The proper classification of emergent and impulsive noise signals is critical for reducing false detections of microearthquakes and developing a complete understanding of ongoing ground motions in the shallow crust. Continuous seismic waveforms contain numerous natural and anthropogenic signals whereas tectonic seismic events occupy only a small percentage of each day. A dense array of 1,100 vertical geophones recorded ground motions at 500 samples per second along the San Jacinto fault zone for 30 days in 2014. The data provides detailed waveforms to detect microearthquakes and observe surface/atmospheric processes that manifest as impulsive and emergent seismic signals. Efforts to detect seismic events using a shallow architecture Random Forest model results in a 72% increase from the hand-picked catalog using regional broadband sensors. Recent studiesutilizing the spatially dense seismic array have demonstrated that ongoing low-amplitude seismic motion is dominated by various weak sources originating at the surface from anthropogenic and atmosphere interaction. Labeling new classes of waveforms from wind generated ground motions, air-traffic, automobiles, and other non-tectonic signals can provide insightful information for designing a machine learning training data set to efficiently classify continuous records. We apply a new methodology that uses subtle changes in correlations to label continuous waveforms as random noise, non-random noise, or a mixture of signals, and produce millions of 4 second labeled waveforms. The training data is generalized by selecting a distribution of sensors from different days to account for the daytime/nighttime variations and site-specific noise amplitude differences. The waveforms are used to calculate the short-time Fourier transform as an input to train a convolutional neural network (ConvNet). The ConvNet contains 4 convolution layers and 2 fully connected layer using rectified linear unit activation functions on each layer and a softmax activation function for the output layer. We focus our efforts on identifying different classes of non-tectonic signals in the non-random noise using unsupervised learning techniques to subdivide the noise signals and build a new training data set that represents the variability in the waveforms. The classification model is applied to continuous waveforms to develop a time series of new classes of noise signals along an active fault. The results of coherent signals across the array provide insight on shallow crustal deformation and surface generated ground motions.