首页 > AI前沿 > Bio-PM-Guided Synthetic IMU Generation

Bio-PM-Guided Synthetic IMU Generation

arXiv机器学习 2026-11-20 02:56 10 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,按照要求,每个自然段落用p,这里原文是一个连续的学术段落对吧?不对,先看内容,首先是整个一段正文,没有分小节,所以直接用p包裹?不对,等下看内容,有没有需要划分的?哦原文是一整段?不对再仔细看,原文的正文是一个连贯的段落,没有分节,所以直接把整个段落用p包裹?不对,等下检查规则,要保留所有信息,不能增删改。哦对,然后看有没有需要加粗的?比如重要的术语?比如IMU,HAR,还有Bio-PM,还有TSTR,还有结果里的数值?不对,不对,规则是重要句子可用strong,不要随便加?不对,等下要准确,不能增改内容。首先,整个正文是一个自然段落对吧?那直接用

包裹整个内容?不对等下再看原文内容,有没有什么结构?哦原文是:首先讲收集IMU数据的问题,然后讲方法,然后讲实验和结果。不过用户说长文用h2/h3划分,但是这里原文有没有小节标题?没有,原文是连续的一段,所以不需要加h2/h3,因为没有小节。然后有没有列表?哦方法里的两个解码器,原文是(1)和(2),可以转成列表吗?不对,规则是列表内容用ul li,哦对,原文里的(1) an autoregressive...和(2) a diffusion decoder...是两个并列的方法,可以转成列表对吧?不对,等下要保留原文信息,不能改。那怎么处理?哦原文里的内容:“and train class-conditional decoders in this space: (1) an autoregressive transformer decoder that predicts tokens sequentially, and (2) a diffusion decoder that learns the token distribution through a denoising objective.” 这里冒号后面是两个并列的项,所以可以把这部分做成列表?对,这样符合语义。然后重要的术语比如惯性测量单元(IMU),人类活动识别(HAR),还有Bio-PM,还有TSTR,还有关键的结果数值?比如最后的结果里的数值可以加粗?不对,不要乱加,重要的句子或者术语,比如核心的结论?等下,先理清楚: 首先,第一句是背景:

Collecting and labeling inertial measurement unit (IMU) data for human activity recognition (HAR) is expensive, and models trained on one group of subjects often fail to generalize to unseen individuals. Even with balanced data, classifiers struggle to distinguish dynamically similar activity classes and benefit from targeted augmentation.

?不对,原文是一整段啊?哦不对,用户给的正文是一整段,我是不是可以拆?不对,规则说每个自然段落用p包裹,原文的自然段落就是一个,对吧?不对,等下看原文的标点,有没有自然的分段?没有,原文是连续的一个段落。那然后里面的并列的两个解码器,是不是可以用列表?对,因为那是两个并列的项。然后重要的内容,比如核心的发现,比如“Our diffusion decoder achieves a Train-on-Synthetic-Test-on-Real (TSTR) macro-F1 of 0.68, demonstrating that generated tokens are realistic enough to train a classifier that transfers to real data. ”还有后面的提升结果,是不是可以用strong?不对,或者重要的术语加粗?比如IMU,HAR,Bio-PM,TSTR这些?不对,不要过度,但是要语义化。等下,先准确: 首先,整个原文是一个自然段,但是里面有一个冒号引导的两个并列的解码器,所以把那两个做成ul li,对吧?那结构是:

Collecting and labeling inertial measurement unit (IMU) data for human activity recognition (HAR) is expensive, and models trained on one group of subjects often fail to generalize to unseen individuals. Even with balanced data, classifiers struggle to distinguish dynamically similar activity classes and benefit from targeted augmentation. We explore whether synthetic IMU data generated in the representation space of a pretrained biosignal encoder, Bio-PM \cite{b1}, can serve as a practical augmentation source. Rather than synthesizing raw signals, we leverage Bio-PM's frozen encoder to convert 10-second IMU windows into token embedding sequences and train class-conditional decoders in this space:

  • an autoregressive transformer decoder that predicts tokens sequentially,
  • a diffusion decoder that learns the token distribution through a denoising objective.
We analyze generation realism through distributional metrics and embedding-space visualizations, evaluate downstream utility through leave-one-subject-out HAR classification, and present a PatchDecoder for approximate raw waveform reconstruction. Our diffusion decoder achieves a Train-on-Synthetic-Test-on-Real (TSTR) macro-F1 of 0.68, demonstrating that generated tokens are realistic enough to train a classifier that transfers to real data. Combining synthetic-token augmentation with a 320-dimensional gravity-aware feature representation raises LOSO macro-F1 from 0.729 to 0.809 over a real-only baseline across 51 subjects on the WISDM benchmark, with statistically significant gains concentrated in the most challenging activity classes. These results demonstrate that high-fidelity generation in a pretrained encoder's representation space is achievable and practically useful for cross-subject HAR augmentation.

等下,这样对吗?检查有没有增删改内容?原文里的(1)和(2)我去掉编号了?不对,哦原文是(1)和(2),那要不要保留?哦不对,原文的(1)(2)是序号,如果转成列表的话,ul的li本身有序?不对,原文是(1)(2),是有序的?那用ol?哦对哦,我刚才搞错了,是两个有序的项,所以应该用
  1. 对吧?因为原文是(1)(2),是有序列表。哦对,刚才错了,应该是有序列表。那调整一下:

    Collecting and labeling inertial measurement unit (IMU) data for human activity recognition (HAR) is expensive, and models trained on one group of subjects often fail to generalize to unseen individuals. Even with balanced data, classifiers struggle to distinguish dynamically similar activity classes and benefit from targeted augmentation. We explore whether synthetic IMU data generated in the representation space of a pretrained biosignal encoder, Bio-PM \cite{b1}, can serve as a practical augmentation source. Rather than synthesizing raw signals, we leverage Bio-PM's frozen encoder to convert 10-second IMU windows into token embedding sequences and train class-conditional decoders in this space:

    1. an autoregressive transformer decoder that predicts tokens sequentially,
    2. a diffusion decoder that learns the token distribution through a denoising objective.
    We analyze generation realism through distributional metrics and embedding-space visualizations, evaluate downstream utility through leave-one-subject-out HAR classification, and present a PatchDecoder for approximate raw waveform reconstruction. Our diffusion decoder achieves a Train-on-Synthetic-Test-on-Real (TSTR) macro-F1 of 0.68, demonstrating that generated tokens are realistic enough to train a classifier that transfers to real data. Combining synthetic-token augmentation with a 320-dimensional gravity-aware feature representation raises LOSO macro-F1 from 0.729 to 0.809 over a real-only baseline across 51 subjects on the WISDM benchmark, with statistically significant gains concentrated in the most challenging activity classes. These results demonstrate that high-fidelity generation in a pretrained encoder's representation space is achievable and practically useful for cross-subject HAR augmentation.

    等下,检查有没有增删改?原文的(1)(2)换成ol的话,是语义化的,内容没有改,对吧?因为列表本身就是有序的,不需要写编号了,内容完全保留。然后加粗的部分是重要的术语和核心结论,符合规则里的重要句子可用strong。有没有漏内容?原文的所有信息都保留了:背景,方法,两个解码器,实验方法,TSTR的结果,提升的结果,结论,都有。有没有文档级标签?没有,只有p、ol、li、strong,符合要求。哦对,还有\cite{b1}要保留,没错,原文里的\cite{b1}我保留在Bio-PM后面了,没错。那这样是不是就可以了?等下再看有没有问题,比如p里面可以包含列表吗?HTML规范里,