U-TSS: a novel time series segmentation model based U-net applied to automatic detection of interference events in geomagnetic field data

Introduction

As innovations in Internet of Things (IoT) technology advance, the scope and complexity of sensor data acquisition have grown, establishing it as a crucial aspect of big data technologies (Yin et al., 2020). An increasing number of devices and sensors can collect and transmit real-time data, typically in the form of time series (Silva et al., 2021). Time series data is widely applied across various fields, including finance, ecology, economics, neuroscience, and physics (Matias et al., 2021; Polge, Robert & Le Traon, 2020). In the era of big data, the growing volume of data has revealed significant limitations of traditional time series analysis methods in handling complex patterns and long-term dependencies. The advancement of deep learning technologies provides new solutions for time series analysis. By employing convolutional neural network (CNN) and recurrent neural network (RNN), deep learning models can automatically extract latent features from time series data and capture complex temporal patterns. A growing body of research suggests that deep learning methods consistently outperform traditional approaches in time series forecasting, time series segmentation (TSS), time series classification and anomaly detection tasks.

Time series segmentation involves partitioning data into non-overlapping, automatically labeled segments. The primary objective of time series segmentation is to identify and delineate change points or event boundaries within the time series. This facilitates the organization of similar or related data segments while isolating dissimilar segments. To enable deep learning models to perform effective time series segmentation, it is essential to first label the data. Much like image segmentation tasks, where pixel-wise annotations aid in learning, properly labeled data plays a crucial role in training deep learning models for time series segmentation. The methods for labeling time series segmentation can be primarily categorized into two approaches: sliding window labeling and dense labeling (Gaugel & Reichert, 2023). Figure 1 illustrates the distinction between sliding window labeling and dense labeling. Sliding window labeling involves dividing the time series data into several fixed-length subsequences, each of which is assigned a label. Although sliding window labeling has demonstrated satisfactory results in many applications, its accuracy is inherently constrained by the size of the time window and the step size. These limitations make it difficult to capture precise event boundaries, particularly for events with variable lengths, leading to coarse segmentation and reduced interpretability in datasets with complex temporal dynamics. In contrast, the dense labeling method provides a precise approach to time series segmentation that does not rely on sliding windows. By assigning labels to each time step, dense labeling provides detailed classification information, enabling it to better adapt to diverse and irregular temporal patterns and accurately capture event boundaries.

Similar to semantic segmentation in computer vision, effective time series segmentation depends on dense labeling and robust model architectures to ensure precise classification. In dense labeling, each individual data point, whether part of a time series or an image pixel, is assigned a specific category. A pivotal architecture in semantic segmentation is the fully convolutional network (FCN) (Long, Shelhamer & Darrell, 2015), which has significantly contributed to end-to-end pixel-level predictions. Based on FCN, many advanced convolutional neural networks have been proposed, including U-net (Ronneberger, Fischer & Brox, 2015), which has been widely used in various segmentation tasks. U-net’s symmetric contracting and expanding paths form a U-shape, and the network uses skip connections to combine positional and semantic information. Segnet (Badrinarayanan, Kendall & Cipolla, 2017) is similar to the U-net network, but it uses indexing for up-sampling to better preserve the boundary feature information. Deeplab (Chen et al., 2017) used dilated convolution and fully connected conditional random fields to improve the segmentation accuracy for image segmentation. Recent research has shown that all of these methods have been quite successful in the field of image segmentation.

Despite the advancements in segmentation techniques for images data, the inherent complexity and unique characteristics of time series data in other fields necessitate specialized segmentation methods. For instance, sleep staging and human activity recognition (HAR) are two major areas of study within time series segmentation, offering a wealth of experimental results and research insights (Yu et al., 2019). Phan et al. (2019) introduced SeqSleepNet, a hierarchical RNN designed for sleep staging, which enables end-to-end training of the network and classifies each time step of the time series to generate an output label sequence. Supratak et al. (2017) developed DeepSleepNet, a model that employs CNN to extract temporal features and utilizes bidirectional long short-term memory networks (BiLSTM) to automatically learn the transition rules between sleep stages from electroencephalogram signals. Perslev et al. (2019) proposed U-Time, a fully feed-forward deep learning algorithm for studying physiological time series segmentation of sleep data. U-Time classifies each time point in the input signal and aggregates them at fixed intervals to produce a final prediction. U-Sleep (Perslev et al., 2021) is an extension of U-Time designed for physiological time series segmentation applications. U-Sleep enables the marking of sleep stages at shorter intervals and facilitates automatic sleep staging. Zhang et al. (2024) developed a deep learning method based on 1D-CNN for human activity recognition, evaluating the model using three public datasets, all of which yielded satisfactory results. Dua, Singh & Semwal (2021) employed an end-to-end model for automatic feature extraction and activity classification. This model, which combines CNN and gated recurrent units (GRU), demonstrated robust classification performance across three publicly available human activity recognition datasets.

Similar to these physiological signal processing tasks, the detection of interference events in geomagnetic field observation data also requires precise time series segmentation methods. These geomagnetic observation data are primarily collected by geomagnetic observation instruments. Geomagnetic observation instruments can be considered as a type of IoT-like device. Compared with typical IoT devices, these instruments are larger in size, but they also have network functions, enabling them to upload the collected data to a server for further analysis. These instruments record the different components of the local magnetic field, where D represents magnetic declination, H the horizontal component, and Z the vertical component of the magnetic field. Geomagnetic observation data plays a crucial role in the study of geomagnetic storms and seismic-magnetic relationships (Beggan et al., 2024; Divett et al., 2017, 2020, 2018). However, as modern infrastructure expands, various sources of interference, including highways, subways, and high-voltage direct current (HVDC) transmission lines, have introduced significant undesired noise into geomagnetic data collection. If these interference events are not detected and preprocessed, geomagnetic field observation data cannot be applied seismic-magnetic relationship research. Among various interference events, HVDC interference events have a wide range of impacts and high frequency. Figure 2 illustrates the distribution of geomagnetic stations across the country and their impact from HVDC interferences, with stations affected by HVDC interference highlighted in red and those unaffected marked in blue. The map clearly shows that the majority of stations in China are influenced by HVDC interference events. These HVDC interference events are not related to geomagnetic activity but are primarily caused by the operation of HVDC transmission systems. During routine operation, HVDC transmission systems induce unbalanced currents in the transmission lines, generating additional magnetic fields that superimpose on the natural geomagnetic field. These unbalanced currents often arise from system faults, testing procedures, or experimental activities. The closer the observation station is to the HVDC transmission line, the more pronounced the interference. The HVDC interference events manifests as staircase-like patterns in the Z-component of the geomagnetic field, with varying durations, amplitudes, and orientations, as shown in Fig. 3. The variability and unpredictability of this interference complicate the task of accurately detecting and analyzing interference events. HVDC interference events have become a focal point in the preprocessing of geomagnetic field observation data.

Current detection methods for HVDC interference events in geomagnetic observation data mainly rely on manual analysis and expert evaluation, typically involving the following steps. When technicians observe anomalies in the data, they first evaluate the abrupt changes in data trends, using their expertise and professional knowledge to determine whether the anomaly is a magnetic interference. Once magnetic interference are ruled out, they examine the observation system, operational records, and environmental conditions to identify other possible sources of interference. After excluding these factors, the focus shifts to analyzing changes in the Z-component. If the variation in the Z-component is significantly larger than that in the D and H components, the event is initially judged to be a HVDC interference event. Additionally, if multiple geomagnetic instruments affected by a line show the same disturbance pattern in the same time period, it is classified as HVDC interference event. However, manual detection methods often vary between individuals. To minimize such variability, identified interference events are reviewed and validated by experienced experts. The manual approach is time-consuming and labor-intensive, making it challenging to meet the demands of intensive data processing in geomagnetic observation networks and highlighting the need for an accurate and widely applicable method for detecting HVDC interference events.

In contrast to sleep staging and human activity recognition, the primary challenge in geomagnetic field observation data preprocessing is the precise detection of interference events. To address these challenges, this study draws inspiration from advancements in time series segmentation and semantic segmentation and proposes a sequence-to-sequence time series segmentation model based on U-net. We apply this model to the detection of HVDC interference events in geomagnetic field observation data, enabling the segmentation of the entire time series and the identification of HVDC interference events of varying durations at any temporal scale. The key contributions of this study are as follows:

• We proposed U-TSS, a novel time series segmentation model based on U-net. This model combines time series segmentation techniques with semantic segmentation to achieve high-precision dense segmentation of interference events in geomagnetic field observation data. This innovative integration effectively leverages the advantages of the U-net architecture, thereby enhancing the ability to capture complex temporal features and patterns.

• The model employs a sequence-to-sequence framework, utilizing one day’s geomagnetic field observation data as input. It implicitly classifies each individual time point in the input data and aggregates these classifications into time series segments of varying lengths to produce the final predictions. This design enables the model to accurately classify each time step, significantly improving the accuracy and robustness of the dense segmentation.

• By employing the Particle Swarm Optimization (PSO) algorithm (Kennedy & Eberhart, 1995) to optimize the model’s hyperparameters, the model’s performance is further enhanced. The introduction of the PSO algorithm not only accelerates the model’s convergence speed but also optimizes the selection of hyperparameters, allowing U-TSS to demonstrate higher efficiency and accuracy when processing geomagnetic field observation data.

In the remainder of the article, we outline the overall structure. “Problem Formalization” formalizes the problem. “Method” provides an overview of the U-TSS architecture and its modules. “Dataset” details the data sources and the sample production process. “Experiments” presents a series of tests conducted with state-of-the-art time series segmentation models. “Results” describes the results of the experiment. “Discussion” makes a discussion, and “Conclusions” concludes this essay.

Problem formalization

U-TSS is a sequence-to-sequence fully convolutional network designed specifically for one-dimensional time series segmentation tasks. Built upon the U-Net architecture, this model effectively addresses the limitations of traditional segmentation models when handling complex time series data. U-TSS processes time series data of arbitrary length as input, efficiently mapping the complete sequence to dense outputs in a single forward pass using fully convolutional layers. This design enhances the model’s capacity to capture global patterns in long sequences while ensuring precise segmentation at each time step, enabling accurate classification of complex temporal signals.

The problem of time series segmentation can be formally defined as follows: Let $X_C\in R^{T\times C}=\left[X_1,X_2,\cdots ,X_C\right]$ represent the multivariate time series data for a specific time interval, where T denotes the number of time steps and C represents the different features of the time series data. In this study, we focus on a univariate case, thus $C=1$. Consequently, the univariate time series data can be expressed as a one-dimensional array $X\in R^{T\times 1}=\left[x_1,x_2,\cdots ,x_T\right]$. To address the segmentation problem, each time step $t$ is associated with a label $Y_t\in R^{T\times 1}$, which indicates the category to which the time step belongs. This labeling approach constitutes a dense annotation method, which is essential for the segmentation task, as it enables the precise categorization of each time point according to its respective class.

In the context of geomagnetic field observation data, the input data $X$ for the U-TSS model consists of one-dimensional time series data with the shape $R^{T\times 1}$, corresponding to the measurements of a specific component across $T$ time steps. The primary task is to map this input data to a sequence of predicted labels ${\hat{Y}}_t\in R^{T\times 1}$, producing an output for each time step. The objective is to accurately identify and localize HVDC interference events in the geomagnetic observation data, resulting in a label sequence where HVDC events are marked as 1 and the BACKGROUND is marked as 0.

This flexibility in assigning labels to every time point is a key feature of U-TSS, allowing it to handle fine-grained temporal segmentation tasks effectively. U-TSS implicitly classifies each time point in the input data and aggregates these classifications into time series segments of varying lengths. This approach enables the accurate detection of HVDC interference events in geomagnetic field observation data, thereby enhancing the accuracy and efficiency of time series segmentation methods.

Method

U-TSS model overview

To achieve precise detection of HVDC interference events in geomagnetic field observation data, U-TSS is adapted from the U-Net model, originally developed for biomedical image segmentation. The U-Net model is structured with both a contracting path and an expanding path, enabling effective feature extraction and reconstruction. The contracting path consists of four contraction stages. Each contraction stage includes two 3 × 3 convolutional layers and one 2 × 2 max pooling layer. After each contraction stage, the number of feature maps doubles while the feature maps are halved in size. The expanding path is located on the right side of the network structure and is made up of four stages. The expanding path consists of four extension stages. Each expansion stage includes one 2 × 2 up-convolutional layer and two 3 × 3 convolutional layers. The number of feature maps is reduced by half through deconvolution, and then connected to the symmetric feature maps from the contracting path on the left side. Because of the difference in size between the contracting and expanding path feature maps, the U-Net model crops the contracting path feature maps to match the size of the symmetric feature maps on the right side. The final output is obtained by applying a 1 × 1 convolutional layer to the entire model.

Figure 4 illustrates the network structure of the U-TSS model, comprising three primary modules: the contracting path, the expanding path, and a dense segmentation classifier. The contracting path focuses on feature extraction by progressively down-sampling the input time series. It captures relevant temporal patterns through convolutional operations and reduces the temporal resolution via pooling layers. The expanding path is designed to up-sampling the feature maps obtained from the contracting path, restoring the original resolution of the input. By incorporating skip connections, the model combines coarse and high-level features with fine details, which enhances its capacity for accurate time series segmentation. The dense segmentation classifier assigns a label to each time step of the input sequence. It leverages the outputs from the expanding path to produce a dense output, generating class probabilities for every time step in the time series. The design of these three modules enables the U-TSS model to effectively achieve precise segmentation and classification of time series.

The U-TSS model modifies the conventional U-Net architecture, which was originally developed for two-dimensional image data, to effectively accommodate one-dimensional time series data. Specifically, the U-TSS employs one-dimensional convolutions to effectively extract features related to HVDC interference events in geomagnetic field observation data. In contrast to two-dimensional convolution, where a sliding window operates over the feature map in both width and height directions, the one-dimensional convolution focuses solely on the width direction. This approach allows for efficient processing of the time series data, as it multiplies and sums values at corresponding positions within a single dimension. The mechanics of one-dimensional convolution are illustrated in Fig. 5. To ensure that the input feature maps of the convolution process are consistent with the output feature maps and to avoid the cropping operation during feature fusion, the U-TSS model uses SAME convolution. The convolution kernel size in Fig. 5 is three and it moves across the input sequence in fixed steps. At each step, the convolution kernel computes the output value by multiplying and summing the elements corresponding to its current position in the input region. Afterwards, the convolution kernel shifts by one step, repeating this operation until the entire input sequence is traversed. Each convolution calculation produces a point in the output feature map, resulting in a new feature map after the convolution process. Assuming that n is the current convolutional layer, the one-dimensional convolutional operation formula for this layer is shown in Eq. (1):

(1) $$Q_j^n=\sum _{i=1}^K{Q}_{j+i-1}^{n-1}\times W_{ij}^n+b_j^n$$where $Q_j^n$ represents the output at the position $j$ after the $n$-th convolutional layer, $K$ is the size of the convolutional kernel, $i$ is the convolution kernel index, and $j$ represents the position index in the output feature map of the $n$-th layer. $\times$ denotes the convolutional operation, $W_{ij}^n$ represents the weight of the convolution kernel at the $i$-th position for the $j$-th output in the $n$-th layer, and $b_j^n$ is the bias corresponding to the output features at position $j$ in layer $n$.

Following the convolution operation, each output value is often transformed nonlinearly by using an activation function in order to improve the network model’s non-linear features. The activation function used in U-TSS is ReLu, which can be represented as Eq. (2):

(2) $$y_j^n=f\left(Q_j^n\right)=max\left\{0,Q_j^n\right\}=\{\begin{array}{c}{Q}_j^n,Q_j^n\ge 0\\ 0,Q_j^n<0\end{array}$$where $Q_j^n$ represents the input value at location $j$ to the activation function from the convolution operation, $y_j^n$ denotes the output value at location $j$ after the activation function is applied in the $n$-th layer.

Dataset

HVDC sample production

The detection of HVDC interference events has primarily relied on manual inspection methods, which are often characterized by substantial time requirements and a susceptibility to human error. Furthermore, the absence of a standardized dataset for these events presents significant challenges for the advancement and validation of automated detection algorithms. Consequently, we produced a dataset specifically for the detection of HVDC interference events in geomagnetic field observation data. In this subsection, we describe the process of sample production in detail. Figure 6 illustrates the sample production process.

Firstly, we selected all HVDC interference events from January 1, 2014, to December 31, 2018, from the manual preprocessing log and recorded the date, start time, end time, and affected instrument of each HVDC event. For each instrument, we individually aggregated all HVDC interference events observed on each specific day into separate records. This step ensures that each instrument’s daily records are consolidated, which is necessary for generating the label file. Missing values in the data were handled using linear interpolation.

Then, the observation data was standardized using the Z-score. In the data standardization process, the mean ( $\mu$) and standard deviation ( $\sigma$) are computed separately for each instrument at each station on a per-day basis. The data for each instrument is standardized using the following equation:

(3) $$Z_i={\displaystyle \frac{X_i-\mu }{\sigma }}$$where $X_i$ is the raw observation at time $i$, $\mu$ is the mean of the observation data for the corresponding instrument at that station on the given day, and $\sigma$ is the standard deviation of the same instrument’s observations on that day. This standardization process does not differentiate between geomagnetically active and inactive days, and its primary aim is to normalize the data, reducing the impact of geomagnetic activity fluctuations. The normalized data is saved as a data file for the HVDC sample, named ‘station_code-instrument_code–date .npy’.

Finally, similar to the sample generation method in semantic segmentation, the U-TSS model needs to label each time point value in the geomagnetic field observation time series sample as BACKGROUND or HVDC. If a time point value is an HVDC interference event, it is labeled as 1, otherwise, it is labeled as 0. After the above operation, we get a label file with the same name as the data file of the HVDC sample, which is stored in the label folder. Therefore, each HVDC sample contains one data file and one label file, and has a consistent length.

Figure 7 shows an HVDC sample from station 12005, instrument 1, on July 3, 2017. The top half of the figure shows geomagnetic field observation data after normalization. The black line represents the BACKGROUND, and the red line indicates the HVDC interference events that occurred on that day. The bottom half of the figure shows the labels corresponding to the data, where 1 indicates HVDC, and 0 indicates the BACKGROUND.

The dataset comprises a total of 9,255 samples, collected from 126 affected observation stations. To enhance the model’s generalization capabilities, the samples were randomly shuffled and divided into three subsets: 7,405 samples for training, 925 for validation, and 925 for testing, following an 8:1:1 ratio. The number of samples per station varies, depending on the frequency and occurrence of HVDC interference events throughout the study period.

Experiments

Results

The accuracy and loss curves of the U-TSS model are shown in Fig. 8. During the training process, the accuracy on both the training and validation sets gradually increased, eventually reaching 99.42% and 94.61%, respectively. The U-TSS model achieved 95.54% accuracy, 88.65% precision, 76.07% recall, and an F1-score of 0.8188 on the testing set.

To evaluate the performance of the U-HVDC, we selected several typical detection methods for comparison. Among these, we incorporated the threshold-based first-order difference (1st-difference) method, a widely used auxiliary approach for the preliminary identification of HVDC interference events. This method relies on experts to manually define thresholds, which are subsequently verified and corrected. Additionally, we also selected some other typical time series segmentation methods to further compare the performance of U-TSS. Although there is limited literature specifically applying deep learning algorithms to detect HVDC interference events, we drew on the similarities between geomagnetic field observation data and time series data in biomedical fields, such as electrocardiogram (ECG) and electroencephalogram (EEG). Therefore, we compared U-TSS with several established segmentation algorithms commonly used in these fields, including the CNN (Acharya et al., 2017), CNN-LSTM (Oh et al., 2018), TinySleepNet (Supratak & Guo, 2020), and U-Time models. In addition, considering that geomagnetic field observation data are classical time series data, we also included several traditional time series classification methods in our comparison. These methods encompass the Encoder model (Serra, Pascual & Karatzoglou, 2018), FCN (Wang, Yan & Oates, 2017), ResNet, and Inception model (Ismail Fawaz et al., 2020). All models were trained and tested on the same datasets to assess their generalization ability.

In the comparative experiments, all models (excepted TinySleepNet and 1st-difference) were evaluated using the optimized hyperparameters obtained from the experiments. Consistent with the training approach of the U-TSS model, an early stopping mechanism was implemented with a maximum of 300 iterations and a patience of 20. Since the 1st-difference method heavily relies on threshold setting, to ensure fairness, we calculated the evaluation indicators for detecting HVDC interference events by setting different thresholds and selected the best result as the detection outcome of this method. Table 2 shows the experimental results of the 1st-difference method. A threshold of 2.0 was chosen, yielding the highest F1-score, establishing it as the optimal outcome for the 1st-difference method.

Table 3 summarizes the comparative experimental results on the test set. The U-TSS model achieves an accuracy of 95.54%, a precision of 88.65%, a recall of 76.07%, an F1-score of 0.8188, and an AUC of 0.9637. In comparison, other methods show lower accuracy: the CNN model achieves 90.24%, the CNN-LSTM achieves 90.82%, and traditional methods like the 1st-difference approach achieve only 73.04%. Several other models, including TinySleepNet, U-Time, Encoder, FCN, Inception, and ResNet, report accuracies in the range of 88.96% to 90.66%, as detailed in Table 3. To provide a more intuitive comparison, the results have also been visualized in Fig. 9. In the bar chart, pink bars represent Accuracy, gray bars denote Recall, green bars indicate Precision, and blue bars correspond to the F1-Score. The results from the experiment indicate that the proposed U-TSS model achieves the highest scores across all evaluation metrics, demonstrating superior overall performance. It surpasses both traditional statistical methods and state-of-the-art deep learning approaches in detecting HVDC interference events.

Figure 10 presents the detection results of various models for an HVDC interference event in geomagnetic observation data collected on May 7, 2023, from station code 14014 and instrument code 1. The U-TSS model demonstrates superior performance, accurately detecting the interference event and precisely identifying its onset and termination. In contrast, the CNN model, while capable of detecting the event, shows less precision in delineating the event’s boundaries. The CNN_LSTM model exhibits similar detection capability to the U-TSS model but suffers from occasional false positives, as evident from the figure. Both U-Time and FCN models successfully identify the start and end times of the HVDC interference event but exhibit reduced accuracy during the event’s duration, with a noticeable number of false detections. Meanwhile, the Encoder, Inception, and ResNet models, although detecting the interference event and its boundaries, display higher false positive rates by misclassifying background data as HVDC interference. Overall, the U-TSS model outperforms all other models in terms of accuracy and reliability for detecting HVDC interference events.

Discussion

Even in some complex cases, the U-TSS model still shows excellent detection performance. Figure 11 shows an example of HVDC interference events detection. The X-axis represents the corresponding time of the observation data on that day. The blue curve is the background data after the Z-score standardization, and the red curve is the HVDC events detected manually after the Z-score standardization. The orange and purple areas represent the probability that the U-TSS model predicts as the BACKGROUND and HVDC. Although the duration, direction, amplitude, and shape of each HVDC interference event are different, the U-TSS model can accurately detect all HVDC interference events, and accurately locate the start and end time of each HVDC interference event.

To validate the detection ability of the U-TSS model in actual HVDC interference events, we randomly selected two days of unused geomagnetic field observation data for experimentation. These days are from instrument 1 at station 14014 on May 6, 2023, and instrument 1 at station 42009 on May 28, 2023. We used the trained U-TSS model to detect HVDC interference events.

As shown in Fig. 12, there is only one HVDC interference event marked by the red curve. The U-TSS model successfully detected this HVDC event and accurately identified the start time and end time of the HVDC event.

Figure 13 shows the detection results of the U-TSS model for the data from instrument 1 at station 42009 on May 28, 2023. There are four HVDC interference events on that day, indicated by the red curves and labeled as 1, 2, 3, and 4. The U-TSS model successfully detected events 1 and 3, while events 2 and 4 were not detected correctly. Currently, there is a certain level of false positive rate, and further improvements are needed to enhance the model’s accuracy.

Compared to the existing manual detection technologies for HVDC interference events, the U-TSS model, as a time series segmentation method, presents several significant advantages:

• This model can significantly reduce the reliance on expert experience and minimize the need for manual intervention, thereby greatly decreasing labor costs. While the model achieves automatic detection of HVDC interference events with high efficiency, occasional human review may still be required to address missed or ambiguous cases, ensuring the accuracy and reliability of the detection results.

• It supports the detection of HVDC interference events of different durations, varying amplitude levels, and different directions.

• It can accurately locate the start time and end time of HVDC interference events.

• It exhibits high detection accuracy and strong generalization capability, making it suitable for all stations without requiring separate training or optimization for each station.

Of course, the U-TSS still has some room for improvement, including:

• The recall of the U-TSS model still needs to be further improved, therefore, the network structure of the U-TSS model can be further optimized, such as introducing LSTM or attention mechanism.

• In practical applications, subway, light rail, and HVDC interference events may occur at the same time, which may reduce the accuracy of the U-TSS model.

Conclusions

With the proliferation of IoT devices and an increasing reliance on sensor networks for real-time monitoring, the necessity for efficient processing of time series data has become paramount. As the deployment of geomagnetic field observation instruments expands and HVDC transmission lines grow, the cost and complexity associated with the manual detection of HVDC interference events are rising. To address these challenges, this article introduces U-TSS, a novel time series segmentation model based on the U-net architecture, applied to the automatic detection of HVDC interference events in geomagnetic field observation data.

U-TSS employs a fully convolutional sequence-to-sequence architecture to perform dense segmentation on one-dimensional time series data, ensuring precise labeling of each time step, and addressing the challenges posed by complex and lengthy temporal dependencies in IoT-driven applications. By implicitly classifying each time point and aggregating these classifications across varying intervals, U-TSS effectively detects events of different durations, enabling fine-grained segmentation even in the presence of complex temporal patterns. Additionally, the PSO algorithm was employed to optimize U-TSS’s hyperparameters, further enhancing its performance. Experiments demonstrate that U-TSS outperforms state-of-the-art models with accuracy, precision, recall, F1-score, and AUC values of 95.54%, 88.65%, 76.07%, 0.8188, and 0.9637, respectively, on the test set.

The key contribution of U-TSS is its ability to accurately segment and detect HVDC interference events in geomagnetic field observation data. This not only significantly reduces the labor cost and time involved in manual detection but also provides an efficient and scalable solution for handling vast amounts of time series data generated by IoT systems. In the future, we plan to extend its application to detect other types of interference events, such as those caused by subways and vehicles.

Supplemental Information