Landslide signal identification system based on titanium dioxide memristor dynamic reservoir calculation

Abstract

The invention discloses a landslide signal identification system based on titanium dioxide memristor dynamic reservoir calculation. The landslide signal identification system is characterized by comprising a data acquisition module, a preprocessing module and a dynamic memristor network which are connected in sequence, the data acquisition module is used for acquiring original seismic wave signals with time sequence characteristics and extracting seismic wave time sequence signals in three directions from the original seismic wave signals; the preprocessing module is used for synchronously preprocessing the seismic wave time sequence signals in three directions to generate seismic wave signals of three channels; each input unit in the dynamic memristor network is used for acquiring a seismic wave signal of one channel; processing the seismic wave signal through a corresponding storage unit to obtain a memristor response current signal of each channel; and finally, integrating all memristor response current signals of the three channels by an output layer to obtain a landslide signal identification result. The landslide monitoring precision of the system can be improved.

Description

Technical Field

[0001] This invention relates to the field of landslide signal recognition technology, and in particular to a landslide signal recognition system based on dynamic reservoir calculation using a titanium dioxide memristor. Background Technology

[0002] Landslides are sudden and highly destructive geological hazards, and accurate detection and prediction are crucial for minimizing casualties and economic losses. On January 22, 2024, a landslide in Zhenxiong County, Yunnan Province, resulted in 44 deaths and more than 42 missing persons. Furthermore, the landslide buried 18 houses, causing significant direct economic losses. This event further underscores the urgency of establishing landslide monitoring and early warning systems to mitigate the impact of such disasters.

[0003] Landslide mechanisms involve the rapid sliding of debris or rock masses, typically generating strong frictional forces. These forces propagate through the ground, producing seismic waves, such as… Figure 2 As shown in Figure C, traditional landslide monitoring methods primarily rely on the differences in seismic wave propagation between the landslide body and surrounding geological materials. Common landslide identification techniques include spectral analysis, seismic-landslide dynamics inversion, microseismic methods, seismic environmental noise correlation, and machine learning methods. In principle, due to the presence of loose debris and soft rock within the landslide body, the seismic wave propagation velocity in the landslide area is typically low, which slows down the seismic wave velocity. Furthermore, the loose material leads to significant seismic wave attenuation, causing changes in amplitude and frequency characteristics, which can be used to detect landslide events. These methods rely on human expertise to define characteristic rules and are susceptible to noise interference and waveform distortion in complex geological environments, which may affect their identification accuracy.

[0004] With the development of deep learning technology, recurrent neural networks (RNNs), due to their ability to model time dependencies, have been applied to landslide waveform analysis, aiming to automate feature extraction and improve the accuracy and efficiency of identification. However, traditional RNNs face challenges when processing long-term sequences, such as the vanishing and exploding gradient problems, especially when detecting seismic waves triggered by landslides. This significantly limits their performance, leading to higher prediction errors. Furthermore, RNNs have high computational complexity, requiring significant resources during training. The backpropagation mechanism in RNNs also increases time overhead, making it difficult to meet the low-latency requirements of real-time early warning systems, such as… Figure 2 As shown in Figure A. Furthermore, RNNs are sensitive to noise and highly dependent on data distribution, which affects their generalization ability in complex and variable environments. Especially under different geological conditions, RNNs often require retraining to maintain optimal performance.

[0005] Disadvantages of existing technologies: Traditional monitoring methods mainly work by analyzing seismic waveforms, but these methods are easily affected by noise and distortion, especially in environments with complex geological conditions, making it difficult to ensure the accuracy and reliability of landslide monitoring, resulting in low landslide monitoring precision. Summary of the Invention

[0006] The present invention provides a landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors, which can improve the accuracy of landslide monitoring.

[0007] To achieve the above objectives, the present invention provides a landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors. The key feature of this system is that it is equipped with a data acquisition module, a preprocessing module, and a dynamic memristor network connected in sequence.

[0008] The dynamic memristor network is provided with an input layer, a storage layer and an output layer connected in sequence. The input layer has three input units arranged in parallel, and the storage layer has three storage units arranged in parallel. The output terminals of the three input units are connected to the input terminals of the three storage units in a one-to-one correspondence. The output terminals of the three storage units are all connected to the input terminals of the output layer.

[0009] The data acquisition module is used to acquire raw seismic wave signals with time series characteristics, extract seismic wave time series signals in three directions from the raw seismic wave signals, and then transmit them to the preprocessing module.

[0010] The preprocessing module is used to synchronously preprocess the seismic wave time series signals in three directions to generate seismic wave signals in three channels and transmit them to the dynamic memristor network.

[0011] Each input unit in the dynamic memristor network is used to acquire the seismic wave signal of one channel respectively; then the seismic wave signal is processed by the corresponding storage unit to obtain the memristor response current signal of each channel; finally, the output layer integrates all the memristor response current signals of the three channels to obtain the landslide signal identification result.

[0012] Through the above design, this invention constructs a reservoir computing system based on dynamic memristors. It utilizes the nonlinear current response and short-term memory characteristics of memristors as the physical reservoir of the system to capture the temporal characteristics of landslide signals. Changes in the reservoir state accurately reflect the dynamic changes of the input seismic wave signal. Because memristors have short-term memory capabilities, they can retain historical information when processing continuous data, thereby improving the system's ability to handle signal time dependence. The nonlinear current response and short-term memory characteristics enable memristors to quickly adapt to rapid changes in landslide signals, improving the accuracy of real-time monitoring and prediction of landslide events.

[0013] Furthermore, by designing parallel reservoir computation, this invention enables the system to comprehensively analyze multidimensional landslide signals, effectively improving the system's ability to extract signal features and further enhancing the system's landslide monitoring accuracy.

[0014] Preferably, the seismic wave time series signals extracted by the data acquisition module in three directions are: east-west (BHE) seismic wave time series signal, north-south (BHN) seismic wave time series signal, and vertical (BHZ) seismic wave time series signal.

[0015] Preferably, the preprocessing module is provided with a mean removal unit, a fast Fourier transform unit, a power spectral density analysis unit, a filtering unit, and a signal segmentation unit connected in sequence.

[0016] The mean-reduction unit is used to subtract the average value of each signal in the seismic wave time series signal in each direction, so that the signal fluctuates near zero. By subtracting the average value of each signal, the mean-reduction unit effectively removes the DC component in the seismic wave time series signal in each direction, thereby enhancing the signal variation characteristics.

[0017] The Fast Fourier Transform (FFT) unit is used to perform a Fast Fourier Transform on the signal processed by the Mean Removal Unit, converting the signal from the time domain to the frequency domain and revealing the energy distribution of the signal in different frequency ranges.

[0018] The power spectral density analysis unit is used to identify the frequency range distribution of the signal after the fast Fourier transform, and to identify landslide-related signals and non-landslide signals in each direction of the signal.

[0019] The filtering unit is used to filter the landslide-related signals and non-landslide signals in each direction signal respectively, and then sample the filtered signals.

[0020] The signal segmentation unit is used to divide the sampled landslide-related signals and non-landslide signals into at least two time slices, each time slice containing signals of m time steps, and then divide all time slices into three channels of seismic wave signals according to the signal direction.

[0021] Preferably, the filtering unit uses a bandpass filter with a frequency range of 5Hz to 50Hz to filter the landslide-related signals; the filtering unit uses a filter with a cutoff frequency of 10Hz to filter the non-landslide signals; and the filtering unit samples the filtered signals at a sampling rate of 1000Hz.

[0022] Based on the spectral analysis results, the filtering unit selected appropriate filtering strategies for different signal types. For landslide-related signals, their frequency components are typically concentrated between 5Hz and 50Hz and fluctuate significantly; therefore, a bandpass filter with a frequency range of 5Hz to 50Hz was used to effectively retain the main frequency components of the landslide signal while removing irrelevant frequencies. For non-landslide signals, their frequency components are lower and change more gradually; a filter with a cutoff frequency of 10Hz was used to attenuate high-frequency noise while maintaining the low-frequency characteristics of the signal. To prevent aliasing, the data was sampled at a sampling rate of 1000Hz to ensure that the signal was fully sampled in the frequency domain and to avoid distortion.

[0023] Preferably, each input unit in the input layer is used to normalize the time slice of the corresponding channel, and then a linear transformation is performed to match the signal in the time slice with the input voltage range of the reservoir memristor model, and the linearly transformed seismic wave signal is transmitted to the corresponding storage unit. The input voltage range of the reservoir memristor model is [-3.0V, 3.0V].

[0024] Preferably, each of the memory cells is composed of a single memristor, which is an Au / TiO2 / Pt dynamic memristor.

[0025] This invention utilizes memristors as physical storage units in reservoir computing systems, specifically dynamic memristor networks, leveraging their nonlinear current response and short-term memory characteristics—inspired by biology—to effectively mimic the behavior of neurons. This makes memristors ideally suited for time-series signal processing. The short-term memory characteristic of memristors allows them to exhibit time- and history-dependent behavior during information processing, thus simulating the learning capabilities of neural networks. Furthermore, memristors can efficiently process information with low power consumption, while offering high integration and excellent scalability, giving them unique advantages in time-series signal processing tasks. Compared to traditional electronic components, memristors possess stronger memory characteristics and higher signal processing capabilities, significantly improving the accuracy and real-time performance of landslide signal identification systems. These characteristics make memristors highly effective in practical applications, particularly in tasks requiring rapid response and historical information dependence.

[0026] Preferably, the storage unit utilizes the nonlinearity and short-term memory effect of the Au / TiO2 / Pt dynamic memristor to process the seismic wave signal from the input layer using the memristor response. At each time step, the reservoir unit responds to the input signal by continuously updating the conductance and iteratively updates the memristor state at each time step based on the conductance response to obtain the memristor response current signal of the corresponding channel.

[0027] After the storage unit completes the iterative processing of m time step signals in a time slice, it obtains the memristor response current signals corresponding to the m time step signals. Then, it flattens all the memristor response current signals in the time slice into a 1×m one-dimensional feature vector and outputs it to the output layer.

[0028] Preferably, the expressions for the conductivity response and current response in the reservoir unit are as follows:

[0029]

[0030]

[0031] Where G represents the conductance of the memristor at the current time step; G′ represents the conductance of the memristor at the previous time step; α, β, r, and K represent fitting parameters, with K taking the positive half-axis value when V>0. p When V < 0, K takes the value of the negative half-axis curve fitting parameter K. n V represents the input voltage, i.e., the seismic wave voltage signal input at the current time step; G0 represents the initial conductance value, when V>0, G0=1; when V<0, G0=0; G th denoted by the conductance threshold; I represents the response current value at the current time step; e represents the base of the natural logarithm function.

[0032] Preferably, the output layer is a multilayer perceptron classifier (MLP), which integrates 1×m one-dimensional feature vectors from three directions of the reservoir to obtain a 3×m two-dimensional feature matrix, and then outputs the landslide signal identification result through the sigmoid activation function.

[0033] The three one-dimensional feature vectors used for integration are seismic wave response signals from different directions collected during the same time period.

[0034] Preferably, the landslide signal identification result ranges from [0,1]. When the value of the landslide signal identification result is greater than the landslide probability threshold, it indicates that a landslide has occurred; otherwise, it indicates that no landslide has occurred. By comparing with the threshold, the occurrence of a landslide can be accurately determined.

[0035] The beneficial effects of this invention are:

[0036] This invention constructs a dynamic memristor network that effectively captures the temporal characteristics of landslide signals by utilizing the nonlinear current response of memristors in the reservoir. This allows for accurate reflection of the dynamic changes in the input seismic wave signal through variations in the reservoir state. Because memristors possess short-term memory, they retain historical information when processing continuous data, thereby improving the system's ability to handle signal time dependence. This characteristic enables memristors to rapidly adapt to rapid changes in landslide signals, improving the accuracy of real-time monitoring and prediction of landslide events.

[0037] Furthermore, by designing parallel reservoir computation, this invention enables the system to comprehensively analyze multidimensional landslide signals, effectively improving the system's ability to extract signal features and further enhancing the system's landslide monitoring accuracy. Attached Figure Description

[0038] Figure 1 This is a block diagram of the system structure of the present invention;

[0039] Figure 2 This is a schematic diagram of a landslide identification RC system based on memristors in the embodiment;

[0040] Figure 3 This is a schematic diagram illustrating the landslide identification process and comparative experimental results in the embodiments;

[0041] Figure 4 The diagram shows the structure and composition characterization of the Au / TiO2 / Pt memristor device in the embodiment.

[0042] Figure 5 This is a schematic diagram of the manufacturing process of the Au / TiO2 / Pt memristor in the embodiment;

[0043] Figure 6 This is a dynamic mode diagram of the Au / TiO2 / Pt memristor in the embodiment;

[0044] Figure 7 This is a single-sided amplitude spectrum analysis curve of the Fast Fourier Transform (FFT) in the embodiment;

[0045] Figure 8 The graph shows the power spectral density (PSD) analysis and the selection of filtering strategies in the embodiments.

[0046] Figure 9 This is a diagram showing the effect of 20 consecutive 0100 pulses on the memristor response current in the embodiment.

[0047] Figure 10 The diagram shows the current response of the Au / TiO2 / Pt memristor to different 4-bit pulse stimulation modes in the embodiment.

[0048] Figure 11 The confusion matrix diagrams are shown for different slice lengths (10 to 100 time steps) in the embodiments;

[0049] Figure 12 This is a summary chart of performance indicators for different slice lengths in the examples. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0051] like Figure 1 As shown, a landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors is provided with a data acquisition module, a preprocessing module, and a dynamic memristor network connected in sequence.

[0052] The dynamic memristor network is provided with an input layer, a storage layer and an output layer connected in sequence. The input layer has three input units arranged in parallel, and the storage layer has three storage units arranged in parallel. The output terminals of the three input units are connected to the input terminals of the three storage units in a one-to-one correspondence. The output terminals of the three storage units are all connected to the input terminals of the output layer.

[0053] The data acquisition module is used to acquire raw seismic wave signals with time series characteristics, extract seismic wave time series signals in three directions from the raw seismic wave signals, and then transmit them to the preprocessing module.

[0054] The preprocessing module is used to synchronously preprocess the seismic wave time series signals in three directions to generate seismic wave signals in three channels and transmit them to the dynamic memristor network.

[0055] Each input unit in the dynamic memristor network is used to acquire the seismic wave signal of one channel respectively; then the seismic wave signal is processed by the corresponding storage unit to obtain the memristor response current signal of each channel; finally, the output layer integrates all the memristor response current signals of the three channels to obtain the landslide signal identification result.

[0056] The data acquisition module extracts seismic wave time series signals in three directions: east-west (BHE) seismic wave time series signal, north-south (BHN) seismic wave time series signal, and vertical (BHZ) seismic wave time series signal.

[0057] The preprocessing module is provided with a mean removal unit, a fast Fourier transform unit, a power spectral density analysis unit, a filtering unit, and a signal segmentation unit connected in sequence.

[0058] The mean-reduction unit is used to subtract the average value of each signal in the seismic wave time series signal in each direction, so that the signal fluctuates near zero.

[0059] The Fast Fourier Transform (FFT) unit is used to perform a Fast Fourier Transform on the signal processed by the Mean Removal Unit, converting the signal from the time domain to the frequency domain.

[0060] The power spectral density analysis unit is used to identify the frequency range distribution of the signal after the fast Fourier transform, and to identify landslide-related signals and non-landslide signals in each direction of the signal.

[0061] The filtering unit is used to filter the landslide-related signals and non-landslide signals in each direction signal respectively, and then sample the filtered signals.

[0062] The signal segmentation unit is used to divide the sampled landslide-related signals and non-landslide signals into at least two time slices, each time slice containing signals of m time steps, and then divide all time slices into three channels of seismic wave signals according to the signal direction.

[0063] The filtering unit uses a bandpass filter with a frequency range of 5Hz to 50Hz to filter the landslide-related signals; the filtering unit uses a filter with a cutoff frequency of 10Hz to filter the non-landslide signals; the filtering unit samples the filtered signals at a sampling rate of 1000Hz.

[0064] Each input unit in the input layer is used to normalize the time slice of the corresponding channel, and then to match the signal in the time slice with the input voltage range of the reservoir memristor model through linear transformation, and then transmit the linearly transformed seismic wave signal to the corresponding storage unit.

[0065] Each of the aforementioned memory cells is composed of a single memristor, which is an Au / TiO2 / Pt dynamic memristor.

[0066] The storage unit utilizes the nonlinearity and short-term memory effect of the Au / TiO2 / Pt dynamic memristor to process the seismic wave signal from the input layer using the memristor response. At each time step, the reservoir unit responds to the input signal by continuously updating the conductance and iteratively updates the memristor state at each time step based on the conductance response to obtain the memristor response current signal of the corresponding channel.

[0067] After the storage unit completes the iterative processing of m time step signals in a time slice, it obtains the memristor response current signals corresponding to the m time step signals. Then, it flattens all the memristor response current signals in the time slice into a 1×m one-dimensional feature vector and outputs it to the output layer.

[0068] The expressions for the conductivity response and current response in the reservoir unit are as follows:

[0069]

[0070]

[0071] Where G represents the conductance of the memristor at the current time step; G′ represents the conductance of the memristor at the previous time step; α, β, r, and K represent fitting parameters, with K taking the positive half-axis value when V>0. p When V < 0, K takes the value of the negative half-axis curve fitting parameter K. n V represents the input voltage, i.e., the seismic wave voltage signal input at the current time step; G0 represents the initial conductance value, when V>0, G0=1; when V<0, G0=0; G th denoted by the conductance threshold; I represents the response current value at the current time step; e represents the base of the natural logarithm function.

[0072] The output layer is a multilayer perceptron classifier (MLP), which integrates 1×m one-dimensional feature vectors from three directions of the reservoir to obtain a 3×m two-dimensional feature matrix, and then outputs the landslide signal identification result through the sigmoid activation function.

[0073] The original data used for training and testing in this embodiment comes from the Significant Landslide Seismic Dataset (2000-2023). This dataset includes seismic observation records of 17 significant landslide events that have occurred since 2000, covering landslides of different sizes and types. This embodiment rigorously screened the dataset, ultimately selecting 14 landslide events, as shown in Table 1.

[0074] Table 1 shows the landslide event information in the dataset.

[0075]

[0076] These landslide events originate from diverse geological environments, such as high-altitude mountains, seismic fault zones, and karst landforms, ensuring broad applicability in landslide research under various geological conditions. From these events, 130 seismic signal data points containing three components (BHE, BHN, BHZ) were extracted. The data were labeled and preprocessed to ensure suitability for further analysis. Furthermore, further pruning and mixing were performed to create a new dataset for landslide signal identification. In the signal processing stage, landslide seismic signals typically exist as time-series data in the directions BHE, BHN, and BHZ. The processed seismic waveform data are separated by direction, and the signal in each direction is pruned and mapped into a pulse sequence input to the corresponding physical reservoir for processing. Once the signal enters the memristor, a nonlinear current response is triggered within the device. The short-term memory characteristic of the memristor allows it to store characteristic information of the input signal in the reservoir state. After processing each reservoir, the responses of all directions are integrated through parallel computation. The state of each reservoir is adjusted according to the input signal, thus reflecting the temporal characteristics of the signal.

[0077] like Figure 3 As shown, Figure 3 A is a flowchart of the input data preprocessing process, which shows the steps from raw data to processed results. Figure 3 B represents the BHE, BHN, and BHZ seismic waveform data of the landslide disaster in Bijie, Guizhou Province, at 3:57 AM on May 8, 2022. Figure 3 C represents the data processing flow of the BHE seismic waveform from the Bijie landslide disaster in Guizhou. From top to bottom, it displays the original BHE signal, the BHE signal after subtracting the average value, and the filtered BHE signal. These signals are categorized into landslide and non-landslide types. The signals are cropped at 100 time steps. Figure 3 D is a schematic diagram of the landslide disaster identification process of the landslide signal identification system.

[0078] Figure 7 This paper demonstrates Fast Fourier Transform (FFT) single-amplitude spectral analysis for analyzing raw signals in the BHE band. In this embodiment, FFT analysis was performed on the raw BHE band signal during the landslide disaster in Bijie, Guizhou Province on May 8, 2022, to explore the signal's frequency characteristics. FFT transforms the signal from the time domain to the frequency domain, revealing the energy distribution of the signal at different frequencies. Figure 7 The single-amplitude spectrum of the signal is shown, indicating that the signal amplitude is relatively high in the low-frequency range (0-50Hz), mainly concentrated in the low-frequency region. Above 100Hz, the signal amplitude drops rapidly, indicating that the energy in the high-frequency range is relatively low during the landslide. This frequency distribution characteristic is of great significance for subsequent signal processing.

[0079] Figure 8This study demonstrates power spectral density (PSD) analysis and the selection of filtering strategies. The Welch method was used to perform PSD analysis on the original signal, revealing the energy distribution at different frequencies. The analysis results show that the signal power is significantly higher in the low-frequency range (0-100Hz) than in the high-frequency range (above 200Hz), and the signal amplitude is also larger in the low-frequency range. With increasing frequency, both power and amplitude decrease significantly. Based on this spectral characteristic, an appropriate filtering strategy was selected: a bandpass filter of 5Hz to 50Hz was applied to the landslide signal to remove high-frequency noise, while a low-pass filter with a cutoff frequency of 10Hz was used for the non-landslide signal to remove unwanted high-frequency components. This spectral analysis and filtering strategy provides a scientific basis for subsequent landslide identification and improves data quality and the performance of the computing system.

[0080] Figure 11 Confusion matrices are presented for different slice lengths (10 to 100 time steps). These matrices show the model's classification performance at different slice lengths, from 10 to 100 time steps. As the slice length increases, the model's ability to correctly classify landslides and non-landslides significantly improves. In particular, the confusion matrix at 100 time steps performs best, with fewer false negatives and false positives, indicating that its sensitivity and specificity are superior to shorter slices. This suggests that increasing the slice length helps the model capture more temporal information, thus achieving more accurate classification. The confusion matrices clearly demonstrate the trend of model performance increasing with slice length.

[0081] Figure 12 This figure summarizes the performance metrics across different slice lengths. It outlines all key performance indicators, including accuracy, recall, specificity, precision, F1 score, and AUC, at each slice length tested. The overall model performance gradually improves with increasing slice length, stabilizing at a time step of 100. At this slice length, the model achieves the highest accuracy (94.58%), recall (91.50%), and AUC (0.9458), demonstrating that a time step of 100 strikes the optimal balance between capturing sufficient temporal information and maintaining model performance. The performance improvement across all evaluation metrics is consistent, confirming that longer slice lengths enable the model to more effectively capture the subtle temporal dynamics of landslides.

[0082] Performance analysis showed that increasing the slice length from 10 to 100 time steps significantly improved the model's ability to identify landslides, as reflected in all metrics—accuracy, recall, specificity, precision, F1 score, and AUC. In particular, a slice length of 100 time steps became the optimal configuration for the model, providing the best overall performance across all metrics. This slice length enabled the model to effectively capture the temporal patterns of seismic waves associated with landslides, thereby improving detection sensitivity and specificity. However, while increasing the slice length to 100 time steps improved performance, the improvement became insignificant beyond 100 time steps. In fact, the performance improvement began to diminish when the slice length exceeded 100 time steps, indicating that the model's ability to capture relevant features had reached a bottleneck. Furthermore, longer slice lengths led to class imbalance in the dataset, i.e., a decrease in the number of positive samples (landslides) and an increase in the number of negative samples (non-landslides). This imbalance caused bias in model predictions, especially in terms of precision, as the model began to favor the majority class (non-landslides), exacerbating the false alarm rate. The importance of data interpolation becomes particularly evident when considering slices with more than 100 time steps. Extending the time step beyond 100 necessitates inserting additional data points, which can distort the original seismic waveform. This interpolation artificially alters the temporal characteristics of the seismic signal, potentially leading to inaccurate model predictions. Since the primary objective of this study is landslide detection based on real seismic data, introducing interpolated values ​​could compromise the integrity of the dataset and degrade model performance. Therefore, choosing 100 time steps as the slice length is optimal. It allows the model to capture sufficient spatiotemporal information for landslide identification while avoiding the risks of class imbalance and data distortion. A slice length of 100 time steps ensures optimal model performance while maintaining data authenticity, improving the accuracy and reliability of landslide detection.

[0083] Figure 11 The results show that as the slice length increases, the confusion matrix becomes more balanced, and misclassifications decrease. Figure 12 The overall performance metrics are summarized, showing that the model performs best at a time step of 100. Finally, Table 2 summarizes the performance metrics for each slice length in detail, highlighting the superiority of the 100-time-step configuration.

[0084] Table 2 Performance indicators for different slice lengths

[0085]

[0086] To evaluate reservoir performance, comparative experiments were conducted, analyzing the system's performance in both reservoir-containing and reservoir-free scenarios. Specifically, the model's performance was evaluated using six key metrics: accuracy, sensitivity, specificity, precision, F1 score, and AUC. Accuracy reflects the proportion of correctly classified samples; sensitivity (or recall) represents the percentage of actual positive samples accurately predicted as positive; specificity measures the proportion of actual negative samples correctly classified as negative; precision represents the proportion of actually positive samples among those predicted as positive; and the F1 score is the harmonic mean of precision and recall, reflecting the model's ability to accurately predict and recall positive samples simultaneously. AUC (Area Under the Curve) represents the area under the ROC curve, measuring the model's classification performance at different thresholds, ranging from [0,1] to 1, where values ​​close to 1 indicate that the model can more effectively distinguish between positive and negative samples.

[0087] Experimental comparison results are as follows Figure 3 As shown in Figure FI, purple represents the results without a data repository, and magenta represents the results with a data repository. Without a data repository, the model's test accuracy is 0.5792, sensitivity is 0.4383, specificity is 0.6673, precision is 0.6102, F1 score is 0.5102, and AUC is 0.5792. This indicates that the model performs poorly overall in the classification task, especially in identifying positive samples (sensitivity) and distinguishing negative samples (specificity). After introducing a data repository, the model's performance significantly improves: test accuracy increases to 0.9467, sensitivity to 0.9167, specificity to 0.9767, precision to 0.9752, F1 score to 0.9450, and AUC to 0.9467. These results demonstrate that the introduction of a data repository significantly improves the model's performance in positive class identification, negative class distinction, and overall classification ability, particularly in AUC and F1 score; the overall performance of the model is significantly improved by the data repository. These improvements, particularly in sensitivity and specificity, highlight the significant impact of the reservoir on model performance. These findings validate the practicality and efficiency of this method in landslide detection tasks, providing solid theoretical support for the future development of landslide monitoring and early warning systems.

[0088] The core concept of Reservoir Computation (RC) is to fix the weights of the reservoir and train only the weights of the output layer. This greatly simplifies the training process and avoids the gradient vanishing and gradient exploding problems common in RNNs. Figure 2As shown in Figure B, unlike the backpropagation algorithm in RNNs, RC processes time-series information through the dynamic response of the reservoir, simplifying the training process and avoiding the complex computation and training steps of traditional RNNs. This characteristic gives RC a significant advantage in processing landslide time-series signals. Due to its efficient real-time performance and strong generalization ability, RC can adapt to various landslide signal processing tasks, especially showing great potential in real-time early warning systems that require rapid response.

[0089] In this embodiment, the memristor used for reservoir calculations adopts a vertical Au / TiO2 / Pt structure, such as... Figure 4 As shown in Figure A, the fabrication process of this device includes three main steps: First, a 20-nanometer-thick platinum film is deposited on a SiO2 / Si substrate using magnetron sputtering to form the bottom electrode. Next, a switching layer is formed on a 70-nanometer-thick titanium dioxide film using pulsed DC sputtering. Finally, a 10-nanometer-thick gold film is deposited as the top electrode using magnetron sputtering. See flowchart below. Figure 5 The sputtering conditions for the platinum bottom electrode were: power 50W, argon flow rate 10 sccm, and sputtering pressure 2.6 × 10⁻⁶. -4 Torr. The sputtering conditions for the titanium dioxide layer were: power 100W, argon flow rate 20sccm, and sputtering pressure 5.8×10. -3 Torr. The sputtering conditions for the gold top electrode were: power 50W, argon flow rate 10 sccm, and sputtering pressure 1.3 × 10⁻⁶. -4 Torr.

[0090] The structure of the Au / TiO2 / Pt memristor was characterized using cross-sectional transmission electron microscopy (TEM), such as... Figure 4 As shown in Figure B, the cross-sectional TEM micrograph clearly shows the layered configuration of the Au and Pt electrodes with the titanium dioxide dielectric layer. Elemental mapping performed by energy-dispersive X-ray spectroscopy (EDS), such as... Figure 4 Figure C shows the spatially uniform distribution of the constituent elements (O, Pt, Ti, Au), thus confirming... Figure 4 The layered structure shown in D. To obtain accurate chemical state information, high-resolution X-ray photoelectron spectroscopy (XPS) was used to study the active layer of titanium dioxide in detail. The obtained XPS spectra ( Figure 4 EF) showed characteristic Ti2p 3 / 2 and Ti2p 1 / 2 The binding energies are 458.9 eV and 464.6 eV, respectively, accompanied by an O atom centered at 530.4 eV. 1s These peaks and spectral features indicate the presence of oxygen defect sites within the crystal framework. These spectral features are highly consistent with the Raman vibrational modes of similar oxide systems previously recorded.

[0091] In this embodiment, a comprehensive electrical characteristic analysis of the Au / TiO2 / Pt memristor was performed. Current-voltage (IV) characteristics were obtained through a voltage scan from 0V to ±3.0V, as shown below. Figure 4 As shown in G. In the negative voltage region, the device exhibits a significant nonlinear current response, with the current rising sharply as the voltage increases. This behavior indicates that the resistance of the memristor changes significantly under negative voltage, clearly switching between a high-resistivity state (HRS) and a low-resistivity state (LRS). This switching phenomenon is attributed to the migration of oxygen vacancies within the titanium dioxide layer under the influence of negative voltage, as previously reported in the literature. Specifically, under negative bias, oxygen vacancies migrate towards the positive electrode, forming a conductive path, thereby reducing the resistance of the device. When 20 consecutive scans of the same negative voltage amplitude (from 0V to -3.0V) are performed, the current gradually increases, as shown in G. Figure 4 As shown in H, the device resistance gradually decreases with increasing voltage scan count, further highlighting its significant nonlinear behavior. This also indicates a significant history-dependent effect, meaning the resistance state is influenced by previous voltage applications.

[0092] Next, the memristor response was tested under a fixed pulse voltage (3.0V amplitude, 200ms pulse width), as follows: Figure 4 As shown in Figure I, the current exhibits significant fluctuations within each pulse cycle, increasing rapidly when the pulse is applied and returning to its initial state after the pulse ends. The peak current response gradually increases with the number of pulses, further confirming the short-term memory effect of the device. These characteristics are a result of the memristor's dynamic properties, whose resistance is affected by internal charge redistribution and the formation or disappearance of conductive filaments within the material. This history-dependent response is crucial for the functionality of RC systems, enabling efficient handling of sequential data, time-series prediction, and pattern recognition tasks. This demonstrates that the memristor's resistive state is influenced not only by the currently applied voltage but also by its previous voltage history, highlighting a significant short-term memory effect. This dynamic memory behavior allows the memristor to mimic the function of a neural network with recurrent connections.

[0093] To further evaluate the potential application of Au / TiO2 / Pt memristors in reservoir RC calculations, this invention conducted detailed tests on the device's 4-bit dynamic response characteristics, such as... Figure 6 As shown in Figure A. To simulate a real-world RC application, the 4-bit pattern was encoded as a pulse stream consisting of 16 different combinations, covering all possible patterns from "0000" to "1111". In the experiment, each pulse cycle was treated as one bit, with a binary "1" represented by applying a high voltage (-6.5V) and a binary "0" represented by applying a low voltage (-0.1V). Figure 6D illustrates the device under pulsed input conditions for four different 4-bit modes, from "0000" to "1100". The device's response characteristics were evaluated by reading the memristor's current response (i.e., postsynaptic current, PSC) at -1V after each single-bit pulse stimulation. During the experiment, lower-amplitude pulse bits ("0") increased the time interval between adjacent higher-amplitude pulse bits ("1"), leading to charge relaxation within the device and a gradual decrease in the current response (PSC). Therefore, the current response became smaller in subsequent pulse stimulations. For example, the current response exhibited decay in the "1001" or "1011" modes. In contrast, for the "1111" mode, the pulse interval facilitated the gradual accumulation of oxygen vacancies in the Au / TiO2 / Pt memristor, thus gradually enhancing the current response. The formation of oxygen vacancies is closely related to the charge migration characteristics of the titanium dioxide material. Specifically, when an electric field is applied, charge carriers accumulate within the material or at the interface, forming space charge regions. Once the electric field disappears, the charge carriers relax and gradually return to a stable thermodynamic state. This process modulates the material's conductivity through charge migration and oxygen vacancy reconstruction. This response characteristic indicates that Au / TiO2 / Pt memristors can simulate synaptic plasticity and exhibit self-regulation capabilities, thus enabling efficient information processing with significant time characteristics. This characteristic is crucial for time information learning and decoding in RC circuits.

[0094] To further verify the stability and repeatability of the Au / TiO2 / Pt memristor in RC, the device was subjected to 20 consecutive operations in "0100" mode. Figure 9 ). A detailed statistical analysis was performed on the current response (PSC) data for each cycle. Figure 6 C). Experimental results show that the current response exhibits a clustered distribution across multiple cycles, with no significant anomalies observed between cycles. This indicates that the Au / TiO2 / Pt memristor demonstrates high cycle stability (C2C stability) in a specific 4-bit mode. These results further confirm the reliability and repeatability of the memristor as an RC physical storage device, demonstrating its ability to maintain stable performance over long periods of operation. This provides important support for the potential applications of memristors in neuromorphic computing and non-volatile memory.

[0095] Next, the current response of Au / TiO2 / Pt memristors in all 4-bit modes from "0000" to "1111" under pulse stimulation is systematically summarized, and statistical and individual analyses are conducted to... Figure 6 B and Figure 6 These responses are detailed in section E; see full data here. Figure 10By applying 16 different input pulse currents, significant differences in the current response were observed in each input mode. This demonstrates the device's excellent ability to separate sequential time information and identify patterns. These results prove that the device can effectively distinguish input modes and plays a crucial role in ensuring the accuracy and reliability of information processing in RC systems. In 4-bit mode, the output current state is affected not only by the current pulse but also by previous input signals. The current responses generated by different pulse combinations exhibit significant differences, indicating that the memristor can effectively distinguish multiple input modes and provide a high-precision dynamic response for information processing. This characteristic is crucial for RC systems, especially in tasks such as learning, storing, and decoding time information, ensuring efficient and reliable data processing. The differentiated responses of the device in different modes with changing input pulse patterns not only verify its high sensitivity to time-series information but also provide strong experimental evidence for its use as a physical reservoir in RC systems. These experiments further confirm the device's potential in neural network simulation and dynamic information processing, particularly in achieving efficient and accurate time information decoding. This device has broad application prospects in the RC field. Through this series of experiments, the potential of Au / TiO2 / Pt memristors as physical reservoirs has been verified, demonstrating their ability to effectively process and simulate time series data, and showcasing their broad application prospects in the RC field.

[0096] Figure 9 The effect of 20 consecutive 0100 pulses on the memristor response current was demonstrated. To further verify the stability and repeatability of the Au / TiO2 / Pt memristor in reservoir calculations, it was run continuously for 20 cycles in "0100" mode. Experimental results show that the current response remains concentrated within each cycle, with no obvious outliers. This indicates that the Au / TiO2 / Pt memristor exhibits high cycle stability (i.e., C2C stability) in a specific 4-bit mode.

[0097] Figure 10 The current response of the Au / TiO2 / Pt memristor to different 4-bit pulse stimulation modes is demonstrated. The current response of the memristor in all 4-bit modes from "0000" to "1111" is systematically summarized. Significant differences in the current response were observed for each input mode when 16 different input pulse currents were applied. This phenomenon indicates that the device performs excellently in separating time information and identifying modes, effectively distinguishing different input modes. This characteristic is crucial for ensuring the accuracy and reliability of information processing in reservoir calculations.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors, characterized in that: It is equipped with a data acquisition module, a preprocessing module and a dynamic memristor network connected in sequence; The dynamic memristor network is provided with an input layer, a storage layer and an output layer connected in sequence. The input layer has three input units arranged in parallel, and the storage layer has three storage units arranged in parallel. The output terminals of the three input units are connected to the input terminals of the three storage units in a one-to-one correspondence. The output terminals of the three storage units are all connected to the input terminals of the output layer. The data acquisition module is used to acquire raw seismic wave signals with time series characteristics, extract seismic wave time series signals in three directions from the raw seismic wave signals, and then transmit them to the preprocessing module. The preprocessing module is used to synchronously preprocess the seismic wave time series signals in three directions to generate seismic wave signals in three channels and transmit them to the dynamic memristor network. Each input unit in the dynamic memristor network is used to acquire the seismic wave signal of one channel respectively; then the seismic wave signal is processed by the corresponding storage unit to obtain the memristor response current signal of each channel; finally, the output layer integrates all the memristor response current signals of the three channels to obtain the landslide signal identification result.

2. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 1, characterized in that: The data acquisition module extracts seismic wave time series signals in three directions: east-west (BHE) seismic wave time series signal, north-south (BHN) seismic wave time series signal, and vertical (BHZ) seismic wave time series signal.

3. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 1, characterized in that: The preprocessing module is provided with a mean removal unit, a fast Fourier transform unit, a power spectral density analysis unit, a filtering unit, and a signal segmentation unit connected in sequence. The mean-reduction unit is used to subtract the average value of each signal in the seismic wave time series signal in each direction, so that the signal fluctuates near zero. The Fast Fourier Transform (FFT) unit is used to perform a Fast Fourier Transform on the signal processed by the Mean Removal Unit, converting the signal from the time domain to the frequency domain. The power spectral density analysis unit is used to identify the frequency range distribution of the signal after the fast Fourier transform, and to identify landslide-related signals and non-landslide signals in each direction of the signal. The filtering unit is used to filter the landslide-related signals and non-landslide signals in each direction signal respectively, and then sample the filtered signals. The signal segmentation unit is used to divide the sampled landslide-related signals and non-landslide signals into at least two time slices, each time slice containing signals of m time steps, and then divide all time slices into three channels of seismic wave signals according to the signal direction.

4. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 3, characterized in that: The filtering unit uses a bandpass filter with a frequency range of 5Hz to 50Hz to filter the landslide-related signals; the filtering unit uses a filter with a cutoff frequency of 10Hz to filter the non-landslide signals; the filtering unit samples the filtered signals at a sampling rate of 1000Hz.

5. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 3, characterized in that: Each input unit in the input layer is used to normalize the time slice of the corresponding channel, and then to match the signal in the time slice with the input voltage range of the reservoir memristor model through linear transformation, and then transmit the linearly transformed seismic wave signal to the corresponding storage unit.

6. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 1, characterized in that: Each of the aforementioned memory cells is composed of a single memristor, which is an Au / TiO2 / Pt dynamic memristor.

7. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 6, characterized in that: The storage unit utilizes the nonlinearity and short-term memory effect of the Au / TiO2 / Pt dynamic memristor to process the seismic wave signal from the input layer using the memristor response. At each time step, the reservoir unit responds to the input signal by continuously updating the conductance and iteratively updates the memristor state at each time step based on the conductance response to obtain the memristor response current signal of the corresponding channel. After the storage unit completes the iterative processing of m time step signals in a time slice, it obtains the memristor response current signals corresponding to the m time step signals. Then, it flattens all the memristor response current signals in the time slice into a 1×m one-dimensional feature vector and outputs it to the output layer.

8. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 7, characterized in that: The expressions for the conductivity response and current response in the reservoir unit are as follows: Where G represents the conductance of the memristor at the current time step; G ′ This represents the conductance of the memristor at the previous time step; α, β, r, and K represent the fitting parameters. When V > 0, K takes the value of the positive half-axis curve fitting parameter K. p When V < 0, K takes the value of the negative half-axis curve fitting parameter K. n V represents the input voltage, i.e., the seismic wave voltage signal input at the current time step; G0 represents the initial conductance value, when V>0, G0=1; when V<0, G0=0; G th denoted by the conductance threshold; I represents the response current value at the current time step; e represents the base of the natural logarithm function.

9. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 7, characterized in that: The output layer is a multilayer perceptron classifier (MLP), which integrates 1×m one-dimensional feature vectors from three directions of the reservoir to obtain a 3×m two-dimensional feature matrix, and then outputs the landslide signal identification result through the sigmoid activation function.

10. The landslide signal identification system based on dynamic reservoir calculation using titanium dioxide memristors according to claim 1 or 9, characterized in that: The value range of the landslide signal identification result is [0,1]. When the value of the landslide signal identification result is greater than the landslide probability threshold, it indicates that a landslide has occurred; otherwise, it indicates that no landslide has occurred.

Images