fully utilizes the complementarity of the two technologies, improving the model's tolerance to weight perturbations while reducing drastic weight value changes caused by single-event flips, thus providing a better protection method for neural network applications in the space environment. This hardening structure also has certain reference value for other deep learning models besides object detection networks.

[0056] (3) This invention adds a Dropout layer after the key convolutional layer of the YOLOv8 model, enabling the model to learn more dispersed and redundant feature representations, reducing reliance on individual weights and improving the model's radiation resistance. Simultaneously, by compressing the 8-bit exponent of the weights into a 4-bit representation, it effectively reduces the drastic changes in weight values ​​caused by single-particle flips, maintaining the model's inference accuracy. This dual protection mechanism significantly improves the reliability and stability of the object detection network in the space environment.

[0057] Special note: This invention is for ease of description only and uses radiation hardening of the YOLOv8 object detection network. Similarly, this invention can also be applied to radiation hardening of other deep learning models, such as convolutional neural networks (CNN), recurrent neural networks (RNN), and transformers. As long as the principle of this invention is used for hardening, it should fall within the scope of this invention. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the bit-flipping module implementation provided in an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram illustrating the dropout function provided in an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the detection head structure with added Dropout layer provided in an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of exponential bit quantization provided in an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of a neural network anti-single-event flip method for a space environment provided in an embodiment of the present invention. Detailed Implementation

[0064] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] See Figure 5 As shown in the figure, this invention discloses a neural network anti-single-event flip method for space environments, including: obtaining a YOLOv8 target detection network model; optimizing the YOLOv8 model using Dropout enhancement and exponential bit quantization respectively; and performing target detection in the space environment using the optimized YOLOv8 model.

[0066] In one specific embodiment, high-energy particle radiation in the space environment can cause single-event upsets (SEIs) in semiconductor devices, causing bits in memory cells to change from 0 to 1 or from 1 to 0, severely impacting the reliability of neural network models. When SEIs occur in memory storing neural network parameters, they can lead to abnormal changes in model parameters, resulting in errors in inference results. This error can have serious consequences, especially in critical tasks such as object detection.

[0067] To address the severe threat to the reliability of neural networks posed by single-event upsets caused by high-energy particle radiation in the space environment, this invention proposes a radiation-resistant optimization scheme for neural networks based on Dropout enhancement and exponential bit quantization. This scheme effectively improves the robustness of the model under radiation conditions while maintaining detection accuracy without significant impact. This scheme requires no special hardware support and is suitable for resource-constrained space applications, providing a reliable technical guarantee for the deployment of deep learning technology in the aerospace field.

[0068] The main idea is to deeply analyze the three major influence mechanisms of single-particle flip on neural network parameters (abnormal changes in parameter values, differences in network structure sensitivity, and cumulative radiation effect) and the key layer in the model that is most sensitive to radiation, and then propose a dual-protection anti-radiation optimization scheme by adopting a training-inference co-optimization strategy.

[0069] Firstly, to address the high sensitivity of the YOLOv8 detection head to single-particle flipping, a precise Dropout mechanism is introduced during the training phase. Dropout layers are added after the key convolutional layers in the bounding box prediction branch (cv2) and the class prediction branch (cv3), forcing the network to learn redundant feature representations and reducing its dependence on single neurons.

[0070] Secondly, to address the issue of drastic parameter value changes caused by exponent flipping in the IEEE 754 floating-point format, an innovative exponent quantization technique is adopted in the inference stage to compress the 8-bit exponent into a 4-bit representation, effectively limiting the destructive range of single-event flips and maximizing radiation resistance while maintaining model performance.

[0071] Single-event upset (SEE) simulation is a crucial step in evaluating the radiation resistance of neural networks. We employ the IEEE 754 floating-point standard for accurate simulation, representing the neural network parameters in binary format, and then performing targeted bit flipping. Specifically, we parse a 32-bit floating-point number into a binary sequence of 1 sign bit, 8 exponent bits, and 23 mantissa bits, implemented using a bit flipping module. The implementation process of this module is as follows: Figure 1 As shown in the figure. Through this simulation method, we can comprehensively evaluate the impact of bit flipping at different positions on the performance of neural networks.

[0072] The impact of single-event flip on neural networks can be described by the following mathematical model: For the original model parameters W, after a single-event flip, the parameters become... For input The original output is The output after flipping is The output change caused by the flip can be expressed as:

[0073]

[0074] Through sensitivity analysis, we found that the detection head part of the YOLOv8 model is particularly sensitive to single-event flip, especially the convolutional layers of the bounding box prediction branch (cv2) and the class prediction branch (cv3).

[0075] Dropout is a powerful regularization technique, such as Figure 2 As shown, by randomly deactivating a portion of neurons during training, we prevent network overfitting and improve generalization ability. Traditional Dropout is mainly used in fully connected layers; we innovatively apply it to key layers of convolutional neural networks to improve the network's robustness to single-particle flipping.

[0076] From a mathematical perspective, the working principle of Dropout can be represented as follows: for the input feature map... After convolutional layer The feature map is obtained after processing. After adding Dropout, the output feature map becomes:

[0077]

[0078] in It is a with Binary mask matrices of the same shape, each element having a probability It is set to 0 independently. This indicates element-wise multiplication.

[0079] During training, this random deactivation mechanism forces the network to learn more distributed and redundant feature representations, avoiding over-reliance on single neurons, thus achieving an "ensemble learning" effect. During inference, all neurons participate in computation, but their outputs are determined according to their deactivation probabilities. Scaling is applied to maintain the desired output:

[0080]

[0081] Based on the sensitivity analysis results, we designed a targeted Dropout enhancement strategy. For example... Figure 3As shown, we add Dropout layers at key locations in the YOLOv8 detector head, including after the first convolutional layer of cv2[0], cv2[1], and cv2[2] in the bounding box prediction branch (cv2), and after the first convolutional layer of cv3[0], cv3[1], and cv3[2] in the class prediction branch (cv3), and after the first convolutional layer of cv3[0], cv3[1], and cv3[2] in the class prediction branch (cv3). This precise Dropout configuration can maximize the resistance to single-particle flip while maintaining model accuracy. From the perspective of network structure, the bounding box prediction branch and class prediction branch of the original YOLOv8 detector head can be represented as:

[0082]

[0083]

[0084] After adding Dropout, the structure becomes:

[0085]

[0086]

[0087] in The layer index is (0, 1, 2). This indicates the output of the previous layer.

[0088] Dropout enhancement strategies improve the robustness of networks to single-event flips (SIFs) through multiple mechanisms. First, by randomly deactivating the network, it is forced to learn multiple paths to extract and process the same features, resulting in redundant representations. When a SIF causes some parameters to become abnormal, other paths can still provide correct feature representations. Second, Dropout reduces the network's dependence on specific neurons, thus reducing the impact of abnormal changes in a single parameter on overall performance. This can be mathematically expressed as: for parameters... Changes Its impact on output satisfy:

[0089]

[0090] in This indicates the effect of parameter changes on the output when Dropout is not used. Furthermore, Dropout is equivalent to training the output during the training process. 1 different subnetwork (of which (This refers to the number of neurons that may be inactivated). These subnetworks share parameters during inference, forming an ensemble effect. When some parameters are affected by single-event flips, other subnetworks can still provide correct predictions. Finally, Dropout essentially introduces noise into the network, training it to adapt to random variations in parameters and activation values. This adaptability makes the network more tolerant of parameter perturbations caused by single-event flips.

[0091] We can formalize Dropout's improvement in robustness to single-event flips as follows: for the set of flip parameters Define robustness index for:

[0092]

[0093] in It is the number of test samples. and These are the model outputs with the original parameters and the flipped parameters, respectively. Experiments show that after using Dropout enhancement, at the same flipping ratio, The value increased significantly.

[0094] Dropout failure probability This is a key hyperparameter that needs to be balanced between robustness and model performance. We determined the optimal value through extensive experiments. Value configuration: Both the bounding box prediction branch (cv2) and the class prediction branch (cv3) are configured as follows. This configuration improves radiation resistance while minimizing the impact on model accuracy. We also found that, with... As the value increases, the model's robustness to single-particle flips initially increases and then decreases, exhibiting an inverted U-shaped relationship. This can be explained as: too small a value results in... The value cannot provide sufficient redundancy, while an excessively large value... An excessively low value will weaken the network's feature extraction ability.

[0095] Exponent quantization is a dedicated quantization method designed for the IEEE 754 floating-point format. In the IEEE 754 standard, a 32-bit single-precision floating-point number consists of 1 sign bit (S), 8 exponent bits (E), and 23 mantissa bits (M). Its value is calculated using the following formula:

[0096]

[0097] The core idea of ​​exponential quantization is to reduce the precision of the exponent, thereby narrowing the range of parameter value changes that may be caused by single-event flips, while maintaining the basic functionality of the model.

[0098] Specifically, our proposed exponent quantization method compresses the original 8-bit exponent representation into a 4-bit representation, achieved through the following steps:

[0099] First, set the original floating-point parameters. Decomposed into sign bit Exponent and last digits

[0100] Quantization mapping of the exponent $E$: Where b=4 is the quantized bit width.

[0101] Dequantization restoration:

[0102] Reconstruct the quantized floating-point number:

[0103] This quantization method can be represented as a mapping function. ,satisfy:

[0104]

[0105] in This is the upper bound of the quantization error. Through theoretical analysis and experimental verification, we have demonstrated that this quantization method can significantly reduce the destructiveness of single-particle flips while maintaining model accuracy.

[0106] like Figure 4 As shown, when a single-event flip occurs in the quantized exponent, its impact is limited to a relatively small range. For the original 8-bit exponent representation, a single-bit flip can cause parameter values ​​to change by as much as [amount missing]. The maximum change is limited to times; however, in 4-bit quantization, the maximum change is limited to times. Within a factor of 1, the destructiveness of the flip is significantly reduced. This can be expressed mathematically as:

[0107]

[0108] in Represents the original parameters In the The value after the bits are flipped Indicates quantization parameters In the The value after the bit is flipped.

[0109] Our innovative approach combines Dropout enhancement with exponential bit quantization to form a training-inference co-optimization strategy. During training, Dropout enhancement enables the network to learn more robust feature representations; during inference, exponential bit quantization reduces the destructive impact of single-particle flips. This dual-protection mechanism significantly improves the reliability and stability of neural networks in the space radiation environment.

[0110] To comprehensively evaluate the effectiveness of our approach, we constructed a test framework simulating single-event flip under space radiation conditions. This framework can precisely control the scale, location, and mode of the flip, simulating the single-event flip effect under radiation environments of varying intensities. Results obtained through numerous comparative experiments are shown in Table 1.

[0111] Table 1. Performance Comparison Analysis of YOLOv8n (Raw vs Dropout vs Exponential Quantization vs Co-optimization)

[0112] Flip rate Base_mAP@50 Dropout_mAP@50 Exponential bit quantization_mAP@50 Collaborative optimization_mAP@50 Performance improvement Protective effect Baseline 0.4565 0.4482 0.4510 0.4495 -1.53% - 1e-5 0.4567 0.4480 0.4530 0.4520 -1.03% +0.79% 1e-4 0.4477 0.4387 0.4490 0.4460 -0.38% +1.63% 5e-4 0.3026 0.3326 0.3580 0.3780 +24.92% +26.74% 1e-3 0.0049 0.0187 0.0350 0.0520 +961.22% +963.04%

[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A neural network-based anti-single-event flip method for space radiation environment, based on a software hardening technology platform for spacecraft target detection AI chips, characterized in that... Includes the following steps: (1) Obtain the target detection neural network model; collect the structural information of the target detection neural network model and determine the key layers that need to be reinforced; (2) Construct a training-inference co-optimization strategy; We determined that a combination of Dropout enhancement and exponential bit quantization would be used to harden the neural network model against radiation. (3) Dropout is used to enhance and optimize the target detection neural network model; A. Identify the key convolutional layers responsible for bounding box prediction in the object detection neural network model, and add a Dropout layer after them; B. Identify the key convolutional layers responsible for class prediction in the object detection neural network model, and add Dropout layers after them; C. Set the inactivation probability p of the Dropout layer to a preset value; (4) The target detection neural network model is optimized using exponential bit quantization; A. Compress the exponent bits in the 32-bit floating-point numbers of neural network model parameters into a smaller representation; B. The following formula is used for exponent quantization calculation: Where E is the original exponent. and These are the minimum and maximum values ​​of the exponent bits, respectively, and b is the quantized bit width; C. Reconstruct the floating-point number after exponent quantization using the following formula: Where S is the sign bit and M is the mantissa. This is the exponent after dequantization; (5) Deploy the optimized target detection neural network model onto the AI ​​chip of the spacecraft to perform target detection in the space radiation environment.

Images