With the global adoption of Internet-of-Things (IoT) technologies, inexpensive IoT devices with various IoT wireless protocols are deployed in license-free ISM bands. These Low-Power Wide-Area Network (LPWAN) signals can be in open or proprietary standards with different security implications. Preambles that are prefixed in the packets are often found in these protocols and are normally used for tasks like signal detection and Automatic Gain Control (AGC). In terms of unintended uses, they can be used for device identification, reverse engineering of protocol features, intrusion detection of wireless attacks, performing jamming attacks, and more. Therefore, it is important to localize and extract the preamble portion of LPWAN signals for physical layer reverse engineering, especially for proprietary protocols and further enhancing the security of the spectrum. In this regard, we provide an automatic preamble extraction system that seeks the common prefix of arbitrary frequency-varying LPWAN signals from the IQ data received by Software-Defined Radios (SDRs). Existing methods usually employ supervised machine learning algorithms with labeled data for training. Our algorithm eliminates the need to use labeled data and only relies on the constant nature of the preamble in the spectrogram to perform extractions. Experimental results demonstrate an average precision (IoU@0.75) of up to 96.3% for LoRa signals with varying bandwidth, spreading factors, data, and preamble lengths.