As the feature sizes of integrated circuits are reduced to the nanometer scale, the total soft error rate (SER) in memory and the proportion of multiple bit upsets (MBUs) are significantly increasing. In order to ensure the information reliability, many error correction codes with strong error correction ability were proposed, such as Reed–Somolon (RS) code and Bose–Chaudhuri–Hocquenghem (BCH) code. However, these error correction codes have limited error correction capability, high algorithm complexity and large data redundancy. In this paper, a novel fault tolerance method for locating and correcting multiple bit errors in memory is proposed based on data similarity. The proposed method uses the inner product as the metric to analyze the similarity of the pre-protected data from the vertical and horizontal dimensions, respectively, and to construct the model of error location and correction. This method performs encoding and decoding in units of blocks and detecting and correcting in units of words, so it can correct any number of bits in a corrupted word with low redundancy overhead. Finally, irradiation tests were conducted on a commercial SRAM, and the feasibility of the proposed method is verified by using heavy ion ${}^{86}$Kr${}^{26+}$ as irradiation source.

随着集成电路的特征尺寸缩小到纳米级，存储器中的总软错误率（SER）和多位翻转（MBU）的比例显着增加。为了保证信息的可靠性，许多纠错能力很强的纠错码被提出，如里德索莫隆（RS）码和博斯乔杜里霍昆亨（BCH）码。但是，这些纠错码纠错能力有限，算法复杂度高，数据冗余度大。本文提出了一种基于数据相似性定位和纠正内存中多位错误的容错方法。所提出的方法使用内积作为度量，分别从垂直和水平维度分析预保护数据的相似性，并构建错误定位和纠正模型。该方法以块为单位进行编码和解码，以字为单位进行检测和纠错，因此可以以低的冗余开销纠正被破坏的字中的任意位数。最后，在商用SRAM上进行了辐照试验，利用重离子验证了所提方法的可行性。${}^{\mathrm{8个}\mathrm{6个}}$氪${}^{\mathrm{2个}\mathrm{6个}+}$作为辐照源。