Parameter fault-tolerant protection method and device with zero additional memory overhead, equipment and medium

Abstract

The application relates to a zero additional memory overhead parameter fault tolerance protection method and device, equipment and medium, and relates to the technical field of satellite remote sensing. According to an exponential redundancy coding strategy, a redundant copy of to-be-protected exponential information and corresponding check information are re-embedded and mapped to a low bit area of a mantissa field to generate reconstructed protection parameters. A two-way even check mechanism is used to determine the credibility of the to-be-protected exponential information and the redundant copy of the protection parameters, and corresponding exponential recovery logic is executed according to the determination result to obtain a target exponential domain. The low bit area of the mantissa field is zeroed or truncated to recombine a target mantissa field, the target exponential domain, a sign bit and the target mantissa field are combined to obtain recovered neural network floating point parameters. The application is compatible with an existing training process, effectively detects and corrects bit flip errors on key bits, and guarantees the numerical stability and original performance of a model during on-orbit operation without additional storage overhead.

Description

Technical Field

[0001] This application relates to the field of satellite remote sensing technology, specifically to a parameter fault-tolerant protection method, apparatus, device, and medium with zero additional memory overhead. Background Technology

[0002] The space environment is filled with a large amount of high-energy particle radiation. This radiation often causes single-event effects in the memory of artificial intelligence (such as remote sensing models) running on satellite payloads, leading to bit flips in model parameters. When critical parameters of the model undergo bit flips, the performance of the remote sensing model deteriorates significantly, severely impacting the processing accuracy and reliability of satellite remote sensing data.

[0003] To address the aforementioned issues and improve the fault tolerance of models, mainstream model fault tolerance methods in recent years have mainly included quantization pruning, fault-tolerant training, and error detection. These methods have achieved some success in model guard position flipping tasks. Despite the advanced performance of these existing technologies, their practical application in satellite remote sensing models in the space environment still faces the following key challenges: First, the protection upper limit of quantization and pruning methods is insufficient. Although the fault tolerance of the model is greatly improved after pruning and quantization, the quantized model still has "critical bits". Once these critical bits suffer bit flip errors, it will still cause a devastating loss to the performance of the remote sensing model, making it unable to meet the high reliability requirements of the extreme environment of space.

[0004] Secondly, the engineering feasibility of fault-tolerant training methods is low. Although introducing corresponding algorithms during the training phase can enhance the model's robustness to errors to some extent, satellite remote sensing models typically have a huge number of parameters, and retraining or fine-tuning requires massive computational resources and time costs, which is usually not feasible in practical aerospace engineering applications.

[0005] Finally, existing error detection and correction methods have limited error correction capabilities. While common error detection methods can effectively detect errors, their accompanying error correction strategies (such as setting the error parameter to zero or limiting the activation value range) are too simplistic and crude, failing to accurately restore the true value of the parameter. They are often unable to cope with the complex and random multi-bit flipping situations in the space environment, making it difficult to fundamentally protect the model's critical parameters.

[0006] In summary, there is an urgent need to propose a fault-tolerant protection scheme that can accurately locate and effectively correct complex multi-bit flip errors without changing the original remote sensing model structure or consuming a lot of resources for retraining, thereby effectively protecting the key parameters of the satellite remote sensing model. Summary of the Invention

[0007] In view of this, this application provides a parameter fault-tolerant protection method, apparatus, device, and medium with zero additional memory overhead. The main purpose is to solve the problem that existing quantized models, when critical bits are subjected to bit flip errors, still suffer devastating performance losses and cannot meet the high reliability requirements of the extreme space environment. Satellite remote sensing models typically have a huge number of parameters, and retraining or fine-tuning requires massive computational resources and time costs. Common error detection methods and their corresponding error correction strategies are too simple and crude, unable to accurately restore the true values ​​of parameters, and often unable to cope with the complex and random multi-bit flip situations in the space environment, making it difficult to fundamentally protect the critical parameters of the model.

[0008] Firstly, this application provides a parameter fault tolerance protection method with zero additional memory overhead, including: Obtain the floating-point parameters of the neural network model to be deployed in the radiation environment, and decompose the floating-point parameters of the neural network into the sign bit, exponent field and mantissa field according to the IEEE 754 standard; Based on the parameter size of the neural network model, the corresponding exponential redundancy coding strategy is adaptively selected; According to the exponential redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, redundant copies of the exponential information to be protected and the corresponding verification information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters. The exponential information to be protected is the low-order exponential fragment or the complete exponential bit string of the exponential field, and the total number of bits of the protection parameters is the same as that of the neural network floating-point parameters. Before model inference or during periodic checks, the information to be protected, the redundant copy, and their respective verification information are parsed from the protection parameters. The credibility of the information to be protected and the redundant copy is determined by a dual-path even parity check mechanism. Based on the determination result, the corresponding index recovery logic is executed to obtain the target index domain. The low-order region of the tail number segment is cleared or truncated to reconstruct the target tail number segment. The target exponent field, the sign bit, and the target tail number segment are combined to obtain the recovered neural network floating-point parameters.

[0009] Secondly, this application provides a parameter fault-tolerant protection device with zero additional memory overhead, comprising: The acquisition module is used to acquire the neural network floating-point parameters of the neural network model to be deployed in the radiation environment, and to decompose the neural network floating-point parameters into the sign bit, exponent field and mantissa field according to the IEEE 754 standard. The selection module is used to adaptively select the corresponding exponential redundancy coding strategy based on the parameter size of the neural network model. The generation module is used to re-embed and map the redundant copy of the exponent information to be protected and the corresponding verification information to the low-order region of the tail number segment according to the exponent redundancy coding strategy, without increasing the total number of bits of the floating-point parameters of the neural network, to generate reconstructed protection parameters. The exponent information to be protected is the low-order exponent fragment or the complete exponent bit string of the exponent field, and the total number of bits of the protection parameters is the same as that of the floating-point parameters of the neural network. The execution module is used to parse the information to be protected, the redundant copy and their respective verification information from the protection parameters before model inference or during periodic checks, determine the credibility of the information to be protected and the redundant copy through a dual-path even parity check mechanism, and execute the corresponding index recovery logic according to the determination result to obtain the target index domain. The combination module is used to clear or truncate the low-order region of the tail number segment to reconstruct the target tail number segment, and combine the target exponent field, the sign bit and the target tail number segment to obtain the recovered neural network floating-point parameters.

[0010] Thirdly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor, when executing the computer program, implements the parameter fault-tolerant protection method with zero additional memory overhead described in the first aspect.

[0011] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the parameter fault-tolerant protection method with zero additional memory overhead described in the first aspect.

[0012] By employing the above technical solutions, this application provides a parameter fault-tolerant protection method, apparatus, device, and medium with zero additional memory overhead. Compared with existing technologies, this application can obtain the neural network floating-point parameters of a neural network model to be deployed in a radiation environment, and according to IEEE... The 754 standard decomposes the floating-point parameters of a neural network into a sign bit, an exponent field, and a mantissa segment. Based on the parameter size of the neural network model, it adaptively selects the corresponding exponent redundancy coding strategy. According to the exponent redundancy coding strategy, without increasing the total number of bits in the neural network floating-point parameters, redundant copies of the exponent information to be protected and the corresponding check information are re-embedded and mapped to the low-order region of the mantissa segment to generate reconstructed protection parameters. The exponent information to be protected is a low-order exponent fragment or a complete exponent string, and the total number of bits in the protection parameters is the same as that of the neural network floating-point parameters. Before model inference or during periodic checks, the exponent information to be protected, the redundant copy, and their respective check information are parsed from the protection parameters. A dual-path even parity check mechanism is used to determine the credibility of the exponent information to be protected and the redundant copy, and the corresponding exponent recovery logic is executed based on the determination result to obtain the target exponent field. The low-order region of the mantissa segment is cleared or truncated to reconstruct the target mantissa segment. The target exponent field, sign bit, and target mantissa segment are combined to obtain the recovered neural network floating-point parameters.

[0013] By employing the aforementioned technical solution, this application completely breaks through the passive protection limits of traditional quantization pruning methods and model-based redundancy methods by "re-embedding redundant copies of the exponent information to be protected and corresponding verification information and mapping them to the low-order region of the tail number segment." This application innovatively recognizes that in IEEE 754 floating-point numbers, the exponent field is the "absolute critical bit" that has the greatest impact on the numerical scale. Without altering the model structure, it proactively extracts the most vulnerable exponent information for dual backup and verification protection; coupled with a "dual-path even parity check mechanism," it can accurately reconstruct the exponent field. This means that even if the most critical bit in the model suffers a devastating radiation attack, this application can accurately restore it through internally embedded redundant copies, raising the protection limit from the dimension of underlying floating-point semantics and meeting the high reliability requirements of the extreme environment of space.

[0014] This application achieves complete "plug-and-play" protection through a parameter fault-tolerant protection method with zero additional memory overhead. From disassembling floating-point parameters to generating reconstructed protection parameters, and then to obtaining the restored floating-point parameters, the entire process is a purely low-level binary bit-level encoding and decoding process. It does not require touching the forward / backward propagation logic of the neural network, modifying the network structure, or consuming any computing power for retraining or fine-tuning. Before launch or while in orbit, the satellite only needs to execute the bit conversion routine of this application once on the pre-trained weight data to obtain radiation resistance, perfectly overcoming the fatal flaw of unacceptable time and computing power costs in fault-tolerant training in aerospace engineering.

[0015] Instead of simply discarding erroneous data during the parsing phase, this application utilizes pre-embedded verification information to determine credibility through a "dual-path even-check mechanism." This mechanism can restore the original numerical appearance of the parameters to the greatest extent possible, avoiding neuron death or sudden loss of network features caused by "setting parameters to zero," thus fundamentally protecting the inference accuracy of the model.

[0016] This application "squeezes" redundant information of the exponent into and maps it to the "lower bit region of the mantissa field" without increasing the total number of bits of the floating-point parameters of the neural network. This takes advantage of the inherent characteristic of the neural network model that is not sensitive to the precision of the lower bits of the mantissa (i.e., the model itself has a certain noise tolerance), and trades the slight loss of precision in the lower bits of the mantissa for the extremely high importance protection of the exponent field, achieving "zero increase" in storage overhead.

[0017] By adaptively selecting the corresponding exponential redundancy coding strategy based on the parameter scale, a "one-size-fits-all" approach is avoided, achieving an optimal balance between protection strength and accuracy loss.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a parameter fault-tolerant protection method with zero additional memory overhead provided in this application embodiment; Figure 2 A structural diagram of a method for zero-additional-storage redundancy protection provided in an embodiment of this application; Figure 3 This is a schematic diagram of a parameter fault-tolerant protection device with zero additional memory overhead, provided in an embodiment of this application. Detailed Implementation

[0022] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0023] The following description, with reference to the accompanying drawings, describes a parameter fault-tolerant protection method, apparatus, device, and medium with zero additional memory overhead according to embodiments of this application.

[0024] This application provides a parameter fault-tolerant protection method, device, equipment, and medium with zero additional memory overhead. Its main purpose is to solve the problem of parameter bit flipping in satellite remote sensing models caused by high-energy particle radiation in the space environment. Without requiring any retraining, fine-tuning, or structural modification of the original remote sensing model, and with full compatibility with existing training processes, it can detect and correct the specific location of multi-bit flipping, thereby protecting the critical bits of key parameters and maintaining the original performance of the model without degradation.

[0025] like Figure 1 As shown, embodiments of this application provide a parameter fault-tolerant protection method with zero additional memory overhead, including: Step 101: Obtain the neural network floating-point parameters of the neural network model to be deployed in the radiation environment, and decompose the neural network floating-point parameters into the sign bit, exponent field and mantissa field according to the IEEE 754 standard.

[0026] This application relates to a zero-storage-overhead redundancy protection technology that utilizes the internal bit semantic structure of floating-point parameters to achieve parameter redundancy embedding, error detection, and recovery. This solution is applicable to spaceborne remote sensing models, on-orbit intelligent interpretation models, multimodal remote sensing large models, and other deep neural networks deployed in radiation-sensitive environments. Unlike traditional solutions that achieve fault tolerance through additional parameter duplication, setting independent error correction regions, or employing triple-modal redundancy, this application does not rely on external storage blocks. Instead, it directly mines usable redundancy within the representation structure of the floating-point parameters themselves. By collaboratively recombining the exponent field and the tail number segment, it achieves the detection and recovery of radiation-induced bit-flip errors without increasing the overall parameter storage scale, thus balancing fault tolerance and storage efficiency in extremely resource-constrained spaceborne scenarios.

[0027] The technical starting point of this application is a systematic statistical analysis of the distribution patterns of parameters in remote sensing mission depth models. By statistically analyzing all parameters of typical remote sensing models, a highly consistent and engineering-valuable empirical prior can be discovered: the absolute values ​​of the vast majority of parameters are strictly less than 2. This phenomenon indicates that, under standard training paradigms, remote sensing model parameters generally exhibit a small-amplitude, concentrated distribution. For IEEE 754 floating-point numbers, when the absolute value of a parameter is less than 2, its most significant bit (MSB) in the exponent field exhibits a stable value pattern. This statistical prior can be further interpreted as follows: in most conventional neural network parameters, the MSB can be considered a naturally redundant bit, providing a structural basis for embedding protective information under zero-additional storage conditions.

[0028] Based on the aforementioned prior knowledge, this application proposes a Zero-Memory-Overhead Redundancy Protection (ZMORP) method. Its basic idea is to re-embed redundant copies of the exponent information and corresponding parity information into the original floating-point bits of the parameters without expanding the parameter word length, increasing model storage footprint, or adding a new independent protection array. Since the mantissa segment is inherently robust to small perturbations to neural network parameters, this application prioritizes using the lower bits of the mantissa as a redundancy container and establishes consistency constraints on the original exponent segment and the redundant exponent segment using even parity checking. Thus, when a single bit or a small number of sparse bits in the parameter are flipped, recovery logic can be executed before inference or during periodic checks to reconstruct the exponent field, thereby avoiding numerical explosion, inter-layer error propagation, and a sharp drop in inference performance caused by damage to high-weight bits. Figure 2 As shown, the mechanism is explained in its entirety.

[0029] Based on this, in the implementation of small- to medium-scale neural networks, this application utilizes the empirical prior that "the highest bit of the exponent is always 0" to construct a lightweight coding structure. Specifically, the system can read the binary representation of the neural network's floating-point parameters and decompose the binary representation of the neural network's floating-point parameters into a sign bit, an exponent field, and a mantissa field according to the IEEE 754 standard.

[0030] Step 102: Adaptively select the corresponding exponential redundancy coding strategy based on the parameter size of the neural network model.

[0031] In the embodiments of this application, this design of adjusting the protection structure based on model statistical properties elevates the Zero-Memory-Overhead Redundancy Protection (ZMORP) method of this application from a single coding technique to a class of floating-point semantic fault-tolerant frameworks. Specifically, it may include: Determine the number of parameters for the neural network model; If the number of parameters in the neural network model is less than a preset threshold, the neural network model is determined to be a small-to-medium scale model, and the exponential redundancy coding strategy is determined to be a lightweight coding strategy. If the number of parameters in a neural network model exceeds a preset threshold (for example, in a large model containing billions of parameters, although the proportion of parameters with an exponent MSB of 1 is extremely low, their absolute number reaches tens of thousands; if forced to be zero, it will introduce serious numerical distortion and accumulate layer by layer in ultra-deep networks), then the neural network model is determined to be a large model, and the exponential redundancy coding strategy is determined to be the coding strategy for large models.

[0032] Step 103: According to the exponential redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, the redundant copy of the exponential information to be protected and the corresponding verification information are re-embedded and mapped to the low-order region of the tail number segment to generate the reconstructed protection parameters. The exponential information to be protected is the low-order exponential segment or the complete exponential bit string of the exponential field, and the total number of bits of the protection parameters is the same as that of the neural network floating-point parameters.

[0033] In the embodiments of this application, based on a lightweight coding strategy, without increasing the total number of bits of the neural network floating-point parameters, redundant copies of the exponent information to be protected and the corresponding check information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters. Specifically, this may include: In response to the lightweight coding strategy, the exponent field is divided into the most significant bit and the least significant bit exponent segment, and the least significant bit exponent segment is used as the exponent information to be protected. Calculate the first even parity value of the low-order exponent segment, and replace the most significant bit in the exponent field with the first even parity value; From the least significant bit of the last digit segment, a first redundancy area and a first parity area of ​​the same length as the low-order exponent segment are drawn out. The low-order exponent segment is copied to the first redundancy area as a redundant copy. The second even parity value of the redundant copy is calculated and stored in the first parity area. The protection parameters are reorganized based on the truncated tail number segment, the first check area, and the first redundancy area.

[0034] In this embodiment, the exponent field is divided into two parts: the most significant bit (MSB) and the remaining low-order exponent segments. Since the MSB is theoretically fixed to 0 in this type of model, it is not reserved as a purely numerical bit, but rather reused as the even parity bit of the original low-order exponent segment. Simultaneously, several least significant bits are allocated from the mantissa segment to store a "redundant copy of the low-order exponent segment." The even parity bit of the redundant copy is then embedded in the preceding or adjacent reserved bit. After the above mapping, a raw floating-point parameter is reconstructed into a protection parameter containing the following logical fields: sign bit, parity bit of the original low-order exponent segment, original low-order exponent segment, truncated mantissa segment, redundant exponent copy parity bit, and redundant exponent copy. This structure, without increasing the total number of bits, simultaneously preserves the original exponent, the exponent backup, and two sets of consistency detection information, thus providing dual-path evidence for subsequent recovery.

[0035] Further, in this small-scale implementation, the encoding process of this application can be described as follows: First, read the binary representation of the original floating-point parameters; then extract the sign bit, the exponent string (i.e., the exponent field), and the mantissa string (i.e., the mantissa segment); next, decompose the exponent string into the most significant bit (MSB) and the remaining bits; calculate the even parity value for the remaining bits; reserve a redundant area of ​​equal length to the low exponent bits and a parity area from the lowest several bits of the mantissa, copy the low exponent bits to the redundant area to obtain the redundant part, and calculate the even parity value for the redundant part; then cover the original MSB with the even parity value of the original exponent bits, and rearrange the truncated high-order mantissa part, the even parity value of the redundant part, and the redundant bits together to form a new mantissa segment; finally, reassemble the sign bit, the replaced exponent field, and the reassembled mantissa segment into the protection parameters. The key here is not simple copying, but controlled rearrangement based on floating-point semantics: exponent protection takes priority over mantissa precision, because in radiation error scenarios, exponent flipping usually has a more significant impact on model accuracy than low-order mantissa errors, while the quantization perturbation caused by mantissa truncation is still within the network's tolerance range.

[0036] In the embodiments of this application, according to the large model encoding strategy, without increasing the total number of bits of the neural network floating-point parameters, redundant copies of the exponent information to be protected and the corresponding verification information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters, which may specifically include: In response to the large model coding strategy, the complete exponent bit string of the exponent field is determined as the exponent information to be protected; From the least significant bit of the last digit segment, a second redundancy area and a second check area of ​​the same length as the complete exponent string are drawn out. The complete exponent bit string is copied to the second redundancy area as a redundant copy. The third even parity value of the complete exponent bit string and the fourth even parity value of the redundant copy are calculated respectively, and the third even parity value and the fourth even parity value are stored in the second parity area. The protection parameters are reorganized based on the truncated tail number segment, the second check area, and the second redundancy area.

[0037] In this embodiment, this application does not mechanically generalize the conclusion that "the highest bit of the exponent is always 0 in small models" to all networks. For large models containing billions of parameters, although the proportion of parameters with an exponent MSB of 1 is extremely low, their absolute number can still reach close to tens of thousands. If the exponent MSB of these parameters is still forcibly set to 0, it will introduce serious numerical distortion, which will accumulate layer by layer in the forward propagation of ultra-deep networks, eventually leading to a sharp decline in inference performance. Therefore, this application further proposes a ZMORP variant for MLLMs, the core idea of ​​which is: no longer reuse the highest bit of the exponent, no longer modify the original exponent field, but perform a complete redundant copy of the entire exponent field, and arrange both even parity checks in the reserved bits of the last digit field. In this way, the original exponent information can be completely preserved while maintaining zero additional storage, avoiding structural damage to a few large-value parameters in large models.

[0038] In this MLLM implementation, the encoding process is as follows: First, extract the complete exponent string and mantissa segment for each floating-point parameter; then, delineate a redundant area of ​​equal length to the exponent and two parity segments from the lower bits of the mantissa; next, copy the exponent to the redundant area as is, and calculate even parity values ​​for the exponent and redundant bits respectively; then, do not rewrite any bit of the original exponent field, but only write two even parity values ​​and the redundant exponent part to the lower bits of the mantissa segment; finally, reassemble the floating-point parameter with the complete original exponent and the rewritten mantissa segment into a protection parameter. Compared to the small model implementation, this scheme sacrifices more precision in the lower bits of the mantissa, but gains complete protection for the entire exponent field, thus making it suitable for large-scale deep multimodal remote sensing models.

[0039] Step 104: Before model inference or during periodic checks, extract the information of the index to be protected, the redundant copies and their respective verification information from the protection parameters. Use a dual-path even verification mechanism to determine the credibility of the information of the index to be protected and the redundant copies, and execute the corresponding index recovery logic according to the determination result to obtain the target index domain.

[0040] In this embodiment of the application, a dual-path even parity check mechanism is used to determine the credibility of the index information to be protected and the redundant copy, and the corresponding index recovery logic is executed according to the determination result to obtain the target index domain. Specifically, it may include: In response to the lightweight coding strategy, even parity is recalculated for the parsed low-order exponent fragment and redundant copy respectively; If only the low-order exponent segment passes the verification, then the low-order exponent segment will be used as the recovered low-order exponent segment. If only the redundant copy passes the verification, the redundant copy will be used as the low-order exponent fragment after recovery. If both the low-order exponent segment and the redundant copy pass the verification, then either the low-order exponent segment or the redundant copy can be selected. If both the low-order exponent segment and the redundant copy fail to be verified, then perform a bit-by-bit comparison on the low-order exponent segment and the redundant copy, retain the bits at the same position, force the bits at different positions to be set to zero, and use the processing result as the recovered low-order exponent segment. Write the recovered low-order exponent fragment back to the exponent field, and explicitly force the most significant bit of the exponent field to be restored to a fixed value, and combine them to generate the target exponent field.

[0041] In this embodiment, during the decoding and recovery phases, this application does not directly trust any segment of exponent information. Instead, it uses dual-path even parity checking to determine the source of trust. During recovery, the original low-order exponent segment, redundant copy, and their corresponding parity bits are first parsed from the protection parameters. Then, the two even parity checks are recalculated. If the original low-order exponent segment verifies correctly while the redundant copy fails, the original segment is considered more trustworthy, and the system uses the original segment to reconstruct the exponent. If the original low-order exponent segment verifies incorrectly while the redundant copy verifies correctly, the redundant copy is not damaged, and the system uses the redundant copy to recover the original exponent. If both paths verify correctly, either can be directly selected as the recovery basis. If both paths fail, the application enters the bit-by-bit consensus recovery process specifically designed in this application. Bit-by-bit consensus refers to comparing the original segment and the redundant segment bit by bit: if corresponding bits are the same, the bit is retained; if corresponding bits are different, the bit is forcibly set to 0. This rule is used to achieve conservative recovery when both sides are questionable but some consistent information still exists. It utilizes the prior distribution of the remote sensing model index MSB and the overall characteristic of small parameter amplitudes to ensure that the restoration results fall within a safe numerical range as much as possible.

[0042] In this embodiment of the application, a dual-path even parity check mechanism is used to determine the credibility of the index information to be protected and the redundant copy, and the corresponding index recovery logic is executed according to the determination result to obtain the target index domain. Specifically, it may include: In response to the large model encoding strategy, even parity is recalculated for the parsed complete exponent bit string and the redundant copy respectively; If only the complete exponent string passes the verification, then the complete exponent string will be used as the recovered complete exponent string. If only the redundant copy passes the verification, the redundant copy will be used as the recovered complete exponent bit string. If both the complete exponent string and the redundant copy pass the verification, then either the complete exponent string or the redundant copy can be selected as the target exponent field. If both the complete exponent string and the redundant copy fail to be verified, then perform a bit-by-bit comparison on the complete exponent string and the redundant copy, retain the bits at the same positions, and force the bits at different positions to be zero, and use the processing result as the recovered complete exponent string. Write the restored complete exponent string back to the exponent field, without performing the operation of zeroing the most significant bit, and directly generate the target exponent field.

[0043] In this embodiment, the recovery process in the MLLM scenario has also been adjusted. The system first reads the original exponent, the redundant exponent copy, and two sets of parity bits; then, it performs even parity recalculation on the exponent bits and the redundant bits respectively; if one path passes and the other fails, the path that passes the parity check is selected as the recovered exponent; if both fail simultaneously, a bit-by-bit consensus mechanism is executed, comparing each bit, retaining it if they match, and conservatively setting it to zero if they do not match; if both pass simultaneously, the original exponent can be used directly, or it can be used after verification with the redundant copy. Unlike the small model, the operation of setting the highest exponent to zero is not performed after recovery here, but the recovered complete exponent is written back as is to retain a few but critical large-value parameters in the large model. That is to say, the "MSB zeroing" in the small model path is a post-processing rule based on statistical priors, while the large model path cancels this rule and changes to "full exponent recovery as is". This design of adaptively selecting the protection structure according to the model size and parameter distribution characteristics makes ZMORP improve from a single coding technique to a floating-point semantic fault-tolerant framework that can be adjusted according to the statistical properties of the model.

[0044] Step 105: Clear or truncate the low-order region of the mantissa segment to reconstruct the target mantissa segment. Combine the target exponent field, sign bit and target mantissa segment to obtain the recovered neural network floating-point parameters.

[0045] In this embodiment, after recovering the low-order exponent segment, the recovery result is written back to the exponent domain. Simultaneously, the reserved bits used for redundant storage in the mantissa are cleared or treated as invalid protection zones, no longer participating in numerical interpretation. Subsequently, for small models, the highest bit of the exponent is explicitly restored to 0. This step is significant: it not only rewrites the protected structure back into normal floating-point semantics but also explicitly encodes the previously obtained parameter distribution prior into a reconstruction rule during the recovery phase. In other words, this application does not passively correct errors but superimposes numerical constraints driven by model statistical knowledge at the end of the error correction process, thereby preventing abnormal amplification of the absolute value of parameters due to radiation flipping in the high-order exponent. This two-level recovery method of "verification recovery + distribution prior constraint" is a key feature that distinguishes this application from pure ECC or pure replication verification schemes.

[0046] From a system implementation perspective, this application can complete protection encoding both during offline model deployment and before loading model parameters into on-chip storage. For each layer's weight tensor, the above encoding process can be performed element-wise when the weights are written to memory, replacing ordinary floating-point weights with protected floating-point representations. During runtime, a full parameter inspection and recovery can be performed before inference, or partial recovery can be performed at key nodes such as layer loading, block scheduling, and tensor retrieval. If the target platform has a dedicated storage controller, this application can be integrated into an online logic of "parameter readout and verification recovery"; if the platform is primarily software-based, a parameter repair routine can be added before the inference graph is executed. Since ZMORP only involves bit extraction, even parity, bit-by-bit comparison, zeroing, and field reassembly, its implementation complexity is relatively low, making it suitable for both hardware state machines and lightweight software kernels.

[0047] The technical advantages of this application are mainly reflected in the following aspects. First, in terms of storage cost, this application does not introduce additional parameter copies and independent error correction blocks, and the protection information is completely embedded within the original parameters. Therefore, the overall model storage scale does not increase, making it particularly suitable for spaceborne platforms with extremely limited storage resources. Second, in terms of error suppression capability, this application prioritizes the protection of the exponential domain, which has the greatest impact on the numerical scale, effectively avoiding the order-of-magnitude abrupt change in parameters caused by radiative flips. Compared with a simple parameter zeroing strategy, this approach better maintains the model's representational capability. Third, in terms of numerical fidelity, this application only uses the low-order mantissa as a redundancy carrier. Neural networks have a strong tolerance for low-order mantissa perturbations, thus achieving significant error tolerance benefits at a relatively low precision cost. Finally, in terms of applicability, this application is not designed for a single model, but rather, through a process of "statistical prior identification—structured redundancy embedding—dual-path verification recovery—recovery rule selection according to model size," it is applicable to different deployment objects, from lightweight remote sensing networks to multimodal large models.

[0048] In addition to the main scheme described in detail in this application, there is an alternative fault-tolerant scheme that is structure-independent but format-aware: Hamming code is introduced into the on-orbit radiation protection system of deep neural network model parameters, and a "parameter-by-parameter granular coding + in-situ storage" strategy is adopted. This method fully utilizes the characteristic that neural networks are insensitive to perturbations of low-weight bits, and achieves single-bit error detection and correction of critical high-weight information bits without introducing any additional storage overhead, thereby achieving high-performance fault tolerance in the extremely resource-constrained spaceborne environment.

[0049] Specifically, this scheme is based on the error correction theory of Hamming codes, and its core constraints are:

[0050] Where n is the information code length and k is the introduced check bits. This formula shows that k check bits can represent... The accompanying states must be sufficient to cover all error locations and error-free states. This application utilizes the characteristic that the low-order bits of the model parameters have little impact on model performance. For float32 format parameters, n is set to 26 and k to 5, discarding the lower 5 bits and using 5 check bits to encode and protect the higher 26 bits of information; for float16 format parameters, n is set to 11 and k to 4, discarding the lower 4 bits and using 4 check bits to protect the higher 11 bits of information. In the decoding stage, the system first uses the accompanying computation mechanism of Hamming code to detect and correct possible single-bit flips, recovering the original higher n bits of information; then, the discarded lower k bits are restored to the standard floating-point format with zero padding for subsequent inference. Since the discarded lower bits themselves contribute little to the model accuracy, and the higher bits remain intact after error correction, the overall model performance is almost lossless.

[0051] In summary, while this alternative can serve as a lightweight backup strategy in resource-constrained spaceborne scenarios, its error correction capabilities are limited, and its overall reliability (especially under high bit error rate environments) is significantly inferior to the fault-tolerance mechanism in the main solution of this application. Therefore, it is positioned as a suboptimal but feasible supplementary technical path, suitable for specific aerospace applications that are extremely sensitive to storage and power consumption and can accept certain performance trade-offs.

[0052] Based on the above Figure 1 The specific implementation of the method shown in this embodiment provides a parameter fault-tolerant protection device with zero additional memory overhead, such as... Figure 3 As shown, the device includes: an acquisition module 31, a selection module 32, a generation module 33, an execution module 34, and a combination module 35; The acquisition module 31 is used to acquire the neural network floating-point parameters of the neural network model to be deployed in the radiation environment, and to decompose the neural network floating-point parameters into the sign bit, exponent field and mantissa field according to the IEEE 754 standard. Selection module 32 is used to adaptively select the corresponding exponential redundancy coding strategy according to the parameter scale of the neural network model. The generation module 33 is used to re-embed and map the redundant copy of the exponent information to be protected and the corresponding verification information to the low-order region of the tail number segment according to the exponent redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, to generate reconstructed protection parameters. The exponent information to be protected is the low-order exponent fragment or the complete exponent bit string of the exponent field, and the total number of bits of the protection parameters is the same as that of the neural network floating-point parameters. The execution module 34 is used to parse the information to be protected, the redundant copy and their respective verification information from the protection parameters before model inference or during periodic checks, determine the credibility of the information to be protected and the redundant copy through a dual-path even parity check mechanism, and execute the corresponding index recovery logic according to the determination result to obtain the target index domain. The combination module 35 is used to clear or truncate the low-order region of the tail number segment to reconstruct the target tail number segment, and combine the target exponent field, the sign bit and the target tail number segment to obtain the recovered neural network floating-point parameters.

[0053] In specific application scenarios, the selection module 32 can be used to determine the number of parameters of the neural network model; if the number of parameters of the neural network model is less than a preset threshold, the neural network model is determined to be a small-to-medium scale model, and the exponential redundancy coding strategy is determined to be a lightweight coding strategy; if the number of parameters of the neural network model is greater than the preset threshold, the neural network model is determined to be a large model, and the exponential redundancy coding strategy is determined to be a large model coding strategy.

[0054] In a specific application scenario, the generation module 33 can be used to respond to the lightweight coding strategy by dividing the exponent field into the most significant bit and the least significant bit exponent segment, and using the least significant bit exponent segment as the exponent information to be protected; calculating the first even parity value of the least significant bit exponent segment, and replacing the most significant bit in the exponent field with the first even parity value; drawing out a first redundancy area and a first parity area of ​​the same length as the least significant bit exponent segment from the least significant bit of the tail number segment, copying the least significant bit exponent segment to the first redundancy area as the redundant copy, calculating the second even parity value of the redundant copy and storing it in the first parity area; and recombining the truncated tail number segment, the first parity area and the first redundancy area into protection parameters.

[0055] In a specific application scenario, the execution module 34 can be used to respond to the lightweight encoding strategy by recalculating the parsed low-order exponent fragment and the redundant copy. If only the low-order exponent fragment passes the parsing, it is used as the recovered low-order exponent fragment. If only the redundant copy passes the parsing, it is used as the recovered low-order exponent fragment. If both the low-order exponent fragment and the redundant copy pass the parsing, either the low-order exponent fragment or the redundant copy can be selected. If both the low-order exponent fragment and the redundant copy fail the parsing, a bit-by-bit comparison is performed on the low-order exponent fragment and the redundant copy, retaining the bits at the same positions and forcibly setting the bits at different positions to zero. The processing result is used as the recovered low-order exponent fragment. The recovered low-order exponent fragment is written back to the exponent field, and the most significant bit of the exponent field is explicitly forced to be restored to a fixed value, and the target exponent field is generated by combining them.

[0056] In a specific application scenario, the generation module 33 can be used to respond to the large model encoding strategy by determining the complete exponent bit string of the exponent field as the exponent information to be protected; dividing a second redundancy area and a second parity area of ​​the same length as the complete exponent bit string from the least significant bit of the tail number segment; copying the complete exponent bit string to the second redundancy area as a redundant copy; calculating the third even parity value of the complete exponent bit string and the fourth even parity value of the redundant copy respectively; and storing the third even parity value and the fourth even parity value in the second parity area; and recombining the truncated tail number segment, the second parity area, and the second redundancy area into protection parameters.

[0057] In a specific application scenario, the execution module 34 can be used to respond to the large model encoding strategy by recalculating the parsed complete exponent bit string and the redundant copy. If only the complete exponent bit string passes the verification, it is used as the recovered complete exponent bit string. If only the redundant copy passes the verification, it is used as the recovered complete exponent bit string. If both the complete exponent bit string and the redundant copy pass the verification, either the complete exponent bit string or the redundant copy is selected as the target exponent field. If both the complete exponent bit string and the redundant copy fail the verification, a bit-by-bit comparison is performed on the complete exponent bit string and the redundant copy, retaining bits at the same positions and forcibly setting bits at different positions to zero, and the processing result is used as the recovered complete exponent bit string. The recovered complete exponent bit string is written back to the exponent field without performing the most significant bit zeroing operation, and the target exponent field is generated directly.

[0058] It should be noted that other corresponding descriptions of the functional units involved in the parameter fault-tolerant protection device with zero additional memory overhead provided in this embodiment can be found in [reference]. Figure 1 The corresponding description in [the document] will not be repeated here.

[0059] Based on the above, Figure 1 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 The method shown.

[0060] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0061] Based on the above, Figure 1 The method shown, and Figure 3 To achieve the above objectives, the present application also provides an electronic device, comprising a storage medium and a processor; the storage medium for storing a computer program; and the processor for executing the computer program to implement the above-described virtual device embodiments. Figure 1 The method shown.

[0062] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0063] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0064] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of parameter-tolerant fault-protected programs with zero additional memory overhead, as well as other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, and to communicate with other hardware and software within the parameter-tolerant fault-protected physical device with zero additional memory overhead.

[0065] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented using hardware. By applying the solution of this embodiment, compared with the prior art, this application can obtain the neural network floating-point parameters of the neural network model to be deployed in the radiation environment, and according to IEEE... The 754 standard decomposes the floating-point parameters of a neural network into a sign bit, an exponent field, and a mantissa segment. Based on the parameter size of the neural network model, it adaptively selects the corresponding exponent redundancy coding strategy. According to the exponent redundancy coding strategy, without increasing the total number of bits in the neural network floating-point parameters, redundant copies of the exponent information to be protected and the corresponding check information are re-embedded and mapped to the low-order region of the mantissa segment to generate reconstructed protection parameters. The exponent information to be protected is a low-order exponent fragment or a complete exponent string, and the total number of bits in the protection parameters is the same as that of the neural network floating-point parameters. Before model inference or during periodic checks, the exponent information to be protected, the redundant copy, and their respective check information are parsed from the protection parameters. A dual-path even parity check mechanism is used to determine the credibility of the exponent information to be protected and the redundant copy, and the corresponding exponent recovery logic is executed based on the determination result to obtain the target exponent field. The low-order region of the mantissa segment is cleared or truncated to reconstruct the target mantissa segment. The target exponent field, sign bit, and target mantissa segment are combined to obtain the recovered neural network floating-point parameters.

[0066] By employing the aforementioned technical solution, this application completely breaks through the passive protection limits of traditional quantization pruning methods and model-based redundancy methods by "re-embedding redundant copies of the exponent information to be protected and corresponding verification information and mapping them to the low-order region of the tail number segment." This application innovatively recognizes that in IEEE 754 floating-point numbers, the exponent field is the "absolute critical bit" that has the greatest impact on the numerical scale. Without altering the model structure, it proactively extracts the most vulnerable exponent information for dual backup and verification protection; coupled with a "dual-path even parity check mechanism," it can accurately reconstruct the exponent field. This means that even if the most critical bit in the model suffers a devastating radiation attack, this application can accurately restore it through internally embedded redundant copies, raising the protection limit from the dimension of underlying floating-point semantics and meeting the high reliability requirements of the extreme environment of space.

[0067] This application achieves complete "plug-and-play" protection through a parameter fault-tolerant protection method with zero additional memory overhead. From disassembling floating-point parameters to generating reconstructed protection parameters, and then to obtaining the restored floating-point parameters, the entire process is a purely low-level binary bit-level encoding and decoding process. It does not require touching the forward / backward propagation logic of the neural network, modifying the network structure, or consuming any computing power for retraining or fine-tuning. Before launch or while in orbit, the satellite only needs to execute the bit conversion routine of this application once on the pre-trained weight data to obtain radiation resistance, perfectly overcoming the fatal flaw of unacceptable time and computing power costs in fault-tolerant training in aerospace engineering.

[0068] Instead of simply discarding erroneous data during the parsing phase, this application utilizes pre-embedded verification information to determine credibility through a "dual-path even-check mechanism." This mechanism can restore the original numerical appearance of the parameters to the greatest extent possible, avoiding neuron death or sudden loss of network features caused by "setting parameters to zero," thus fundamentally protecting the inference accuracy of the model.

[0069] This application "squeezes" redundant information of the exponent into and maps it to the "lower bit region of the mantissa field" without increasing the total number of bits of the floating-point parameters of the neural network. This takes advantage of the inherent characteristic of the neural network model that is not sensitive to the precision of the lower bits of the mantissa (i.e., the model itself has a certain noise tolerance), and trades the slight loss of precision in the lower bits of the mantissa for the extremely high importance protection of the exponent field, achieving "zero increase" in storage overhead.

[0070] By adaptively selecting the corresponding exponential redundancy coding strategy based on the parameter scale, a "one-size-fits-all" approach is avoided, achieving an optimal balance between protection strength and accuracy loss.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0072] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. 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 this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A parameter fault-tolerant protection method with zero additional memory overhead, characterized in that, The method includes: Obtain the floating-point parameters of the neural network model to be deployed in the radiation environment, and decompose the floating-point parameters of the neural network into the sign bit, exponent field and mantissa field according to the IEEE 754 standard; Based on the parameter size of the neural network model, the corresponding exponential redundancy coding strategy is adaptively selected; According to the exponential redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, redundant copies of the exponential information to be protected and the corresponding verification information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters. The exponential information to be protected is the low-order exponential fragment or the complete exponential bit string of the exponential field, and the total number of bits of the protection parameters is the same as that of the neural network floating-point parameters. Before model inference or during periodic checks, the information to be protected, the redundant copy, and their respective verification information are parsed from the protection parameters. The credibility of the information to be protected and the redundant copy is determined by a dual-path even parity check mechanism. Based on the determination result, the corresponding index recovery logic is executed to obtain the target index domain. The low-order region of the tail number segment is cleared or truncated to reconstruct the target tail number segment. The target exponent field, the sign bit, and the target tail number segment are combined to obtain the recovered neural network floating-point parameters.

2. The method according to claim 1, characterized in that, The step of adaptively selecting the corresponding exponential redundancy coding strategy based on the parameter size of the neural network model specifically includes: Determine the number of parameters in the neural network model; If the number of parameters in the neural network model is less than a preset threshold, then the neural network model is determined to be a small-to-medium scale model, and the exponential redundancy coding strategy is determined to be a lightweight coding strategy.

3. The method according to claim 2, characterized in that, According to the exponential redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, the redundant copy of the exponential information to be protected and the corresponding check information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters, specifically including: In response to the lightweight coding strategy, the exponent field is divided into the most significant bit and the least significant bit exponent segment, and the least significant bit exponent segment is used as the exponent information to be protected; Calculate the first even parity value of the low-order exponent segment, and replace the most significant bit in the exponent field with the first even parity value; From the least significant bit of the last digit segment, a first redundancy area and a first parity area of ​​the same length as the low-order exponent segment are drawn out. The low-order exponent segment is copied to the first redundancy area as the redundancy copy. The second even parity value of the redundancy copy is calculated and stored in the first parity area. The protection parameters are recombined based on the truncated tail number segment, the first check area, and the first redundancy area.

4. The method according to claim 3, characterized in that, The process of determining the credibility of the index information to be protected and the redundant copy through a dual-path even parity check mechanism, and executing the corresponding index recovery logic based on the determination result to obtain the target index domain, specifically includes: In response to the lightweight encoding strategy, even parity is recalculated for the parsed low-order exponent segment and the redundant copy, respectively. If only the low-order exponent segment passes the verification, then the low-order exponent segment is used as the recovered low-order exponent segment. If only the redundant copy passes the verification, then the redundant copy will be used as the recovered low-order exponent fragment. If both the low-order exponent segment and the redundant copy pass the verification, then either the low-order exponent segment or the redundant copy can be selected. If both the low-order exponent segment and the redundant copy fail to be verified, then a bit-by-bit comparison is performed on the low-order exponent segment and the redundant copy, retaining the bits at the same position, forcibly setting the bits at different positions to zero, and using the processing result as the recovered low-order exponent segment. Write the recovered low-order exponent fragment back to the exponent field, and explicitly force the most significant bit of the exponent field to be restored to a fixed value, and combine them to generate the target exponent field.

5. The method according to claim 1, characterized in that, The step of adaptively selecting the corresponding exponential redundancy coding strategy based on the parameter size of the neural network model specifically includes: Determine the number of parameters in the neural network model; If the number of parameters in the neural network model is greater than a preset threshold, then the neural network model is determined to be a large model, and the exponential redundancy coding strategy is determined to be a large model coding strategy.

6. The method according to claim 5, characterized in that, According to the exponential redundancy coding strategy, without increasing the total number of bits of the neural network floating-point parameters, the redundant copy of the exponential information to be protected and the corresponding check information are re-embedded and mapped to the low-order region of the tail number segment to generate reconstructed protection parameters, specifically including: In response to the large model encoding strategy, the complete exponent bit string of the exponent field is determined as the exponent information to be protected; A second redundancy region and a second check region of the same length as the complete exponent string are drawn from the least significant bit of the tail digit segment. The complete exponent bit string is copied to the second redundancy area as a redundant copy. The third even parity value of the complete exponent bit string and the fourth even parity value of the redundant copy are calculated respectively, and the third even parity value and the fourth even parity value are stored in the second parity area. The protection parameters are recombined based on the truncated tail number segment, the second check area, and the second redundancy area.

7. The method according to claim 6, characterized in that, The process of determining the credibility of the index information to be protected and the redundant copy through a dual-path even parity check mechanism, and executing the corresponding index recovery logic based on the determination result to obtain the target index domain, specifically includes: In response to the large model encoding strategy, even parity is recalculated for the parsed complete exponent bit string and the redundant copy, respectively. If only the complete exponent bit string passes the verification, then the complete exponent bit string is used as the recovered complete exponent bit string. If only the redundant copy passes the verification, then the redundant copy is used as the recovered complete exponent bit string; If both the complete exponent string and the redundant copy pass the verification, then either the complete exponent string or the redundant copy can be selected as the target exponent field. If both the complete exponent string and the redundant copy fail to be verified, then a bit-by-bit comparison is performed on the complete exponent string and the redundant copy, retaining the bits at the same position and forcibly setting the bits at different positions to zero, and the processing result is used as the recovered complete exponent string. Write the restored complete exponent string back to the exponent field without performing the operation of zeroing the most significant bit, and directly generate the target exponent field.

8. A parameter fault-tolerant protection device with zero additional memory overhead, characterized in that, include: The acquisition module is used to acquire the neural network floating-point parameters of the neural network model to be deployed in the radiation environment, and to decompose the neural network floating-point parameters into the sign bit, exponent field and mantissa field according to the IEEE 754 standard. The selection module is used to adaptively select the corresponding exponential redundancy coding strategy based on the parameter size of the neural network model. The generation module is used to re-embed and map the redundant copy of the exponent information to be protected and the corresponding verification information to the low-order region of the tail number segment according to the exponent redundancy coding strategy, without increasing the total number of bits of the floating-point parameters of the neural network, to generate reconstructed protection parameters. The exponent information to be protected is the low-order exponent fragment or the complete exponent bit string of the exponent field, and the total number of bits of the protection parameters is the same as that of the floating-point parameters of the neural network. The execution module is used to parse the information to be protected, the redundant copy and their respective verification information from the protection parameters before model inference or during periodic checks, determine the credibility of the information to be protected and the redundant copy through a dual-path even parity check mechanism, and execute the corresponding index recovery logic according to the determination result to obtain the target index domain. The combination module is used to clear or truncate the low-order region of the tail number segment to reconstruct the target tail number segment, and combine the target exponent field, the sign bit and the target tail number segment to obtain the recovered neural network floating-point parameters.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the parameter fault-tolerant protection method with zero additional memory overhead as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the parameter fault-tolerant protection method with zero additional memory overhead as described in any one of claims 1 to 7.

Images