With the rapid development and application of the Internet of Things (IoT) and indoor activities, high-precision indoor positioning technology based on location services has a wide range of applications. Given that GPS and Beidou signals lead to signal attenuation when penetrating buildings, it is impossible to realize accurate indoor positioning, and an effective indoor positioning method is urgently required to compensate for the vacancy of high-precision indoor positioning. Compared with other indoor positioning methods, visible light indoor positioning, based on the received signal strength, can be used as an effective indoor positioning method owing to its advantages of low cost, high precision, and ease of deployment. However, the existence of multipath effect, shadow change, receiver thermal effect, and other problems can lead to fluctuations in the strength of the received indoor visible light positioning signal, and thereby, resulting in large positioning errors. Furthermore, in extant studies, researchers tend to solely examine one positioning unit and assume that it can be completely copied to other positioning units. However, the migration from one positioning unit to other positioning units may lead to high positioning errors due to different LED positions, varying noise levels, and other differences. Therefore, it is important to solve the jitter problem of the received signal and improve the accuracy of indoor positioning.

To address the problem of jitter in received signals, in this study, a convolutional neural network was proposed based on an attention mechanism (Fig. 4) to reduce the impact of fluctuations in received signals. First, a fast Fourier transform was used to preprocess the received time-domain signal strength values, and the power spectra of the signals were obtained. A CNN with an attention mechanism was used to extract the features of the signal power spectrum, and a channel attention module (Fig. 5) was used to increase the weights for each channel to reduce the influence of redundant information on the network. This in turn reduces the influence of signal fluctuations on positioning accuracy. To solve the problem, in which migrating to different localization units may decrease the localization accuracy, transfer learning was used to migrate the network trained in the first localization unit to other localization units. Based on the premise that the general features of each location unit are similar, the parameters of the attention convolutional neural network trained in the first unit were maintained as unchanged, and only the final fully connected layer was updated. This can reduce the cost of the training network without changing the location accuracy.

A simulation environment of 10 m×10 m×3 m (Table 1) is divided into four positioning units. The proposed algorithm is simulated and compared with the convolutional neural network algorithm. The simulation results (Fig. 7?Fig. 10) indicate that the proposed algorithm can realize 3D positioning with an average error of 3.54 cm in a positioning unit of 5 m× 5 m×3 m. The average error of the CNN algorithm is 4.25. Furthermore, through the introduction of transfer learning, the proposed neural network model can be deployed more easily in other positioning units with an average error of 3.67 cm. Additionally, the training time of the neural network in the other positioning units is significantly shortened, which can effectively reduce the time and computing costs during network deployment. A 1.5 m×1.2 m×1.2 m positioning platform (Table 2) is built and divided into two positioning units for testing (Fig. 12). A comparison with the comparative algorithms shows that the algorithm proposed in this study can effectively reduce the impact due to the fluctuation of the received signal strength, with an average error of 3.32 cm and 90% of the errors are within 4.12 cm. When transfer learning is deployed to the second location unit, the average error is 3.35 cm, and the location performance does not deteriorate. Based on simulations and experiments, it is proven that the proposed algorithm exhibits excellent performance in terms of convergence speed and anti-jitter of the received signal.

In this study, a convolutional neural network (CNN) algorithm based on an attention mechanism is proposed to realize three-dimensional indoor positioning. By preprocessing the received signal strength in the time domain and iterating the channel information using the attention mechanism, the algorithm effectively improves the positioning accuracy and fitting rate of the neural network and reduces the training cost of the neural network. To reduce the positioning accuracy after application to different positioning units, transfer learning is deployed to other positioning units. Compared with the convolutional neural network model, the proposed algorithm can effectively improve the positioning accuracy and fitting rate. Compared with the comparative algorithms, the proposed algorithm is not only more accurate, but can also maintain positioning accuracy when deployed to other positioning units. Simulation and experimental results show that the proposed algorithm can effectively reduce the influence of the received signal strength fluctuation on the positioning accuracy and improve the positioning accuracy.