基于多层感知机的无线射频标签数量估计方法

In recent years, with the price decline of RFID tags and the growth of supply chain demand, the RFID technology has moved towards the next full deployment stage, which requires the deployment of tags on individual commodities to improve the accuracy of supply chain tracking. This trend will lead to an exponential increase in the number of RFID tags, increasing the pressure on reader for inventory counting. Usually, the reader needs to identify 96-bit ID of a tag for checking if they exist. However, transmitting ID via low-rate RFID channel takes too much time. Considering that many existing RFID applications do not require explicit ID information but quantity information for production and logistics decision-making, the tag cardinality estimation problems have therefore attracted the most attention in recent years. Most of the existing tag cardinality estimation schemes only need tags to respond a binary or short messages and let the reader observe the statical features of the time frame for performing quantity estimation. The researchers have proposed various estimators, which reveals the explicit relationship between the statical feature of the time frame and the number of tags, including the number of empty/non-empty slots, the index of first empty /non-empty slot, the average length of runs of empty/nonempty slots, the difference between empty/non-empty slots. Although all these estimators arc unbiased estimators of tag quantity, it is still challenging to derive an accurate and stable estimation result due to the large estimation variance caused by random noise. Moreover, all these estimators have limited applicable range, and their performance degrade quickly once the actual tag quantality out of the estimation range. To this end, this paper designs an Al-driven tag estimation method based on multi-layer pcrccptrons (MLP), a simple fully-connected neural network, which prompts the accuracy and stability of estimator by fusing diversified statical features of time frame. We also propose a time frame generator to simulate the slot status of time frame with given tag quantality and frame length. By generating numerous simulated time frames of the same length with respect to each tag quantality and feeding them into the training process of the neural networks, we can obtain a MLP estimator for the certain length of time frame. Additionally, to overcome the limited range of the MLP estimator, we designed a two-sage MLP estimation protocol based on the idea of sampling theory. Specifically, the first stage following the binary tree search idea to gradually increases the length of the masks for decreases the number of tags activated by the matched mask. The reader can obtain a roughly quality estimation results based on the index of first empty slots and sets a reasonable sampling probability for the second stage accordingly. Such a sampling process limits the number of tags replying within the working range of the MLP estimator and promise the estimator can provide an accurate estimation result under various tag populations. The extensive simulation results under various scenario demonstrate that the proposed MLP estimation protocol can improve the estimation accuracy by at least 21% compared with existing state-of-the-art methods.