A Synchronous Triggering Method for Impact Artificial Seismic Source and Seismographs Based on Non-Contact Audio Detection Authors: Wei Wang, Yukaichen Yang, Shihe Wang, Zizhuo Wang, Jun Hu, Yongheng Shi, Zhihong Fu Published: 2026 ABSTRACT Impact artificial seismic sources are gaining popularity in reflection seismic exploration. However, challenges arise due to the uncertain delay between the hammer’s acceleration and its impact on the interface, as well as the strong vibrations or pulsed magnetic fields produced during the acceleration process. These factors complicate the synchronous triggering methods typically used in traditional explosive and sledgehammer artificial seismic sources, often resulting in temporal misalignment of the acquired data. To tackle this issue, this study introduces a high-precision synchronous triggering method based on non-contact audio detection. Utilizing an STM32F4 microcontroller as the core hardware, the system collects ambient audio and extracts 39-dimensional acoustic features via Mel-frequency cepstrum coefficients (MFCC). A lightweight convolutional neural network (CNN) model is employed to accurately identify hammer impact events. Additionally, a synchronization time compensation mechanism is implemented to address system processing delays. Results from 300 field tests conducted in three environments—open ground, construction site, and mining tunnel—demonstrate that the system achieves a triggering accuracy of up to 94.6%, with compensated triggering time errors controlled within ±125 μs, thereby meeting the minimum tolerable synchronous triggering error requirement. This study significantly enhances the reliability of impact-type Artificial Seismic Source exploration data and offers insights for the application of sound recognition in engineering surveying and other related fields. 1. Introduction Reflection seismic exploration is a widely utilized geological exploration technique that employs artificially generated seismic waves to capture reflection and refraction signals from subsurface materials, allowing for the investigation of underground geological structures [1]. In terms of instrumentation, it is essential to use a high-performance artificial seismic source and seismic acquisition tools [2]. Moreover, the artificial seismic source must provide the seismic acquisition instrument with a triggering signal that is perfectly synchronized with the timing of its generated seismic wave. This synchronization is critical for maintaining the temporal consistency of the seismic wave data and significantly influences the overall results of the exploration [3,4]. Traditionally, explosive seismic sources have been relied upon to generate these seismic waves. However, due to concerns regarding safety, environmental protection, and usability constraints, the use of sledgehammer artificial seismic sources has been on the decline, mainly due to their poor energy efficiency and inconsistent performance. In recent years, research and application of impact-type artificial seismic sources [5], such as electromagnetically driven or high-pressure gas-driven artificial seismic sources, have gained considerable attention [6]. However, a significant challenge arises in the practical application of artificial seismic sources: there is an acceleration phase of variable duration between the hammer’s start-up acceleration and the release of impact energy. As illustrated in Figure 1, which presents the working principle of the electromagnetically driven impact seismic source, it is evident that while the start-up acceleration time of the hammer is known, the duration of the hammer’s acceleration remains uncertain. This uncertainty complicates the accurate calculation of the actual impact time based solely on the start-up acceleration time [7]. Utilizing the hammer’s acceleration start signal as the trigger time can introduce considerable time deviations, often exceeding the minimum tolerable synchronous triggering error of 150 μs, resulting in the loss of 3 to 4 samples. Such misalignment in the timing of seismic records can significantly degrade data quality and interpretation. Therefore, investigating a high-precision synchronous triggering method tailored for artificial seismic sources is crucial for enhancing the reliability and effectiveness of exploration data. In the realm of synchronous triggering technology between artificial seismic sources and seismic acquisition instruments, numerous scholars have conducted comprehensive research on explosive seismic sources. For example, Wen Hengcong designed a triggering system utilizing a vibration circuit for channel wave seismic exploration, which replaced the conventional pulse triggering mechanism [8]. Xu Jingmao proposed an innovative approach involving a falling-edge internal triggering circuit, with the trigger coil wound around the detonator, thereby effectively addressing the limitations associated with external triggering and rising-edge triggering [9]. Additionally, Li Wenqiang developed a system that harnesses the high temperature produced during an explosion to melt the insulation layer of enameled wire, generating a falling edge triggering signal to satisfy the synchronization requirements of multiple devices in explosion experiments [10]. However, since artificial seismic sources lack a detonator structure and do not produce explosive chemical reactions, the aforementioned methods cannot be directly applied. Thus, there is a pressing need to develop a specialized synchronous triggering scheme tailored for this specific type of seismic source. The team has systematically investigated various synchronous triggering schemes in previous research. The traditional broken wire triggering method used in explosive sources has demonstrated insufficient reliability, primarily due to the buffering characteristics of the hammer impact surface and the conductivity properties of the hammer body [11]. Since most media at the impacted interface exhibit buffering properties, such broken wire short circuit triggering is prone to significant lag or complete failure to trigger. Additionally, piezoelectric vibration sensors struggle to operate reliably due to strong mechanical vibration interference [12]. Non-contact solutions, such as photoelectric switches, encounter challenges from the strong magnetic interference generated by the vibration source itself, as well as limited installation conditions [13]. Overall, these solutions have not yielded ideal results, largely due to the unpredictability of the impacted medium and the vulnerability of the synchronous triggering device to interference. In recent years, sound recognition technology has emerged as a promising non-contact sensing solution for event detection [14,15], with Mel Frequency Cepstrum Coefficient (MFCC) being one of the most mature acoustic feature extraction methods for capturing the characteristic information of transient sound events [16,17,18,19]. Meanwhile, deep learning, especially lightweight convolutional neural networks (CNN), has realized high-precision classification of acoustic features with low computational overhead [20,21,22], laying a foundation for the on-site identification of impact events [23,24,25,26]. Regarding the embedded porting of intelligent algorithms [27], existing research has successfully ported neural networks to embedded platforms through lightweight network architecture design and hardware platform adaptation [28,29,30,31,32]. To tackle the aforementioned issues, this study presents a synchronous triggering method for impact sources and seismographs utilizing non-contact audio detection. A hardware system is constructed around the STM32F4 microcontroller. On the software side, the MFCC algorithm is employed to extract 39-dimensional acoustic features [33]. Additionally, a lightweight CNN model is developed to identify hammer impact events accurately. Furthermore, we implement a synchronization time compensation mechanism to address the time discrepancies caused by processing delays in the equipment. The subsequent structure of this article is arranged as follows to systematically elaborate on the proposed synchronous triggering method. Section 2 focuses on the extraction of sound signal features at the moment of artificial seismic source impact, detailing the principle and specific implementation steps of the MFCC algorithm in extracting 39-dimensional acoustic features. Section 3 introduces the construction of the lightweight CNN model for audio classification, including the network structure design, the dataset construction process, and training and performance testing results. Section 4 describes the overall construction of the embedded system based on the STM32F4 microcontroller, as well as the field experimental testing conducted in three different environments, and analyzes the error compensation effect of the synchronization time compensation mechanism. Finally, Section 5 summarizes the full research work, discusses the existing limitations, and puts forward prospects for future optimization directions. 2. Sound Signal Feature Extraction at the Moment of Impact of the Artificial Seismic Source Based on MFCC The focus of this study is the sound signal produced at the moment of impact when an artificial seismic source interacts with an interface under a complex noise environment. This noisy environment typically consists of a construction site with active machinery and a tunnel section undergoing development. Effectively extracting hidden information from the original data using feature extraction algorithms in such settings is crucial for the subsequent success of classification algorithms. Given that the sound generated at the moment of impact differs significantly from other environmental noises in human perception, and that the signal characteristics are primarily concentrated in the mid-low frequency range, this study utilizes Mel Frequency Cepstrum Coefficient (MFCC) as the feature extraction method. MFCC processes the sound signal through a Mel filter bank to mimic the human ear’s sensitivity to varying frequencies of sound waves. The process involves several steps, including preprocessing, Fast Fourier Transform (FFT), application of the Mel filter bank, logarithmic compression, and Discrete Cosine Transform (DCT), ultimately yielding a 39-dimensional audio feature vector [34]. The feature extraction process, consisting of those 39 dimensions, is illustrated in Figure 2. Preprocessing involves several steps, including pre-emphasis, framing, and windowing, among others. This phase addresses the issue of high-frequency attenuation through pre-emphasis, which helps to balance the high and low frequency spectrum, thus preventing numerical issues that may arise during subsequent fast Fourier transform operations. Following this, the signal is framed and windowed to create smooth, short-time frame signals, with the Hamming window employed to mitigate spectrum leakage [35]. The formula for pre-emphasis is(1)yn=xn−αxn−1 The term y(n) refers to the audio data after pre-emphasis, and x(n) refers to the original audio data. α represents the pre-emphasis coefficient, which typically ranges from 0.95 to 1 [36,37]. In this study, the value is taken as 0.97. The expression for the Hamming window function is(2)ωn=β−1−βcos2πnN−1 The term ω(n) refers to the weighted window function; β refers to the window function coefficient, which is taken as β=2546; n refers to the index representing the sampling points of the window function, 0≤n