As an essential tool in science and technology, wavelength detection plays a vital role in analytical chemistry, bio-sensing and optical communication. The traditional spectrometers based on dispersing components or resonant cavities greatly suffer from bulky size, high power consumption and fabrication imperfection. With the rapid development of micro processing, novel types of high-performance and portable spectrometers emerged, and the pursuit of pushing the performance to limit remains unsettled. Based on the signal transmission theory in multimode fiber, the intensity interference patterns are resulting from the mode interference effect were established in adiabatic and collimated model. In the experimental measurement, the tapered region with a slowly varying slope (about 0.01) was introduced near the end of the fiber to ensure that the side radiation signal could be collected. To estimate the number of modes supported in different structures, both the theoretical and numerical simulations are consistent with the experimental tendency. Using the confocal microscope system we made, the interference patterns are stored by continuous scanning a narrow-band laser. The calibration matrix corresponding to the device’s unique characteristics is obtained by region selection, vector splicing and singular value decomposition. The following wavelength detection process can be divided into two steps: the rough calibration matrix within the working bandwidth is obtained after the rough scanning of the wavelength in 1nm scale, and the wavelength units with the non-zero value are selected as the target after inner product correlation operation with the degraded one-dimensional signal intensity vector. This initial procedure provides the criterion of optimizing the structural parameters. On this basis, fine scanning is performed to obtain the refine calibration matrix. The three largest principal components are selected and defined as the final detected wavelength based on the minimum Euclidean distance. The inner product correlation operation combined with the principal component analysis can improve the wavelength detection resolution to 20 pm with the accuracy rate of 96.7%. The detection efficiency is fifty times higher than other nonlinear spectral reconstruction algorithms reported. The experimental results show that the working bandwidth is at least from 400 to 700 nm, and the device size is only π×(20 μm)²×0.5 mm. The practical feasibility and photon detection are also investigated, considering its further application. Compared with its counterparts, this device has a significant improvement in high performance, portability and low cost, it also integrates with an efficient algorithm in wavelength detection procedures. Both device and theory could be widely used in real-time wavelength detection of optical fiber transmission systems.