Anomaly detection system and program using electromagnetic wave radar

Abstract

This system provides an anomaly detection system that enables non-contact monitoring of internal conditions, is less susceptible to environmental influences, and can identify the location of anomalies in real time. [Solution] An anomaly detection system using electromagnetic wave radar, comprising: a radar signal processing unit that applies range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; a time series signal generation unit that divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time series waveform data; a feature extraction unit that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing; and an anomaly detection unit that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the features.

Description

Technical Field

[0001] The present invention relates to an abnormality detection system and a program using an electromagnetic wave radar.

Background Art

[0002] In recent years, in factories, warehouses, infrastructure facilities, etc., the importance of monitoring the status of facilities and articles and detecting abnormalities at an early stage has been increasing. In conventional abnormality detection systems, optical sensors, temperature sensors, etc. have been used, but these sensors are easily affected by dust and smoke, and there is also a problem that they can only detect surface changes of the object. In addition, in order to ensure the safety of workers in factories and warehouses, detecting the presence of people in the work space and preventing dangerous situations in advance has also become an important issue. In the detection by conventional monitoring cameras, there is a problem that it is easily affected by lighting conditions and dead angles, and it is difficult to distinguish a person from other heat sources with a temperature sensor.

[0003] Patent Document 1 discloses a sensor device that combines a filter bank, normalization means, and recognition means to identify an object in a detection area using a radio wave sensor and reduce the influence of background noise.

[0004] In addition, Patent Document 2 discloses an imaging device using high-resolution millimeter waves, and a technique that enables distance measurement in the FMCW radar method while suppressing the generation of standing waves by using a broadband millimeter wave noise signal.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, Patent Document 1 primarily focuses on frequency domain analysis using a filter bank and specific background noise removal, and was unable to perform advanced anomaly detection through time-series learning processing. Furthermore, Patent Document 2 is primarily aimed at distance measurement and does not describe processes such as time-series analysis or anomaly identification, making it difficult to detect complex anomalies.

[0007] Therefore, the present invention aims to provide an anomaly detection system that enables non-contact monitoring of internal conditions, is less susceptible to environmental influences, and can identify the location of anomalies in real time. [Means for solving the problem]

[0008] To solve the above problems, the present invention provides an anomaly detection system using electromagnetic wave radar, comprising: a radar signal processing unit that applies range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; a time series signal generation unit that divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time series waveform data; a feature extraction unit that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing; and an anomaly detection unit that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the features. [Effects of the Invention]

[0009] According to the present invention, it is possible to provide an anomaly detection system that enables non-contact monitoring of internal conditions, is less susceptible to environmental influences, and can identify the location of anomalies in real time. [Brief explanation of the drawing]

[0010] [Figure 1] This is a block diagram showing an example of the overall configuration of the anomaly detection system according to the first embodiment. [Figure 2]This is a configuration diagram showing the hardware configuration of the anomaly detection device according to the first embodiment. [Figure 3] This flowchart shows an example of the range Doppler image generation process by the radar signal processing unit according to the first embodiment. [Figure 4] This is an explanatory diagram illustrating an example of the concept of grid division of a range Doppler image according to the first embodiment. [Figure 5] This is a flowchart showing an example of the overall processing of the anomaly detection system according to the first embodiment. [Figure 6] This is a flowchart showing an example of the processing flow in the abnormality detection unit according to the first embodiment. [Figure 7] This figure shows an example of a display that visualizes the anomaly detection results for each grid according to the first embodiment. [Figure 8] This is a flowchart showing an example of the overall processing of the anomaly detection system in the second embodiment of the present invention. [Modes for carrying out the invention]

[0011] The present invention will be described below with reference to the drawings. However, the present invention is not limited to the embodiments shown below, and can be modified, added, altered, or deleted to the extent that a person skilled in the art can conceive of it. Any embodiment that achieves the function and effect of the present invention is included within the scope of the present invention.

[0012] <First Embodiment> FIG. 1 is a block diagram showing the overall configuration of the abnormality detection system 1 of the present embodiment. As a target for abnormality detection, for example, the presence of a person in the detection area of an electromagnetic radar can be considered. The abnormality detection system 1 includes an electromagnetic radar 10, a radar signal processing unit 20, a time-series signal generation unit 30, a feature amount extraction unit 40, and an abnormality detection unit 50. Range FFT and Doppler FFT are applied to the received IQ signal to generate a range-Doppler image. This range-Doppler image is grid-divided, and inverse Fourier transform (iFFT) is applied to the signals in each grid area to generate time-series waveform data. Filter bank processing is applied to extract feature amounts using a dynamic time-series model including reservoir computing, and an abnormality is detected based on the reconstruction error obtained by applying principal component analysis (PCA) to this feature amount.

[0013] The feature amount extraction unit 40 preferably includes a process of applying a non-linear reservoir using an Echo State Network (ESN).

[0014] In addition, the feature amount extraction unit 40 preferably includes a process of applying an autoregressive neural network (AR-NN) that takes into account the autoregressive property of time-series data and dynamically updates the state based on the autoregressive characteristics based on the time-series data up to the previous time, and a process of applying a non-linear activation function (Rectified Linear Unit: ReLU, sigmoid function, or hyperbolic tangent function) to the output of the autoregressive neural network.

[0015] In addition, the abnormality detection unit 50 preferably performs abnormality detection for each of the grids.

[0016] In addition, the abnormality detection unit 50 preferably performs abnormality detection based on the variance in the principal component space obtained by the principal component analysis.

[0017] Further, the electromagnetic wave radar 10 is a millimeter wave radar using the FMCW method, and it is preferable that the radar signal processing unit 20 generates sector-shaped data.

[0018] Further, the feature amount extraction unit 40 preferably further includes a learning unit that extracts feature amounts from the waveform data using a filter bank composed of filters in a plurality of frequency bands, and performs learning for anomaly detection using the time-series pattern of the feature amounts.

[0019] The anomaly detection system 1 includes an electromagnetic wave radar 10, a radar signal processing unit 20, a time-series signal generation unit 30, a feature amount extraction unit 40, an anomaly detection unit 50, and a learning unit 60, and realizes highly accurate anomaly detection by combining these components.

[0020] The electromagnetic wave radar 10 is a radar device for transmitting and receiving electromagnetic waves to detect the state of an object. Specifically, it is configured as a millimeter wave radar using the FMCW method.

[0021] The FMCW method millimeter wave radar adopts the frequency-modulated continuous wave method and enables highly accurate distance measurement and speed measurement.

[0022] The radar signal processing unit 20 applies range FFT and Doppler FFT to the IQ signal received from the electromagnetic wave radar 10 to generate a range-Doppler image 22. Further, the radar signal processing unit 20 generates sector-shaped two-dimensional data.

[0023] The range-Doppler image 22 is image data including two-dimensional information on distance and speed. This data is grid-divided in the time-series signal generation unit 30 and processed for each grid area 31.

[0024] The time-series signal generation unit 30 performs a process of grid-dividing the range-Doppler image 22 and applying inverse Fourier transform (iFFT) to the signals in each grid area 31 to generate time-series waveform data 32.

[0025] The grid region 31 refers to each region obtained by dividing the range Doppler image 22 into a grid, and anomaly detection is performed at this unit. Processing in each grid region 31 makes it possible to identify the location of the anomaly.

[0026] The feature extraction unit 40 applies filter bank processing and extracts features using a dynamic time series model including reservoir computing 42. In particular, it performs nonlinear feature extraction using an Echo State Network.

[0027] The feature extraction unit 40 may also include a function to analyze the micro-Doppler effect, particularly that caused by human movement. Periodic movements associated with walking or working by the human body appear as characteristic patterns on the range Doppler image 22. By detecting these patterns and combining frequency analysis using a filter bank with time-series analysis using reservoir computing, the presence of a person can be detected with high accuracy. In particular, nonlinear feature extraction using an ESN is effective in recognizing complex time-series patterns associated with human movement.

[0028] A filter bank consists of filters across multiple frequency bands and is used to extract features from time-series waveform data. Each filter captures features from a different frequency band.

[0029] Reservoir computing is a type of dynamic time series model and a computational method used for feature extraction. It has a large-scale neural network internally and efficiently extracts nonlinear features from time series data.

[0030] The Echo State Network is a special neural network used in the feature extraction unit 40, and includes a process that applies a nonlinear reservoir. This allows it to capture the complex dynamic features of time-series data.

[0031] The Echo State Network uses a nonlinear reservoir to enable nonlinear feature extraction. Its internal connection weights are randomly set and used in a fixed state.

[0032] The anomaly detection unit 50 detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the feature quantities. Furthermore, by performing anomaly detection for each grid region 31, the location of the anomaly can be identified.

[0033] The learning unit 60 performs learning for anomaly detection using the time-series patterns of the features. It learns using only normal data and constructs a model that serves as the basis for anomaly detection.

[0034] <Second Embodiment> As a second embodiment, the feature extraction unit 40 may also include a process that applies an autoregressive neural network (AR-NN) that takes into account the autoregressive properties of time series data, dynamically updating the state based on the autoregressive characteristics of the time series data up to the previous time point, and applying a nonlinear activation function (Rectified Linear Unit: ReLU, sigmoid function, or hyperbolic tangent function) to the output of the autoregressive neural network. [Examples]

[0035] The present invention will be described in more detail below with reference to examples, but the present invention is not limited to these examples.

[0036] The anomaly detection system 1 of the present invention, as shown in Figure 1, will be described below. The anomaly detection system 1 comprises an electromagnetic wave radar 10, a radar signal processing unit 20, a time-series signal generation unit 30, a feature extraction unit 40, an anomaly detection unit 50, and a learning unit 60. The electromagnetic wave radar 10 is configured as an FMCW type millimeter-wave radar, which irradiates an object with frequency-modulated continuous waves and receives the reflected waves. Millimeter-wave radar is less susceptible to the effects of dust and smoke and can detect the internal state of an object.

[0037] Figure 2 shows the hardware configuration of the anomaly detection device 100 related to the anomaly detection system 1. The anomaly detection device 100 is an information processing device comprising a processor 2, memory 3, storage unit 4, user interface unit 5, interface unit 6, and communication bus 7. The storage unit 4 is provided with a storage area 8 for storing programs necessary for anomaly detection. An electromagnetic wave radar 10 is connected to the interface unit 6, enabling data processing and recording for the entire system. This allows the execution of a series of processes from signal detection by the electromagnetic wave radar 10 to anomaly detection. The processor 2 is, for example, a CPU (Central Processing Unit), and the memory 2 includes RAM (Random Access Memory: volatile memory) and ROM (Read Only Memory: non-volatile memory). The processor 2 executes various calculations according to the programs loaded into the memory 2 and controls the entire system. The RAM is used as the work area of ​​the processor 2 and temporarily stores programs and data necessary for the processing of the processor 2. The ROM non-volatilely stores the basic control programs and various parameters of the processor 2. The memory unit 4 is a non-volatile high-capacity storage device such as an SSD (Solid State Drive) or HDD (Hard Disk Drive), and the user interface unit 5 includes, for example, a display, keyboard, and mouse. This hardware configuration enables the anomaly detection device 100 to achieve high-speed and stable processing capabilities, and to efficiently process information from the electromagnetic wave radar 10. Specifically, the anomaly detection device 100 implements the functions of the radar signal processing unit 20, the time-series signal generation unit 30, the feature extraction unit 40, and the anomaly detection unit 50 by executing a program.

[0038] Referring to Figure 3, the specific processing flow by the radar signal processing unit 20 will be described in detail. The radar signal processing unit 20 takes the IQ signal output from the electromagnetic wave radar 10, for example, an FMCW type millimeter wave radar, as input and generates a range Doppler image 22 including distance and velocity information by performing multiple signal processing operations in a continuous and stepwise manner. Each processing step will be described in detail below.

[0039] Firstly, in step ST1, the electromagnetic waves transmitted by the electromagnetic wave radar 10 are reflected by the target object, and the reflected waves are received and output as an IQ signal. This IQ signal contains basic data for analyzing the movement and distance of the target object, based on the signal strength and phase information from the object.

[0040] Next, in step ST2, the received IQ signal is subjected to range FFT processing. In this process, the Fast Fourier Transform is applied to convert the received signal in the distance direction and calculate the frequency spectrum. As a result, it becomes possible to accurately identify the distance information of each reflection point based on the frequency changes obtained from the reflected signal of the object.

[0041] Next, in step ST3, Doppler FFT processing is applied to the data from which distance information has been extracted by range FFT processing. This Doppler FFT processing analyzes the temporal frequency fluctuations of the reflected signal and extracts velocity information of the object. In other words, by utilizing the Doppler effect, it is possible to determine the velocity vector, which includes the direction of whether the object is approaching or moving away from the sensor.

[0042] Furthermore, in step ST4, the results of these two FFT processes are integrated to construct a two-dimensional range Doppler image 22 containing distance and velocity information of the object. This image associates velocity information with the horizontal axis and distance information with the vertical axis of the coordinate system, allowing for the visualization of spatial and dynamic information. In particular, this range Doppler image 22 is represented as fan-shaped two-dimensional data, enabling efficient capture of the object's behavior in its surrounding dynamic environment.

[0043] By executing this processing flow, a range Doppler image 22 is generated, retaining its spatial and temporal characteristics, and is then subjected to the next stage of grid division and time series analysis processing. The image data thus prepared forms the basis for accurately detecting the presence or absence of abnormal conditions.

[0044] The following describes in detail the process of dividing the range Doppler image 22 into grid regions 31 in the anomaly detection system 1 of the present invention, focusing on Figure 4.

[0045] As shown in Figure 4, the range Doppler image 22 is two-dimensional image data containing distance information and velocity information, and in this invention, each grid region 31 serves as the basis for anomaly detection as an individual unit. The range Doppler image 22 to be divided is regularly divided into a grid of horizontal angle bins and vertical distance bins.

[0046] The range Doppler image 22 generated by the radar signal processing unit 20 is divided into grid regions 31 by the time-series signal generation unit 30. In this division process, each grid region 31 is divided equally, and each is treated as an independent unit of analysis. The number of divisions can be set as appropriate depending on the target system and application, and can be configured in shapes such as 10x10 or 100x100.

[0047] Next, an inverse Fourier transform (iFFT) is applied to each grid region 31 to generate time-series waveform data 32. This process makes it possible to extract signal components corresponding to specific distance and velocity ranges within each grid region 31 along the time axis. The obtained time-series waveform data retains dynamic information about the object and serves as foundational data for detecting anomalies in real time.

[0048] Furthermore, as shown in Figure 4, which displays subdivided range Doppler images 22 of each grid region 31, this process enables further refinement of anomaly detection. This allows for the provision of detection results with increased spatial resolution and the identification of more localized anomalies in the target object.

[0049] In the anomaly detection system 1 of the present invention, by using time-series waveform data obtained through such grid division and inverse Fourier transform processing, it becomes possible to efficiently capture the internal state and dynamic changes of an object and perform anomaly detection with high accuracy.

[0050] Referring to Figure 5, the overall processing flow of the anomaly detection system 1 will be explained in detail step by step based on a flowchart.

[0051] In the first step (SU1), the radar signal processing unit 20 applies range FFT and Doppler FFT processing to the IQ signals transmitted and received from the electromagnetic wave radar 10 to generate a range Doppler image 22. This image has a two-dimensional structure that includes spatial distance information and velocity information of reflectors. This allows for the acquisition of data that comprehensively captures the behavior of the target object.

[0052] Next, in step (SU2), the time-series signal generation unit 30 divides the generated range Doppler image 22 into grid regions 31. This division process divides the image into small regional units, and by applying an inverse Fourier transform (iFFT) to the signal in each region, time-series waveform data is generated. This time-series data becomes detailed data that reflects the position and movement of the object in each grid unit.

[0053] In the following step (SU3), the feature extraction unit 40 performs frequency division processing on the time-series waveform data using a filter bank. This filter bank, for example, covers the range from 1 Hz to 100 Hz and consists of multiple bandpass filters divided into 10 bands. The data for each band extracted by each filter becomes the basic data for analyzing the behavior and characteristics of the object from a frequency perspective.

[0054] In step (SU4), reservoir computing 42 is run based on the data extracted by filtering. This process utilizes an Echo State Network (ESN) to construct a large-scale recurrent neural network, particularly one with a nonlinear reservoir 44 internally. This network contains, for example, more than 1000 neurons and is excellent at extracting complex dynamic features from time-series data. This process yields important features for anomaly detection.

[0055] In the subsequent step (SU5), the anomaly detection unit 50 applies principal component analysis (PCA) to the features obtained from the ESN. This PCA process reduces the dimensionality of the features and calculates the variance and reconstruction error in the principal component space 51. This makes it possible to sensitively detect anomaly locations that have characteristics different from normal.

[0056] Finally, in step (SU6), the location of the anomalies and the counting of the objects are performed based on the anomalies detected by PCA. The results of this process are output in a format that allows for intuitive understanding of the presence and severity of anomalies, for example, by displaying the anomalies for each grid region 31.

[0057] In this way, the anomaly detection system 1 of the present invention can analyze the dynamic behavior of an object in detail from both spatial and temporal perspectives and detect anomalies with high accuracy.

[0058] Referring to Figure 6, the processing flow in the anomaly detection unit 50 will be explained step by step.

[0059] First, the anomaly detection unit 50 receives the features extracted by the feature extraction unit 40 (step SV1). At this time, the features obtained in each grid region 31 are processed individually. This method makes it possible to examine the anomaly state on a grid-by-grid basis.

[0060] Next, principal component analysis (PCA) is applied to the received features (step SV2). PCA reduces the dimensionality of the features and transforms only the most important components into the principal component space. This process reduces the redundancy of the feature data and improves the efficiency of anomaly detection.

[0061] Next, the reconstruction error and variance within the principal component space are calculated (step SV3). The reconstruction error is the difference between the original feature and the feature reconstructed from the principal component space, and serves as an indicator to improve the sensitivity of anomaly detection. On the other hand, the variance represents the variation in principal component values ​​for each grid region 31, and this is used to further distinguish anomalies in detail.

[0062] If these indicators exceed pre-set thresholds, an anomaly is detected in the corresponding grid region 31 (step SV4). For example, if the reconstruction error exceeds 3σ under normal conditions, it is determined to be an anomaly. Since this anomaly determination is performed independently for each grid region 31, it is possible to accurately pinpoint the spatial location of the anomaly.

[0063] Finally, the judgment results are visualized and output (step SV5). Grid areas 31 judged as abnormal are displayed in black, for example, while areas judged as normal are displayed in white, for example. Furthermore, grayscale representation according to the degree of abnormality is also possible, allowing for a visual understanding of the severity of the abnormality.

[0064] Referring to Figure 7, the visualization of anomaly detection results will be explained. The judgment results from the anomaly detection unit 50 are visualized for each grid region 31. Normal areas are displayed in white, and areas where anomalies are detected are displayed in black. Furthermore, it is possible to display the results in grayscale depending on the degree of anomaly, allowing for an intuitive understanding of the location and importance of the anomaly.

[0065] Referring to Figure 8, the processing flow of the anomaly detection system 1 in the second embodiment will be described. In this embodiment, an autoregressive neural network is used instead of an Echo State Network. The autoregressive neural network uses time-series data from the past 10 points in time as input and performs feature extraction while considering the autoregressive nature of the time-series data by applying nonlinear activation functions such as ReLU and sigmoid functions. This makes it possible to capture changes in the time-series pattern in more detail. Other processing is the same as in the first embodiment.

[0066] <Aspect 1> An anomaly detection system using electromagnetic wave radar, comprising: a radar signal processing unit 20 that applies range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; a time series signal generation unit 30 that divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time series waveform data; a feature extraction unit 40 that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing; and an anomaly detection unit 50 that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the features.

[0067] According to this embodiment, in an anomaly detection system using electromagnetic wave radar, a range Doppler image is generated by the radar signal processing unit, and time-series data is acquired in grid units by the time-series signal generation unit, thereby enabling anomaly detection while ensuring detailed spatial resolution. Furthermore, by utilizing filter bank processing and reservoir computing in the feature extraction unit, feature extraction that takes into account the complex dynamic changes of time-series data becomes possible. In addition, by performing anomaly detection based on reconstruction error using principal component analysis in the anomaly detection unit, highly accurate anomaly detection is achieved. As a result, the accuracy of anomaly detection is improved compared to conventional methods, and a more reliable anomaly detection system can be provided.

[0068] <Aspect 2> The feature extraction unit 40 includes a process that applies a nonlinear reservoir using an Echo State Network (ESN). An anomaly detection system according to embodiment 1, characterized by the features described above.

[0069] According to this embodiment, by utilizing an Echo State Network (ESN), it becomes possible to extract features that take into account the long-term dependencies of time-series data. By applying a nonlinear reservoir, more complex dynamic changes can be captured, and thus an improvement in the accuracy of anomaly detection can be expected. In particular, since nonlinearities that cannot be captured by ordinary linear models can be taken into account, the occurrence of anomalies can be detected with higher accuracy.

[0070] <Aspect 3> The anomaly detection system according to embodiment 1, characterized in that the feature extraction unit 40 includes a process of applying an autoregressive neural network (AR-NN) that takes into account the autoregressiveness of time series data and dynamically updating the state based on the autoregressive characteristics of the time series data up to the previous time, and a process of applying a nonlinear activation function (Rectified Linear Unit: ReLU, sigmoid function, or hyperbolic tangent function) to the output of the autoregressive neural network.

[0071] According to this embodiment, by applying an autoregressive neural network (AR-NN), dynamic feature extraction that takes into account the autoregressive characteristics of time-series data becomes possible. Furthermore, by applying a nonlinear activation function, complex data patterns that cannot be captured by linear regression models can be extracted, thereby improving the accuracy of anomaly detection. In particular, anomalies with strong temporal dependencies can be effectively detected, resulting in more adaptive anomaly detection than before.

[0072] <Aspect 4> The anomaly detection system according to any one of embodiments 1 to 3, characterized in that the anomaly detection unit 50 performs anomaly detection for each grid.

[0073] According to this embodiment, by performing anomaly detection on a grid basis, it becomes possible to detect localized anomalies with high accuracy. In particular, when monitoring large areas, it becomes easier to identify the location of anomalies, and appropriate responses can be implemented quickly. Furthermore, by improving the accuracy of anomaly detection, false positives can be reduced, and the reliability of the entire system can be enhanced.

[0074] <Aspect 5> The anomaly detection system according to any one of embodiments 1 to 3, characterized in that the anomaly detection unit 40 performs anomaly detection based on the variance in the principal component space obtained by principal component analysis.

[0075] According to this embodiment, by utilizing the variance of the principal component space obtained by principal component analysis (PCA), the anomaly of the data can be determined with high accuracy. When an anomaly occurs, the variance in the principal component space changes significantly, making it an effective indicator for anomaly detection. This allows for a clear distinction between normal and abnormal data, enabling highly accurate anomaly detection while suppressing false positives.

[0076] <Aspect 6> The anomaly detection system according to any one of embodiments 1 to 3, characterized in that the electromagnetic wave radar 10 is an FMCW type millimeter wave radar, and the radar signal processing unit 10 generates fan-shaped data.

[0077] According to this embodiment, by employing a millimeter-wave radar using the FMCW method, high distance resolution and Doppler resolution can be ensured, enabling the acquisition of detailed dynamic information of the target object. Furthermore, by generating fan-shaped data in the radar signal processing unit, environmental information over a wide area can be processed efficiently, expanding the anomaly detection range. As a result, compared to conventional radar systems, more precise anomaly detection is possible, providing a highly reliable system.

[0078] <Aspect 7> An anomaly detection system according to any one of embodiments 1 to 3, characterized in that the feature extraction unit 40 further comprises a learning unit 60 that extracts features from waveform data using a filter bank consisting of filters of multiple frequency bands and performs learning for anomaly detection using the time-series pattern of the features.

[0079] According to this embodiment, feature extraction using a filter bank allows for detailed analysis of information across different frequency bands, thereby improving the sensitivity of anomaly detection. Furthermore, by incorporating a learning unit, it becomes possible to learn the time-series patterns of features and continuously improve the accuracy of anomaly detection. This results in an anomaly detection system that is more adaptable to environmental changes, and is expected to reduce false positives and improve detection performance.

[0080] <Aspect 8> A program to cause a computer to function as: a radar signal processing means that applies range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; a time-series signal generation means that divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time-series waveform data; a feature extraction means that applies filter bank processing and extracts features using a dynamic time-series model including reservoir computing; and an anomaly detection means that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the features.

[0081] According to this embodiment, by implementing this program on a computer, the operation of the anomaly detection system can be realized on a software basis. This makes it possible to flexibly build and operate the system without depending on the hardware configuration. Furthermore, by updating the program, the anomaly detection algorithm can be improved, enabling continuous performance improvement. In addition, by combining it with a cloud environment or a distributed processing system, it becomes easy to apply it to large-scale data processing and real-time anomaly detection, providing a more advanced anomaly monitoring system.

[0082] <Pattern 9> An anomaly detection device comprising a processor and memory, wherein the processor performs radar signal processing, which involves applying range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; time series signal generation, which involves grid-dividing the range Doppler image and applying inverse Fourier transform (iFFT) to the signal in each grid region to generate time series waveform data; feature extraction, which involves applying filter bank processing and extracting features using a dynamic time series model including reservoir computing; and anomaly detection, which involves detecting anomalies based on reconstruction errors obtained by applying principal component analysis (PCA) to the features.

[0083] According to this embodiment, by applying range FFT and Doppler FFT to the received IQ signal, a range Doppler image in the frequency domain can be generated, allowing for a detailed understanding of the spatial and temporal characteristics of the signal. Furthermore, by dividing the range Doppler image into a grid and applying the inverse Fourier transform (iFFT) to the signal in each grid region, waveform data in the time domain can be obtained, enabling detailed time-series analysis. In addition, feature extraction using a dynamic time-series model including filter bank processing and reservoir computing allows for highly accurate analysis of the complex temporal changes of the signal. Moreover, by performing anomaly detection using reconstruction error by principal component analysis (PCA), minute anomalies that are difficult to detect with conventional methods can be detected with high sensitivity, thereby improving the accuracy of the anomaly detection device.

[0084] <Aspect 10> An anomaly detection method comprising: a radar signal processing step of applying range FFT and Doppler FFT to a received IQ signal to generate a range Doppler image; a time series signal generation step of dividing the range Doppler image into a grid and applying inverse Fourier transform (iFFT) to the signal in each grid region to generate time series waveform data; a feature extraction step of applying filter bank processing and extracting features using a dynamic time series model including reservoir computing; and an anomaly detection step of detecting anomalies based on reconstruction errors obtained by applying principal component analysis (PCA) to the features.

[0085] According to this embodiment, by applying range FFT and Doppler FFT, the frequency characteristics of the received signal can be clarified and visualized as a range Doppler image. Furthermore, by dividing the range Doppler image into a grid and applying the inverse Fourier transform (iFFT) to each grid region, the temporal signal changes in each region can be analyzed in detail. This enables local anomaly detection, and an improvement in system accuracy can be expected. In addition, by applying filter bank processing and using a dynamic time series model including reservoir computing, features of complex time series data with nonlinearity can be extracted with high accuracy. Furthermore, anomaly detection using reconstruction error by PCA makes it easier to detect characteristic pattern changes in the signal, enabling more sensitive anomaly detection. [Explanation of symbols]

[0086] 1. Anomaly detection system 10 Electromagnetic wave radar 20 Radar signal processing unit 22 Range Doppler images 30. Time-series signal generation unit 31 grid areas 40 Feature Extraction Unit 50 Anomaly detection unit 60 Learning Department

Claims

1. An anomaly detection system using electromagnetic wave radar, A radar signal processing unit that applies range FFT and Doppler FFT to the received IQ signal to generate a range Doppler image, A time-series signal generation unit divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time-series waveform data. A feature extraction unit that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing, The system includes an anomaly detection unit that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the aforementioned feature quantities. An anomaly detection system characterized by the following features.

2. The feature extraction unit includes a process that uses an Echo State Network (ESN) and applies a nonlinear reservoir. The anomaly detection system according to feature 1.

3. The feature extraction unit, By applying an autoregressive neural network (AR-NN) that takes into account the autoregressive nature of time-series data, A process that dynamically updates the state based on autoregressive properties derived from time-series data up to the previous time point, The process involves applying a nonlinear activation function (Rectified Linear Unit: ReLU, sigmoid function, or hyperbolic tangent function) to the output of the autoregressive neural network, An anomaly detection system according to claim 1, characterized by comprising:

4. The anomaly detection unit performs anomaly detection for each grid. An anomaly detection system according to any one of claims 1 to 3.

5. The anomaly detection unit performs anomaly detection based on the variance in the principal component space obtained by the principal component analysis. An anomaly detection system according to any one of claims 1 to 3.

6. The aforementioned electromagnetic wave radar is an FMCW type millimeter-wave radar. The radar signal processing unit generates fan-shaped data. An anomaly detection system according to any one of claims 1 to 3.

7. The feature extraction unit extracts features from the waveform data using a filter bank consisting of filters of multiple frequency bands. The system further comprises a learning unit that performs learning for anomaly detection using the time-series patterns of the aforementioned features. An anomaly detection system according to any one of claims 1 to 3.

8. Computers, Radar signal processing means that applies range FFT and Doppler FFT to the received IQ signal to generate a range Doppler image. A time-series signal generation means that divides the range Doppler image into a grid and applies an inverse Fourier transform (iFFT) to the signal in each grid region to generate time-series waveform data. A feature extraction means that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing. An anomaly detection means for detecting anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the aforementioned feature quantities. A program designed to function as such.

9. Equipped with a processor and memory, The aforementioned processor Radar signal processing that applies range FFT and Doppler FFT to the received IQ signal to generate a range Doppler image, Time-series signal generation involves dividing the range Doppler image into a grid and applying an inverse Fourier transform (iFFT) to the signal in each grid region to generate time-series waveform data. Feature extraction involves applying filter bank processing and extracting features using a dynamic time series model that includes reservoir computing. Anomaly detection is performed by detecting anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the aforementioned feature quantities. An anomaly detection device characterized by performing the following actions.

10. Range FFT and Doppler FFT are applied to the received IQ signal. A radar signal processing process that generates a range Doppler image, A time-series signal generation process involves dividing the range Doppler image into a grid and applying an inverse Fourier transform (iFFT) to the signal in each grid region to generate time-series waveform data. A feature extraction process that applies filter bank processing and extracts features using a dynamic time series model including reservoir computing, An anomaly detection step that detects anomalies based on the reconstruction error obtained by applying principal component analysis (PCA) to the aforementioned feature quantities, An anomaly detection method including

Images