Abnormality detection method for propulsion device for railway vehicles
Patent Information
- Application Number
- KR1020250047795
- Authority / Receiving Office
- KR · KR
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2026-10-21
Smart Images
Figure PAT00002_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a method for detecting abnormalities in a propulsion system for railway vehicles, and more specifically, to a method for detecting abnormalities in a propulsion system for railway vehicles that collects and preprocesses temperature and current data of the propulsion system and compares them with a threshold value that distinguishes between normal and abnormal conditions to detect whether the system is normal or abnormal. Background Technology
[0002] Generally, railway vehicles are equipped with a traction motor, or propulsion device, installed to receive voltage supplied from a main generator or overhead line to tow (drive) the vehicle.
[0003] The above-mentioned traction motors for railway vehicles are classified into indirect drive type traction motors that drive the drive shaft of the wheel through a power transmission device, and direct drive type traction motors that drive the drive shaft directly by being coupled to the drive shaft of the wheel.
[0004] These traction motors for railway vehicles adopt an air-cooling method that allows them to cool during operation to minimize heat generation caused by continuous use.
[0005] Furthermore, various components constituting railway vehicles, such as traction motors, primarily utilize a preventive maintenance method of replacing them before they fail in order to maintain safety and efficiency.
[0006] This method is a measure to prevent sudden breakdowns of the vehicle and minimize downtime, and involves replacing or inspecting parts at set intervals.
[0007] However, conventional propulsion systems for railway vehicles had the problem of incurring excessive replacement costs because they relied on a preventive maintenance method, that is, replacing parts in advance before a failure occurred.
[0008] In addition, conventional railway vehicle propulsion systems have a problem in that there are significant constraints on developing anomaly detection models because sufficient data regarding failures is not secured, as there is little to no data regarding actual failure situations.
[0009] In addition, while conventional railway vehicle propulsion systems can apply supervised learning-based anomaly detection methods, they suffer from a problem of reduced accuracy regarding data that deviates from previously classified faults. Prior art literature
[0010] Registration No. 10-2110480 Publication No. 10-2023-0163614 The problem to be solved
[0011] Accordingly, the present invention has been devised to solve the aforementioned problems, and the objective of the present invention is to provide a method for detecting anomalies in a railway vehicle propulsion system, wherein a traction motor, which is a propulsion device, learns the pattern of normal data by utilizing an unsupervised learning-based deep learning model, and automatically detects an 'anomaly' when new data differs from the normal pattern based on this, thereby enabling early detection of unexpected abnormal situations or potential failures. means of solving the problem
[0012] To achieve the above objective, the present invention provides a method for detecting abnormalities in a propulsion system for a railway vehicle, comprising: (a) collecting temperature and current data of the railway vehicle propulsion system; (b) preprocessing raw data regarding the temperature and current of the propulsion system; (c) performing learning using a reconstruction error from the preprocessed data; (d) calculating an abnormality score for abnormal data based on normal data of the temperature and current; and (e) determining whether the propulsion system is normal or abnormal by comparing the calculated abnormality score with a threshold value.
[0013] The present invention relates to a method for detecting anomalies in a propulsion system for a railway vehicle, wherein in step (c), the learning is characterized in that two autoencoder models of the USAD (Unsupervised Anomaly Detection) algorithm learn normal data.
[0014] The present invention relates to a method for detecting abnormalities in a propulsion system for a railway vehicle, wherein in step (d), the abnormality score is calculated by two autoencoder models of the USAD algorithm learning normal data using temperature and current data of the railway vehicle propulsion system as input, and calculating the abnormality score based on reconstruction error for new data.
[0015] The present invention relates to a method for detecting abnormalities in a propulsion system for a railway vehicle, wherein in step (e), the determination of normality and abnormality is characterized by automatically detecting whether there is an abnormality by evaluating how different new data of the propulsion system is from the pattern of normal data through a learned model.
[0016] In addition, the present invention may further provide a railway vehicle operated by a propulsion device abnormality detection method. Effects of the invention
[0017] As described above, the present invention collects and preprocesses temperature and current data of a propulsion device and compares them with a threshold value that distinguishes between normal and abnormal to detect whether the device is normal or abnormal, thereby maximizing the lifespan of the propulsion device and significantly reducing excessive replacement costs.
[0018] In addition, the present invention has the effect of determining whether to detect anomalies with high reliability even for data that deviates from previously classified faults. Brief explanation of the drawing
[0019] FIG. 1 is a flowchart of a method for detecting abnormalities in a propulsion device for a railway vehicle according to a preferred embodiment of the present invention. FIG. 2 is a schematic diagram showing the structure of an autoencoder of a method for detecting abnormalities in a propulsion device for a railway vehicle according to a preferred embodiment of the present invention. Figure 3 is a schematic diagram showing the structure of a Generative Adversarial Network. FIG. 4 is a schematic diagram showing the autoencoder training structure of a method for detecting abnormalities in a railway vehicle propulsion system according to a preferred embodiment of the present invention. Figure 5 is a schematic diagram showing the structure of Adversarial Training. FIG. 6 is a schematic diagram showing a graph of temperature data generation for a method of detecting abnormalities in a propulsion device for a railway vehicle according to a preferred embodiment of the present invention. FIG. 7 is a schematic diagram showing an anomaly score graph of a method for detecting anomalies in a propulsion device for a railway vehicle according to a preferred embodiment of the present invention. FIG. 8 is a schematic diagram showing the results of a performance evaluation of a method for detecting abnormalities in a propulsion device for a railway vehicle according to a preferred embodiment of the present invention. Specific details for implementing the invention
[0020] Hereinafter, preferred embodiments according to the present invention will be described in more detail with reference to the attached drawings.
[0021] Here, components having the same function in all the drawings below use the same reference numerals, and repetitive descriptions are omitted. Additionally, it is specified that the terms described below are defined in consideration of their functions in the present invention and should be interpreted according to their inherent common meanings.
[0022] As illustrated in FIGS. 1 to 8, the method for detecting abnormalities in a heating and cooling device for a railway vehicle according to the present invention is broadly divided into the following steps: (a) collecting temperature and current data of a railway vehicle propulsion device (S10); (b) preprocessing raw data regarding the temperature and current of the propulsion device (S20); (c) performing learning using a reconstruction error from the preprocessed data (S30); (d) calculating an abnormal score for abnormal data based on normal data of temperature and current (S40); and (e) determining whether the propulsion device is normal or abnormal by comparing the calculated abnormal score with a threshold value (S50).
[0024] Step (a) above collects temperature and current data of the railway vehicle propulsion system. (S10)
[0025] To this end, temperature and current sensors are installed in the railway vehicle propulsion system to collect raw data.
[0026] In other words, the present invention utilized multivariate time-series data in seconds and used temperature / current variables for analysis.
[0027] In addition, only the time when the propulsion system is in operation was considered, and since there is a variation in the distribution of temperature variables depending on whether or not there is an influence of the driving wind, the analysis was conducted by distinguishing between them.
[0028] Here, when collecting measurement data through the temperature sensor and current sensor, it is desirable to receive only normal data.
[0029] The reason is that receiving only normal data as input results in higher reconstruction accuracy compared to receiving abnormal data.
[0030] Meanwhile, the propulsion system of a railway vehicle is a device that includes a traction motor, which plays a key role in controlling the movement and speed of the train, and moves the railway vehicle by utilizing propulsion force.
[0031] In this way, the propulsion system plays an important role in ensuring the safe movement of passengers and enabling efficient operation by precisely controlling the speed and motion of railway vehicles.
[0032] In addition, it is a key device that optimizes energy consumption through stable acceleration and deceleration of the train, helping to reduce long-term operating costs while maintaining vehicle performance.
[0033] Therefore, a monitoring and anomaly detection model for the propulsion system is essential.
[0035] Step (b) above preprocesses raw data regarding the temperature and current of the propulsion device. (S20)
[0036] In other words, the collected raw temperature and current data of the propulsion device undergo pre-processing such as clipping to reduce the data sample size to save data processing time, filtering to remove unwanted signals included in the signal, smoothing to smooth data that deviates significantly from the average, and normalization to adjust the data size so that it falls within a defined numerical range.
[0037] Here, the clipping operation is a task to reduce the data sample size in order to save data processing time for selected temperature and current variables.
[0038] Also, the filtering operation is a task that filters out unwanted signals included in the signal by applying a sliding window with a window size of 20 and a stride of 2 to the data set.
[0039] In addition, the smoothing operation is a process of splitting normalized data into training (70%) and validation (30%) sets to weaken or remove data that deviates significantly from the mean.
[0040] In addition, since there was no anomaly labeling data, to verify the algorithm's performance, anomaly data corresponding to both Context and Point anomaly perspectives was generated by multiplying a specific variable by a linear weight, and verification was conducted.
[0041] To elaborate on the validation, the Train Set and Test Set were constructed from the entire dataset in a 7:3 ratio. After performing Min-Max scaling preprocessing for each variable, anomalous data was generated within the same interval of the Test Set.
[0042] In addition, since the above data generally differ significantly in numerical values, it is necessary to normalize the data to values between 0 and 1.
[0044] Step (c) above performs learning using reconstruction error from the preprocessed data. (S30)
[0045] In other words, learning is performed using the deviation of the output reconstructed from the preprocessed input temperature and current data.
[0046] Here, machine learning is broadly divided into supervised learning and unsupervised learning. Unsupervised learning is a method of training a model in a state where there are no correct answers (labels) in the data.
[0047] It allows the model to learn the patterns or structures of the input data using only unlabeled data, focusing on understanding the intrinsic structure of the data.
[0048] The goal of unsupervised learning is to find groups with similar characteristics within data or to understand data distributions and patterns. For example, customer purchase data can be used to identify customer clusters that exhibit similar purchasing behaviors.
[0049] In this case, the data contains only information about 'items purchased by customers,' but there are no labels to divide customers into predefined groups.
[0050] Based on this data, unsupervised learning algorithms automatically classify customers into groups with similar characteristics.
[0051] Representative unsupervised learning algorithms include K-means clustering, hierarchical clustering, principal component analysis (PCA), and autoencoders.
[0052] Unsupervised learning-based anomaly detection models are primarily trained using only normal data to enable the model to understand normal patterns and structures, and are used to detect abnormal data (anomaly data) based on this understanding.
[0053] Representative unsupervised learning models widely used for anomaly detection include methods based on autoencoders and Generative Adversarial Networks (GANs), as shown in Figures 2 and 3.
[0054] An autoencoder has a neural network structure that compresses input data into a low-dimensional latent space and then restores it back to the original dimension. In this process, the network is trained to restore the input data as effectively as possible. An autoencoder trained exclusively on normal data exhibits low reconstruction error for normal data. However, it fails to restore abnormal data that it has not been trained on, resulting in high reconstruction error.
[0055] Therefore, if the reconstruction error exceeds a specific threshold, the data can be considered abnormal.
[0056] This method has the advantage of effectively detecting anomalies using only normal data, even in situations where there is little to no fault data.
[0057] Generative Adversarial Networks (GANs) are models composed of two neural networks: a Generator and a Discriminator. The Generator attempts to deceive the Discriminator by generating random samples, while the Discriminator attempts to distinguish between real data and generated data.
[0058] The two networks learn competitively, and the generator produces increasingly realistic data.
[0059] In GAN-based anomaly detection, a generator trained using normal data can produce samples that closely follow the distribution of normal data. Since the discriminator learns the difference between normal data and the generated data, when new data is given, it detects it as an anomaly if it does not fall within the distribution of normal data.
[0060] Anomaly detection methods using GANs have the advantage of being able to effectively learn complex data distributions, making them useful for detecting anomalies in high-dimensional data such as images or video data.
[0061] Among these, the present invention has two autoencoder models of the USAD (Unsupervised Anomaly Detection) algorithm learn normal data.
[0062] In other words, unsupervised anomaly detection is characterized by the ability of an unsupervised learning-based deep learning model to learn patterns even from unlabeled data. Therefore, it learns normal patterns and structures using only normal data, and subsequently acquires the ability to detect new data that differs from these normal patterns as anomalies.
[0063] This method enables the early detection of potential system problems even in the absence of failure data and can contribute to reducing unnecessary parts replacement or maintenance costs.
[0064] Furthermore, since unsupervised learning-based models independently identify the distribution and characteristics of data, they can effectively resolve the issue of insufficient fault data that arises in existing fault detection methods.
[0065] In this invention, even by utilizing only normal data, the model can define the criteria for a normal state and automatically detect abnormal situations that deviate from it, thereby improving the ability to predict and respond to failures before they occur.
[0066] Therefore, using an anomaly detection method based on an unsupervised learning-based deep learning model can improve the maintenance of railway vehicle propulsion systems in a more intelligent and efficient manner, and can be an important solution to minimize unexpected accidents or operational interruptions caused by failures.
[0067] In other words, the present invention enables the detection of abnormalities in railway vehicle propulsion systems by utilizing the USAD (UnSupervised Anomaly Detection) algorithm.
[0068] The USAD (Unsupervised Anomaly Detection) algorithm is an unsupervised learning-based anomaly detection method. It is a model designed to detect anomalies more effectively by combining the concepts of autoencoders and Generative Adversarial Networks (GANs).
[0069] This algorithm learns patterns in normal data and uses a combination of two approaches to detect abnormal data.
[0070] USAD demonstrates powerful performance in detecting abnormal conditions by utilizing only normal data, especially in situations where fault data is scarce or non-existent.
[0071] The operating principle of USAD consists of two structures: Autoencoder Training and Adversarial Training, as illustrated in Figures 4 and 5.
[0072] First, Autoencoder Training uses two autoencoders to train the data.
[0073] The first autoencoder (Autoencoder 1) learns the patterns of normal data through the process of compressing input data into a lower dimension and then restoring it. This autoencoder learns what characteristics the input data has and attempts to restore it to its original form as much as possible.
[0074] The second autoencoder (Autoencoder 2) takes the data restored by the first autoencoder as input and restores it to be closer to the original data.
[0075] In other words, it learns in a direction that reduces errors occurring during the restoration process of the first autoencoder. This dual autoencoder structure plays an important role in helping to learn normal data more accurately.
[0076] Next, Adversarial Training applies principles similar to GANs (Generative Adversarial Networks) to the interaction between autoencoders.
[0077] In GANs, the generator attempts to deceive the discriminator by creating fake data, while the discriminator has a competitive learning structure designed to distinguish between fake and real data.
[0078] In a similar manner at USAD, the first autoencoder attempts to reconstruct the input data as closely as possible to the original data, and the second autoencoder learns to minimize the reconstruction error of the first autoencoder.
[0079] The first autoencoder acts as a "generator" and is trained to effectively reconstruct the patterns of normal data, while the second autoencoder acts as a "discriminator" to evaluate how close the reconstructed data is to the normal data.
[0080] Through such competitive learning, the two autoencoders learn the characteristics of normal data more precisely.
[0081] Finally, the USAD algorithm detects anomalies by calculating the reconstruction error between the outputs of the two trained autoencoders and the actual input data.
[0082] Normal data maintains a low reconstruction error even after passing through two autoencoders, but abnormal data exhibits a high reconstruction error during the process of passing through the two autoencoders.
[0083] Therefore, USAD considers data abnormal if the reconstruction error exceeds a specific threshold. This result occurs because abnormal data does not follow the patterns of normal data.
[0084] That is, abnormal data was generated in a section corresponding to approximately 10% of the Test Set data, and training was conducted assuming that the Train Set contained only normal data.
[0085] Hyperparameters such as the Anomaly Score threshold and Window Size were set to optimal values through various simulations.
[0087] Step (d) above calculates an abnormal score for abnormal data based on normal data of temperature and current. (S40)
[0088] In other words, the trained USAD model calculates an anomaly score for artificially generated anomaly data based on data from the normal state.
[0089] The two training methods explained earlier can be expressed as a single loss function as follows.
[0090]
[0091] The larger this value is, the more it is identified as an outlier, and it is also called the Anomaly Score. For reference, a loss function is a function that measures the difference between the model's predicted value and the actual value (target value) during the process of training machine learning and deep learning models.
[0092] The value of the loss function indicates how well the model is predicting, and the model is trained to minimize this value.
[0093] In other words, the loss function evaluates the model's performance by quantifying the error or loss between the model's predicted result and the actual value, and plays a role in adjusting the model's parameters (e.g., weights and biases) to reduce that error.
[0094] By integrating these methods, it is possible to learn patterns in normal data more precisely and effectively detect abnormal data. Autoencoders learn the characteristics of normal data through the process of compressing and restoring input data, while the competitive learning method of GANs enables the model to acquire the ability to adapt to various patterns.
[0095] This combined approach significantly improves anomaly detection performance and provides higher accuracy and reliability compared to existing single models.
[0096] Second, it offers high data efficiency. Since the USAD algorithm is based on unsupervised learning, it can sufficiently perform training using only normal data.
[0097] In general, in anomaly detection problems, abnormal data (e.g., failure data) is very rare or often absent.
[0098] In this situation, USAD can effectively train a model using only normal data, so it can be used even when labeled abnormal data is scarce.
[0099] This is particularly useful in fields such as railway rolling stock, manufacturing processes, and financial transactions, enabling the ability to detect and respond to unexpected problems early.
[0100] Third, generalization ability is strengthened. USAD has a structure in which two autoencoders learn competitively, allowing it to effectively learn patterns from various normal data.
[0101] This enables the model to adapt well to new situations or unexpected data. For example, a USAD model that has learned all the various changes in the normal state has a higher ability to detect unexpected abnormal situations even if they occur.
[0102] Therefore, USAD is a powerful model that can be flexibly applied in various data environments and can overcome the limitations of existing anomaly detection models.
[0103] Fourth, it has a wide and flexible scope of application. The USAD algorithm can be used to solve anomaly detection problems in various industries, such as finance, network security, manufacturing, healthcare, and logistics.
[0104] Since this algorithm trains the model using only normal data and can effectively detect abnormal data, it can be applied even in situations where there is no labeled data or almost no abnormal data.
[0105] In addition, it delivers high performance even with complex high-dimensional data, providing the flexibility to process various types of data, such as images, video, audio, and time-series data.
[0106] Finally, it can play a crucial role in early warning and preventive maintenance. By accurately learning patterns in normal data to detect abnormal situations early, USAD helps identify problems in advance and take appropriate measures.
[0107] This enhances the operational stability of the system, reduces costs, and provides the effect of preventing potential risks in advance.
[0108] In this invention, multivariate time-series data in seconds was utilized, and temperature and current variables were used for analysis. Furthermore, only the time when the propulsion device was in operation was considered, and since there is a variation in the distribution of temperature variables depending on the influence of driving wind, the analysis was conducted by distinguishing between them.
[0109] In addition, since there was no anomaly labeling data, to verify the algorithm's performance, anomaly data corresponding to both Context and Point anomaly perspectives was generated by multiplying a specific variable by a linear weight, and verification was conducted.
[0110] To elaborate on the validation, the Train Set and Test Set were constructed from the entire dataset in a 7:3 ratio. After performing Min-Max scaling preprocessing for each variable, anomalous data was generated within the same interval of the Test Set.
[0111] That is, abnormal data was generated in a section corresponding to approximately 10% of the Test Set data, and training was conducted assuming that the Train Set contained only normal data.
[0112] Hyperparameters such as the Anomaly Score threshold and Window Size were set to optimal values through various simulations. The analysis results are shown in Figure 6.
[0113] To explain the analysis results, first, as shown in Fig. 6, the result of artificially generating abnormal data in a specific section of temperature data can be seen. In Fig. 6, the green line represents the original normal data (existing data), and the red line represents the newly generated abnormal data.
[0114] The aim is to evaluate how different the data is from normal data and whether such differences can be detected by inserting anomalous data through artificial changes in specific sections.
[0115] In this way, abnormal data was generated in a similar manner for the same interval not only for temperature data but also for the remaining variables (current, etc.).
[0116] In other words, for each variable, data exhibiting patterns different from the normal state within specific time periods or ranges was intentionally generated, and this was evaluated to see if the model could effectively detect it.
[0117] Next, as shown in Fig. 7, the Anomaly Score derived by applying the USAD (unsupervised anomaly detection) algorithm to the anomaly data interval can be known.
[0118] The USAD algorithm calculates anomaly scores to distinguish between normal and abnormal data.
[0119] In other words, the trained USAD model calculates an anomaly score for artificially generated anomaly data based on data from the normal state.
[0121] Step (e) above determines whether the propulsion device is normal or abnormal by comparing the calculated abnormal score with a threshold value. (S50)
[0122] In other words, based on the set threshold, if the abnormal score exceeds the threshold, the data is determined to be abnormal data, and if it is below the threshold, it is determined to be normal data.
[0123] In this way, based on the threshold set by the inventor, if the abnormal score exceeds the threshold, the corresponding data is determined to be abnormal data, and if it is below the threshold, it is determined to be normal data.
[0124] In this invention, the calculated abnormal score is used by the model to determine whether each data point is normal or abnormal.
[0125] As can be seen from the graph in Figure 7, the Anomaly Score is significantly higher in the section where abnormal data was artificially inserted. This means that the section was recognized as an abnormal pattern by the model when compared to other normal data sections.
[0126] On the other hand, in sections where no anomalous data was inserted, the Anomaly Score remained low, indicating that the data in these sections was recognized as a normal pattern by the model.
[0127] Therefore, Figure 7 is an important result showing that the USAD algorithm can effectively distinguish between normal data and abnormal data.
[0128] As shown in Figure 8, the results regarding the performance evaluation indicators of the USAD algorithm can be seen. This allows us to verify how accurately the USAD algorithm recognizes normal data as normal and identifies abnormal data as abnormal through three key performance indicators.
[0129] These metrics play an important role in evaluating the accuracy and effectiveness of a model and provide criteria for determining how reliably the model can operate in real-world environments.
[0130] For reference, precision refers to the ratio of data that is actually anomaly among the data predicted as "anomaly" by the model. In other words, it measures how much of the data identified as anomaly is actually anomaly out of all the data classified as such by the model.
[0131] And, recall refers to the ratio of data that the model correctly identified as "anomalies" among the data that are actually "anomalies."
[0132] In other words, recall is an indicator of how well the model detected actual anomaly data. Finally, the F1-Score is the harmonic mean that considers both precision and recall. It focuses on balancing precision and recall and evaluates how well the balance between the two metrics is maintained.
[0133] In conclusion, the results in Figure 8 confirm that the USAD algorithm demonstrated excellent performance in recognizing normal data as normal and accurately detecting abnormal data.
[0134] This proves that the algorithm is a highly reliable method that can be effectively used in various real-world situations, and serves as important evidence demonstrating its potential for application in multiple fields.
[0135] Therefore, the present invention can monitor the status of a railway vehicle propulsion system in real time and detect abnormal situations quickly and accurately through such a method.
[0136] The present invention described above is not limited by the aforementioned embodiments and attached drawings, and it will be obvious to those skilled in the art that various substitutions, modifications, and changes to equivalent alternative embodiments are possible within the scope of the technical concept of the present invention.
Claims
Claim 1 A method for detecting abnormalities in a propulsion system for a railway vehicle, comprising: (a) collecting temperature and current data of the propulsion system for the railway vehicle; (b) preprocessing raw data regarding the temperature and current of the propulsion system; (c) performing learning using a reconstruction error from the preprocessed data; (d) calculating an abnormality score for abnormal data based on normal data of the temperature and current; and (e) determining whether the propulsion system is normal or abnormal by comparing the calculated abnormality score with a threshold value. Claim 2 A method for detecting anomalies in a railway vehicle propulsion system according to claim 1, wherein in step (c), the learning is characterized by two autoencoder models of the USAD (Unsupervised Anomaly Detection) algorithm learning normal data. Claim 3 A method for detecting abnormalities in a railway vehicle propulsion system according to claim 1, wherein in step (d), the abnormal score is calculated by two autoencoder models of the USAD algorithm learning normal data using temperature and current data of the railway vehicle propulsion system as input, and calculating the abnormal score based on reconstruction error for new data. Claim 4 A method for detecting abnormalities in a propulsion device for a railway vehicle according to claim 1, wherein in step (e), the determination of normal and abnormal is characterized by automatically detecting whether there is an abnormality by evaluating how different new data of the propulsion device is from the pattern of normal data through a learned model. Claim 5 A railway vehicle operated by the propulsion device abnormality detection method described in any one of claims 1 to 4.