Keywords

1 Introduction

In the Banded Iron Formation (BIF) hosted iron ore deposits located in the Pilbara region of Western Australia, natural gamma logs are frequently used to determine the location of stratigraphic boundaries [1]. The location of these boundaries is required for accurate modelling of the deposit. These deposits consist of layers of BIF and shale, with sections of the BIF enriched to create a minable high grade ore [2]. This paper studies a typical Marra Mamba type iron ore deposit, with iron ore in the Mount Newman Member overlain by the shale in the West Angelas Member.

The natural gamma logs are one of several geophysical logs that are typically collected in exploration holes, along with density and magnetic susceptibility. As thousands of holes may be drilled for a single deposit, manually interpreting them can be an arduous process. Additionally, it is prone to inconsistencies between different interpreters as well as errors. Therefore an automatic method that can assist the interpretation is greatly desired. A computer-aided method is a better option as it can provide a fast, objective, and reliable analysis. Many studies have been conducted on the development of automatically identified boundaries based on machine learning applications, including several studies on these types of deposit [3, 4]. However, the process of choosing a set of optimal features to classify a signature from the signal is very difficult. Therefore, a deep learning technique is presented in this study to overcome the challenges faced by conventional automated systems. A deep learning network involves several stages of learning processes, including an input layer, hidden layers, and an output layer [5]. The network takes unprocessed data as the input and learns the representative features that needed for classification without user input. The network is trained using the backpropagation algorithm.

In this study we investigated the capability of a long short-term memory (LSTM) network to classify several geological signatures from a gamma log. LSTM is a type of recurrent neural network mostly used to analyse time series sequence data [6]. In this study the authors investigated the capability of a LSTM network to classify several geological signatures from a gamma log sequence data along the down hole samples. The work demonstrates a classifier with three different outputs that can differentiate AS1-AS2 signals and NS3-NS4 signals from long gamma sequence.

2 Data Used

Natural gamma logs from a typical Marra Mamba style iron ore deposit were used. This deposit contains two natural gamma signatures of particular interest. The first signature is produced by the AS1-AS2 shales at the base of the West Angelas Member (Fig. 1a) and marks the transition between the shale and the Mount Newman ore below. The second signature comes from the NS3-NS4 shale bands (Fig. 1b) that mark a transition between two geological units within the Mount Newman Member. 42 examples of each of these signatures were chosen, along with 42 examples of other signatures from different parts of the natural gamma logs.

Fig. 1
figure 1

Typical examples of the natural gamma signatures for a AS1-AS2, and b NS3-NS4. c, d Contain two examples of other signatures, there is a large variety of signatures in this group

Full size image

3 Methodology

3.1 Long Short-Term Memory (LSTM)

In order to train a deep neural network to classify each signal along the depth of sequence data, a sequence-to-sequence LSTM network was used as proposed in Matlab R2019a [7]. The existing method was used to make prediction for each time step of the sequence. However, in this study, the method was used to make prediction of depth sequence data.

The proposed method consists of a LSTM layer, with tanh as the activation function. LSTM layer was set to have 200 hidden units and output the full sequence. The final layer, which is a fully connected (FC) layer with a softmax activation function, is used as the classifier with three classes. The classifier uses the output from the LSTM as the input and predicts the class label of the gamma signatures for a given sequence. The ‘adam’ optimizer was used and training was happened with different number of epochs.

3.2 Training and Validation

In order to analyze the capability of the LSTM architecture for identifying gamma signatures, a range of tests were performed. Out of 42 sequences, 30 sequences were used as training samples and rest were used as test samples. Two training libraries were used. In the first library all samples had the same signal order. The other library had two different sequence sets: AS1-AS2, other and NS3-NS4 in order, NS3-NS4, AS1-AS2 and other in order.

Figure 2 shows twelve test samples each containing a combination of AS1-AS2 and NS3-NS4 signatures along with signatures from other sections of the logs. The test sequences have the AS1-AS2, NS3-NS4 and other signals in different order. The first six test sequences have the signals in the order AS1-AS2, other and NS3-NS4, and the other six sequences have the order NS3-NS4, AS1-AS2 and other. To test the robustness of this method to noise, the test was re-run with random noise added to the test samples.

Fig. 2
figure 2

Test sequences with the combination of AS1-AS2 (orange), NS3-NS4 (purple) signatures and other (yellow). Samples with (a) no noise and (b) noise added. Blue lines show the joins between the different signatures

Full size image

In total four tests were completed. Case 1 and Case 2 were trained using libraries 1 and 2 respectively and were both stopped at epoch 200. Case 3 was trained with library 2 and stopped at epoch 228. Case 4 used the same training as Case 3, but predicted the signature classes of samples with added noise.

4 Results and Discussion

The accuracy and loss of the proposed method from the training as a function of epochs can be seen in Fig. 3. Each plot is associated with different training library and different epochs. In case one library 1 was used to train the net. Then the net was used to predict the signatures in the original test samples (no noise added) (Fig. 2a). In other cases, library 2 was used in the training process with epochs 200 and 228. Prediction accuracy was tested on both original and noisy samples (Fig. 2). In all scenarios a high accuracy were achieved during the training process (Fig. 3). The performance is summarized in Figs. 4 and 5.

Fig. 3
figure 3

Performance from the proposed LSTM method. Left: accuracy versus epochs; Right: Loss versus epochs; a Training with library 1: AS1-AS2, Other, NS3-NS4, training accuracy at epoch 200. b Training with library 2: order 1: AS1-AS2, Other,NS3-NS4; order2: NS3-NS4, AS1-AS2, other at epoch 200. c Training with library 2 manually terminated at epoch 228

Full size image
Fig. 4
figure 4

Comparing predicted class labels (blue) with given class labels (red) for different scenarios: a Training with library 1, epoch 200, b Training with library 2, epoch 200, c Training with library 2, epoch 228, d Training with library 2, epoch 228 and prediction on noisy samples. X axis: depth; Y axis: class labels, 1:3 represent the AS1-AS2 signatures, other signatures and NS3-NS4 signatures respectively

Full size image
Fig. 5
figure 5

Summaries of the prediction accuracy on 12 samples for each case. Case1: Training with library 1, epoch 200, Case 2 Training with library 2, epoch 200, Case 3 Training with library 2, epoch 228, Case 4 Training with library 2, epoch 228 and prediction on noisy sample

Full size image

In Fig. 4a it can be seen that the classification accuracies are significantly higher in the first 6 sequences compared to the remaining six sequences. This is because the net used in prediction was trained with library 1, which has all signals in the same order. Without the inclusion of different signal orders in the training process, the LSTM method failed to identify signals where the order was different. In comparison, when training with mixed order samples the LSTM closely predicted the signals in 11 out of 12 samples (Fig. 4b). The accuracy of the prediction was further improved at epoch 228, where the author stopped the training where accuracy of the network peaked (Fig. 4c). The same trained network obtained relatively good accuracy in the noisy test samples as well (Fig. 4d.). In all scenarios test sample 7 yielded a much lower accuracy. This is because the AS1-AS2 signature in that sample is quite different to the other AS1-AS2 signatures.

5 Conclusion

In this paper, a deep learning based solution using LSTM is developed for natural gamma signature identification. The proposed method learns patterns from gamma sequences, and could successfully identify the significant AS1-AS2 and NS3-NS4 signatures. With only a few training samples it generally achieved a relatively high accuracy. However, the accuracy was dependent on the signal order present in the training samples. Further evaluation is needed with increasing training samples with different sequence lengths.