KR102628542B1 - Artificial intelligence-based marine rescue signal automatic identification system and method - Google Patents

Abstract

The artificial intelligence-based automatic identification system and method for marine rescue signals can quickly and accurately notify rescue organizations of rescue signals through artificial intelligence analysis using radio wave classification without the complicated process of using hearing in the internal structure of the distress signal receiver. In an embodiment, the audibility of the rescue signal can be maximized by transforming the electrical signal input to the receiver into a spectrogram to remove noise. In addition, in the embodiment, the procedures required for the rescue initiation process, from generating a rescue signal to recognizing the rescue signal from a rescue organization, are reduced, thereby enabling rapid rescue activities in the event of a disaster. In addition, in the embodiment, through the installation of a receiver, multiple channels can be listened to simultaneously, so that rescue signals that may occur at various frequencies and channels can be transmitted to rescue organizations without being missed.

Description

Artificial intelligence-based marine rescue signal automatic identification system and method {ARTIFICIAL INTELLIGENCE-BASED MARINE RESCUE SIGNAL AUTOMATIC IDENTIFICATION SYSTEM AND METHOD}

This disclosure relates to an artificial intelligence-based automatic identification system and method for marine rescue signals. Specifically, when a rescue signal is input to a receiver, it relates to a system and method for automatically identifying the rescue signal using radio wave classification.

Unless otherwise indicated herein, the material described in this section is not prior art to the claims of this application, and is not admitted to be prior art by inclusion in this section.

It is most important in the rescue process for rescue organizations to quickly recognize marine rescue signals. As one of the international community's efforts to listen to rescue signals, the International Maritime Organization (IMO) under the UN designated VHF 16 as a distress safety listening channel through the International Telecommunication Union, and each rescue organization around the world listens to rescue signals generated at sea 24/7. It is mandatory. In Korea, the Coast Guard listens to rescue signals 24 hours a day.

Referring to Figure 1, which shows a conventional distress and rescue system, conventionally, when a ship in distress occurs in the ocean, a rescue signal is transmitted through a transmitter provided on the ship and the rescue signal is transmitted through surrounding ships. Afterwards, the Coast Guard recognizes the rescue signal and begins rescue operations. The rescue signal is the first signal that initiates rescue activities, and listening to the rescue signal properly is very related to securing the golden time for rescue.

As shown in Figure 2, which shows ships and marine traffic on a map, there are over 100,000 registered vessels near Korea, and since it is the place with the largest maritime traffic volume in the world, including neighboring China and Japan, various frequencies are used. and channels exist. However, the current method of listening to rescue signals allows one person to listen to only one channel. Additionally, there are limitations to listening to analog rescue signals, as only three people per day need to monitor the entire sea area. In addition, in an environment where listening is performed 24 hours a day, 365 days a year without a break, you may be exposed to blind spots in rescue signals due to decreased concentration due to accumulated fatigue, and wireless communication may have problems with reception distance and/or problems due to bad weather due to natural disasters or aging of equipment and speakers. There is a problem with the method of listening to rescue signals, which drastically reduces sensitivity.

Referring to Figure 3, which shows the status of distress communication reception and processing at the International Safety Communication Center, it can be seen that the number of unconfirmed cases is 5-7 times higher than the number of actual distress calls in the first half of 2020, 21, and 22. If the number of unconfirmed confirmations included actual rescue signals, this could lead to a safety accident.

1. Korean Patent Registration No. 10-1998521 (2019.07.03) 2. Korean Patent Registration No. 10-1764816 (2017.07.28)

The artificial intelligence-based marine rescue signal automatic identification system and method according to the embodiment automatically identifies the rescue signal using radio wave classification when the rescue signal is input to the receiver. In the embodiment, voice rescue signal keywords such as mayday, please help me, fire, and sinking are collected, and the voice rescue signal is data processed into wavelengths at the electrical signal stage. Afterwards, when a rescue signal comes into the receiver using an artificial intelligence algorithm, the rescue signal is recognized at the electrical signal stage and notified to the rescue organization.

Additionally, in the embodiment, the noise removal (denoising) effect is increased through a spectrogram, and noise can be removed by transforming the electrical signal input to the receiver into a spectrogram.

An artificial intelligence-based marine rescue signal automatic identification system according to an embodiment includes: a rescue signal transmission device that converts the rescue signal into an electronic signal when a rescue signal including the voice of a rescue requester is input, and transmits the converted electronic signal; and an automatic rescue signal identification device that receives the electronic signal transmitted from the rescue signal transmission device, converts it into an electric signal, and visualizes the electric signal in the form of vision and sound through a spectogram. Includes.

In an embodiment, a rescue signal automatic identification device; Collects voice rescue signal keywords including mayday, please help me, fire, and sinking, processes the collected voice rescue signals into wavelength data at the electrical signal stage, archives the processed wavelength data, and sends the rescue signal to the receiver using an artificial intelligence algorithm. When input, the rescue signal is recognized at the electrical signal stage through the archived wavelength data and notified to the rescue agency.

In an embodiment, a rescue signal automatic identification device; Identify analysis elements including time, frequency, and amplitude of the converted electrical signal, create a spectrogram visualizing the identified analysis elements as a graph, and divide the generated spectogram into regular time intervals. preprocessing module; Deep classifies input data containing segmented spectogram images at regular time intervals, calculates the accuracy of the classification value as a probability, selects keywords used when requesting rescue, and learns to recognize the selected keywords. running module; and a post-processing module that extracts and outputs display objects and notification sounds according to the calculated probability value. Includes.

The artificial intelligence-based marine rescue signal automatic identification system and method as described above can quickly and accurately report rescue signals to rescue organizations through artificial intelligence analysis using radio wave classification without the complicated procedure of using hearing in the internal structure of the distress signal receiver. You can.

In an embodiment, the audibility of the rescue signal can be maximized by transforming the electrical signal input to the receiver into a spectrogram to remove noise.

In addition, in the embodiment, the procedures required for the rescue initiation process, from generating a rescue signal to recognizing the rescue signal from a rescue organization, are reduced, thereby enabling rapid rescue activities in the event of a disaster.

In addition, in the embodiment, through the installation of a receiver, multiple channels can be listened to simultaneously, so that rescue signals that may occur at various frequencies and channels can be transmitted to rescue organizations without being missed.

The effects of the present invention are not limited to the effects described above, and should be understood to include all effects that can be inferred from the configuration of the invention described in the detailed description or claims of the present invention.

Figure 1 is a diagram showing a conventional distress and rescue system.
Figure 2 is a map showing ships and maritime traffic
Figure 3 is a diagram showing the status of distress communication reception and processing at the International Safety Communication Center
Figure 4 is a diagram showing the configuration of an artificial intelligence-based marine rescue signal automatic identification system according to an embodiment.
Figure 5 is a diagram for comparing the automatic identification device for rescue signals according to an embodiment with the prior art.
Figure 6 is a diagram showing the data processing configuration of an artificial intelligence-based automatic marine rescue signal identification device according to an embodiment.
Figure 7 is a diagram showing a spectogram according to an embodiment
Figure 8 is a diagram for explaining the data archiving process according to an embodiment
Figure 9 is a diagram showing the signal flow of an artificial intelligence-based marine rescue signal automatic identification system according to an embodiment.
Figure 10 is a diagram showing the data processing process of the automatic rescue signal identification device according to an embodiment
Figure 11 is a diagram showing the data processing flow of automatic identification of artificial intelligence-based marine rescue signals according to an embodiment
Figure 12 is a diagram showing wavelength data generated by an automatic rescue signal identification device in an embodiment

The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. The present embodiments are merely provided to ensure that the disclosure of the present invention is complete and to provide common knowledge in the technical field to which the present invention pertains. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

In describing embodiments of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the gist of the present invention, the detailed description will be omitted. The terms described below are terms defined in consideration of functions in embodiments of the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification.

Figure 4 is a diagram showing the configuration of an artificial intelligence-based marine rescue signal automatic identification system according to an embodiment. Referring to FIG. 4, an artificial intelligence-based marine rescue signal automatic identification system according to an embodiment may be configured to include a rescue signal automatic identification device 200 and a rescue signal transmission device 100.

The rescue signal transmission device 100 is a device that collects signals from rescue seekers in need of emergency rescue. When a rescue signal including the rescue seeker's voice is input, the rescue signal transmission device 100 converts the rescue signal into an electronic signal, and converts the converted electronic signal into a rescue signal. Transmitted to the automatic identification device 200.

The rescue signal automatic identification device 200 receives the electronic signal transmitted from the rescue signal transmission device 100, converts it into an electric signal, and visualizes the electric signal in the form of vision and sound. Here, the electrical signal may be amplified between each stage of the transmission process. In an embodiment, the automatic rescue signal identification device 200 collects voice rescue signal keywords such as mayday, please help me, fire, and sinking, and processes the collected voice rescue signals into wavelength data at the electrical signal stage. In the embodiment, when a rescue signal is input to the receiver by archiving the processed wavelength data, the artificial intelligence algorithm recognizes the rescue signal at the electrical signal stage and informs the rescue organization. In an embodiment, the saved wavelength data is compared with the wavelength data of the rescue signal, and the rescue signal can be recognized according to the comparison result.

Figure 5 is a diagram for comparing an automatic rescue signal identification device according to an embodiment with the prior art. Figure 5(a) is a diagram showing a conventional distress signal identification device, and Figure 5(b) is a diagram showing a distress signal identification device according to an embodiment.

Referring to FIG. 5, the automatic rescue signal identification device according to the embodiment analyzes radio signals through artificial intelligence without frequency amplification or demodulation to quickly and accurately identify rescue signals and inform rescue organizations of them.

The artificial intelligence-based marine rescue signal automatic identification device according to the embodiment automatically identifies the rescue signal using radio wave classification when the rescue signal is input to the receiver. In the embodiment, voice rescue signal keywords such as mayday, please help me, fire, and sinking are collected, and the voice rescue signal is data processed into wavelengths at the electrical signal stage. Afterwards, when a rescue signal comes into the receiver using an artificial intelligence algorithm, the rescue signal is recognized at the electrical signal stage and notified to the rescue organization. Additionally, in the embodiment, the noise denoising effect is increased through a spectrogram, and the electrical signal input to the receiver is transformed into a spectogram to remove noise.

Figure 6 is a diagram showing the data processing configuration of an artificial intelligence-based automatic marine rescue signal identification device according to an embodiment.

Referring to FIG. 6, the automatic rescue signal identification device 200 according to the embodiment may be configured to include a pre-processing module 210, a deep learning module 220, and a post-processing module 230. The term 'module' used in this specification should be interpreted to include software, hardware, or a combination thereof, depending on the context in which the term is used. For example, software may be machine language, firmware, embedded code, and application software. As another example, hardware may be a circuit, processor, computer, integrated circuit, integrated circuit core, sensor, Micro-Electro-Mechanical System (MEMS), passive device, or a combination thereof.

In an embodiment, the preprocessing module 210 collects voice rescue signal keywords including mayday, please help me, fire, and sinking, and processes the collected voice rescue signal into wavelength data at the electrical signal stage. In the case of marine accidents, they are often caused by bad weather, etc., and at this time, various noises such as waves, wind, and tempest are involved in the rescue signal of the rescuer. Therefore, pre-filtering of these rescue signals will have a significant impact on the automatic recognition rate of the rescue signal. You can. Through pre-filtering, the recognition rate can be improved when a rescue signal is input to the rescue signal recognition stage. Afterwards, the deep learning module 220 archives the processed wavelength data. Archiving refers to storing necessary records in a file format on a storage medium for a specific period of time. As shown in FIG. 8, in the embodiment, when a rescue signal is input to the receiver using an artificial intelligence algorithm, the electrical signal stage is generated through the stored wavelength data. recognizes rescue signals and informs rescue organizations. Figure 12 shows wavelength data generated by the automatic rescue signal identification device 200 in an embodiment. In the embodiment, the voice rescue signal is processed into wavelength data at the electrical signal stage. In an embodiment, the saved wavelength data is compared with the wavelength data of the rescue signal, and the rescue signal can be recognized according to the comparison result. In an embodiment, the wavelength coincidence rate between the stored wavelength data and the rescue signal is converted into a rescue signal occurrence probability value, and a warning sound or visual object is output according to the calculated probability value to notify the rescue organization of the rescue signal.

In addition, the preprocessing module 210 identifies analysis elements including time, frequency, and amplitude of the converted electrical signal, generates a spectrogram visualizing the identified analysis elements in a graph, and generates a spectrogram that visualizes the identified analysis elements in a graph. Grams are divided into regular time intervals. A spectogram visualizes the spectrum of sound and represents it as a graph, one of the visualization objects. FIG. 7 is a diagram showing a spectogram according to an embodiment. Referring to FIG. 7, the x-axis of the spectogram represents time, the y-axis represents frequency, and the z-axis represents amplitude.

In an embodiment, the preprocessing module 210 changes the voice data from the time domain to the frequency domain through Fourier Transform (FT), performs a denoising process, and then performs a denoising process. Enter into the deep learning module.

Specifically, the preprocessing module 210 divides the data before being converted to the frequency domain according to the specified window size, transforms the divided data through DFT (Discrete Fourier Transform), and converts the converted information. is connected and input into a deep learning network (DL network) based on an encoder-decoder network to perform denoising. Additionally, in the embodiment, the preprocessing module 210 may input a raw wave form of an electrical signal. In an embodiment, data augmentation may be performed by combining clean speech and noise. In an embodiment, denoising can be performed on the amplified data through an encoder-decoder network based on U-net with Long Short-Term Memory (LSTM).

The deep learning module 220 classifies input data including segmented spectogram images at regular time intervals, calculates the accuracy of the classification value as a probability, and selects keywords used when requesting rescue. Learn to recognize keywords. In an embodiment, a rescue signal recognition model can be implemented by performing machine learning based on a deep learning neural network based on a customized training data set optimized for identifying rescue signal keyword information. In an embodiment, the deep learning module 220 trains a deep learning neural network including at least one of a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and a Bidirectional Recurrent Deep Neural Network (BRDNN). By training with a data set, a keyword recognition model and accuracy calculation model can be implemented.

Additionally, in the embodiment, the deep learning module 220 may classify image input data using image recognition through artificial intelligence machine learning in the process of classifying divided spectogram images at regular time intervals. Artificial intelligence image recognition is a type of computer vision technology in which a machine recognizes objects and understands scenes from photos just like a human. In an embodiment, for image recognition, a data processing process may be performed to classify and detect objects included in the image, and to identify and segment the objects in pixel units. Additionally, in the embodiment, the deep learning module 220 performs an out of distribution detection process other than learning to process unlearned patterns in addition to noise response. Non-learning distribution data detection is to identify whether the image input to artificial intelligence is learned probability distribution data. In the embodiment, stability and reliability can be increased by filtering out or exception processing images that are difficult for the artificial neural network to judge through detection of distribution data other than learning. In the embodiment, in order to detect non-learning distribution data, a probability value indicating how confident one is in the deep learning decision is calibrated or non-learning distribution data is generated and learned using a generative adversarial network (GAN). This helps improve detection accuracy.

Additionally, in the embodiment, in order to reduce the size of the model while maintaining image recognition accuracy, image input values can be classified using lightweight deep learning technology that simplifies computation. In the embodiment, for image recognition, a convolutional filter is modified in a convolutional neural network (CNN) to reduce the operation dimension, or pruning and weight values to delete weights of a neural network that do not have a significant impact are performed. A quantization process is performed to simplify calculations by reducing the number of floating point numbers, enabling data lightweighting. Additionally, in the embodiment, the output of a pre-trained large neural network is imitated and learned by a small neural network to simplify computation and maintain accuracy.

The post-processing module 230 extracts and outputs display objects and notification sounds according to the calculated probability value. In an embodiment, the post-processing module 230 may set a plurality of thresholds according to the calculated probability value and create a range in which the display object and the notification sound match with the set threshold. In the embodiment, if the probability value is calculated to be less than 0 percent, which is the first threshold, there is no response, and if the probability value is calculated to be more than 50 percent, which is the first threshold, but less than 60 percent, which is the second threshold, blue is output, and If it is more than 60 percent, which is the second threshold, but less than 70 percent, which is the third threshold, green is output. If the calculated probability value is more than 70 percent, which is the third threshold, but less than 80 percent, which is the fourth threshold, yellow is output, and if the calculated probability value is more than 80 percent, which is the fourth threshold, but less than 90 percent, which is the fifth threshold, orange is output. there is. If the calculated probability value exceeds 90 percent, which is the fifth threshold, in the embodiment, a red color may be output, and if the calculated probability value exceeds a certain probability (e.g., 80 percent), an alarm sound may be output. Since the probability value calculated in the embodiment may not be calculated linearly, the first to fifth thresholds that determine the matching range with the audiovisual object may be set differently depending on the rescue signal. Depending on the threshold, this post-processing helps to catch signals that need to be brought to the attention of the listening worker among unidentifiable (unknown) signals without increasing the fatigue of the rescue signal listening worker. These thresholds are not set arbitrarily, but as shown in Figure 3, when the ratio of 'unidentifiable signals' to 'actual distress' in a certain period increases, each threshold is lowered to draw the attention of listening workers even at slightly lower probability values. It can be set to do so. In this case, the occurrence of a rescue signal will be communicated to the listening worker at a slightly higher frequency. Through this post-processing, automatic recognition of artificial intelligence marine rescue signals through computers and foster recognition through listening workers can form a collaborative relationship.

Below, we will sequentially explain the artificial intelligence-based automatic identification method for marine rescue signals. Since the operation (function) of the artificial intelligence-based marine rescue signal automatic identification method according to the embodiment is essentially the same as the function of the artificial intelligence-based marine rescue signal automatic identification system, descriptions overlapping with FIGS. 4 to 8 will be omitted.

Figure 9 is a diagram showing the signal flow of an artificial intelligence-based marine rescue signal automatic identification system according to an embodiment.

Referring to FIG. 9, in step S100, when a rescue signal including the voice of a rescue requester is input from the rescue signal transmission device, the rescue signal is converted into an electronic signal in step S200, and the converted electronic signal is transmitted to an automatic rescue signal identification device. send. In step S300, the automatic rescue signal identification device receives the electronic signal transmitted from the rescue signal transmission device, and in step S400, it can be converted to an electric signal and amplified. In the S500 stage, the electrical signal (if amplified, the amplified electrical signal) is visualized in the form of vision and sound through a spectogram.

Figure 10 is a diagram showing the data processing process of the automatic rescue signal identification device according to an embodiment.

Referring to Figure 10, in step S910, voice rescue signal keywords including mayday, please help me, fire, and sinking are collected in the preprocessing module, and in step S920, the collected voice rescue signal is processed into wavelength data in the electrical signal step. In the S930 step, the wavelength data processed by the deep learning module is archived, and in the S940 step, when a rescue signal is input to the receiver, the electrical signal is generated through an artificial intelligence algorithm that recognizes the rescue signal by comparing the archived wavelength data and the input rescue signal. Recognizes a rescue signal. In step S950, the recognized rescue signal is transmitted to the rescue organization.

Figure 11 is a diagram showing the data processing flow of artificial intelligence-based automatic identification of marine rescue signals according to an embodiment.

Referring to FIG. 11, in step S941, analysis elements including time, frequency, and amplitude of the converted electrical signal are identified, and in step S942, a spectrogram visualizing the identified analysis elements in a graph is generated. In step S943, the generated spectogram is divided into regular time intervals. In step S944, input data including segmented spectogram images at regular time intervals are classified, and the accuracy of the classification value is calculated as a probability. In step S945, keywords used when requesting rescue are selected and learned to recognize the selected keywords. In step S946, the display object and notification sound are extracted and output according to the calculated probability value.

In the embodiment, in step S942, analysis elements including time, frequency, and amplitude of the converted electrical signal are identified, a spectrogram visualizing the identified analysis elements as a graph is generated, and the generated spectogram is Divide at regular time intervals. A spectogram visualizes the spectrum of sound and represents it as a graph, one of the visualization objects. FIG. 7 is a diagram showing a spectogram according to an embodiment. Referring to FIG. 7, the x-axis of the spectogram represents time, the y-axis represents frequency, and the z-axis represents amplitude.

In step S944, the voice data is changed from the time domain to the frequency domain through Fourier Transform (FT), a denoising process is performed, and then input into the deep learning module. You can.

Specifically, in the embodiment, the data before being converted to the frequency domain is divided according to the specified window size, the divided data is transformed through DFT (Discrete Fourier Transform), and the transformed information is connected. Denoising is performed by inputting it into a deep learning network (DL network) based on an encoder-decoder network. Additionally, in the embodiment, the raw wave form of the electrical signal may be input. In an embodiment, data augmentation may be performed by combining clean speech and noise. In an embodiment, denoising can be performed on the amplified data through an encoder-decoder network based on U-net with Long Short-Term Memory (LSTM).

In step S945, input data including segmented spectogram images at regular time intervals are classified, the accuracy of the classification value is calculated as a probability, and keywords used when requesting rescue are selected to recognize the selected keywords. learn In an embodiment, a rescue signal recognition model can be implemented by performing machine learning based on a deep learning neural network based on a customized training data set optimized for identifying rescue signal keyword information. In an embodiment, the deep learning module 220 trains a deep learning neural network including at least one of a Deep Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and a Bidirectional Recurrent Deep Neural Network (BRDNN). By training with a data set, a keyword recognition model and accuracy calculation model can be implemented.

In step S946, a plurality of threshold values are set according to the calculated probability value, and a range in which the display object and the notification sound are matched can be created using the set threshold values. In the embodiment, if the probability value is calculated to be less than 0 percent, which is the first threshold, there is no response, and if the probability value is calculated to be more than 50 percent, which is the first threshold, but less than 60 percent, which is the second threshold, blue is output, and If it is more than 60 percent, which is the second threshold, but less than 70 percent, which is the third threshold, green is output. If the calculated probability value is more than 70 percent, which is the third threshold, but less than 80 percent, which is the fourth threshold, yellow is output, and if the calculated probability value is more than 80 percent, which is the fourth threshold, but less than 90 percent, which is the fifth threshold, orange is output. there is. If the calculated probability value exceeds 90 percent, which is the fifth threshold, in the embodiment, a red color may be output, and if the calculated probability value exceeds a certain probability (e.g., 80 percent), an alarm sound may be output. Since the probability value calculated in the embodiment may not be calculated linearly, the first to fifth thresholds that determine the matching range with the audiovisual object may be set differently depending on the rescue signal.

The artificial intelligence-based marine rescue signal automatic identification system and method as described above can quickly and accurately report rescue signals to rescue organizations through artificial intelligence analysis using radio wave classification without the complicated procedure of using hearing in the internal structure of the distress signal receiver. make it possible

In an embodiment, the audibility of the rescue signal can be maximized by transforming the electrical signal input to the receiver into a spectrogram to remove noise.

In addition, in the embodiment, the procedures required for the rescue initiation process, from generating a rescue signal to recognizing the rescue signal from a rescue organization, are reduced, thereby enabling rapid rescue activities in the event of a disaster.

In addition, in the embodiment, through the installation of a receiver, multiple channels can be listened to simultaneously, so that rescue signals that may occur at various frequencies and channels can be transmitted to rescue organizations without being missed.

The disclosed content is merely an example, and various modifications and implementations may be made by those skilled in the art without departing from the gist of the claims, so the scope of protection of the disclosed content is limited to the above-mentioned specific scope. It is not limited to the examples.

Claims (10)

In the artificial intelligence-based automatic identification system for marine rescue signals,
When a rescue signal including the voice of a rescue requester is input, a rescue signal transmission device that converts the rescue signal into an electronic signal and transmits the converted electronic signal; and
An automatic rescue signal identification device that receives the electronic signal transmitted from the rescue signal transmission device, converts it into an electric signal, and visualizes the electric signal in the form of vision and sound through a spectogram; includes
The rescue signal automatic identification device; Is
We collect voice rescue signal keywords including mayday, please help me, fire, and sinking, and process the collected voice rescue signals into wavelength data at the electrical signal stage.
By archiving the processed wavelength data, when a rescue signal is input to the receiver using an artificial intelligence algorithm, the rescue signal is recognized at the electrical signal stage through the archived wavelength data and notified to the rescue agency.
The rescue signal automatic identification device; Is
Identify analysis elements including time, frequency, and amplitude of the converted electrical signal, create a spectrogram visualizing the identified analysis elements in a graph, and divide the generated spectogram into regular time intervals. Preprocessing module;
Classifies input data including the divided spectogram images at regular time intervals, calculates the accuracy of the classification value as a probability, selects keywords used when requesting rescue, and learns to recognize the selected keywords. Deep learning module; and
A post-processing module that extracts and outputs display objects and notification sounds according to the calculated probability value; Including,
The post-processing module; silver
Set a plurality of thresholds according to the calculated probability value, create a range in which the display object and notification sound match with the set threshold, and
The preprocessing module; silver
Voice data is changed from the time domain to the frequency domain through Fourier Transform (FT), a denoising process is performed, and then input into the deep learning module.
The data before being converted to the frequency domain is divided according to the specified window size, the divided data is converted through DFT (Discrete Fourier Transform), and the converted information is connected to create an encoder-decoder network. Denoising is performed by inputting it into a deep learning network (DL network) based on -Decoder network.
After data augmentation by combining clean speech and noise, the amplified data is connected to an encoder-decoder network based on U-net with LSTM (Long Short-Term Memory). -Decoder network) performs noise removal (denoising),
The deep learning module; silver
A keyword recognition model that classifies input data containing segmented spectogram images at regular time intervals, calculates the accuracy of the classification value as a probability, and selects keywords used when requesting rescue to recognize the selected keywords. and train an accuracy calculation model,
A rescue signal recognition model is implemented by performing machine learning based on a deep learning neural network based on a customized training data set optimized for identifying rescue signal keyword information.
In the process of classifying segmented spectogram images at regular time intervals, image input data is classified through image recognition using artificial intelligence machine learning,
The rescue signal automatic identification device; Is
By comparing the saved wavelength data and the rescue signal wavelength data, the rescue signal is recognized according to the comparison result, and the wavelength matching rate between the saved wavelength data and the rescue signal is converted into a rescue signal occurrence probability value, and a notification sound or visual object is generated according to the calculated probability value. An artificial intelligence-based automatic marine rescue signal identification system that outputs and transmits rescue signals to rescue organizations.
delete delete delete delete In the artificial intelligence-based automatic identification method of marine rescue signals,
(A) When a rescue signal including the voice of a rescue requester is input from the rescue signal transmission device, converting the rescue signal into an electronic signal and transmitting the converted electronic signal;
(B) receiving the electronic signal transmitted from the rescue signal transmission device in an automatic rescue signal identification device, converting it into an electric signal, and visualizing the electric signal in the form of vision and sound through a spectogram; includes
Step (B) above; Is
(B-1) Collecting voice rescue signal keywords including mayday, please help me, fire, and sinking, and processing the collected voice rescue signals into wavelength data at the electrical signal stage;
(B-2) archiving the processed wavelength data and, when the rescue signal is input to the receiver using an artificial intelligence algorithm, recognizing the rescue signal from the electrical signal through the archived wavelength data; and
(B-3) transmitting the recognized rescue signal to the rescue organization; includes
Step (B-1) above; Is
Identify analysis elements including time, frequency, and amplitude of the converted electrical signal, create a spectrogram visualizing the identified analysis elements in a graph, and divide the generated spectogram into regular time intervals. step; Including,
Step (B-2) above; Is
Classifies input data including the divided spectogram images at regular time intervals, calculates the accuracy of the classification value as a probability, selects keywords used when requesting rescue, and learns to recognize the selected keywords. step; and
Extracting and outputting a display object and a notification sound according to the calculated probability value; includes
Extracting and outputting a display object and a notification sound according to the calculated probability value; Is
Set a plurality of thresholds according to the calculated probability value, create a range in which the display object and notification sound match with the set threshold, and
Identify analysis elements including time, frequency, and amplitude of the converted electrical signal, generate a spectrogram visualizing the identified analysis elements as a graph, and divide the generated spectogram into regular time intervals. steps; Is
Voice data is changed from the time domain to the frequency domain through Fourier Transform (FT), and input data is generated by performing a denoising process.
Step (B-1) above; Is
The data before being converted to the frequency domain is divided according to the specified window size, the divided data is converted through DFT (Discrete Fourier Transform), and the converted information is connected to create an encoder-decoder network. Denoising is performed by inputting it into a deep learning network (DL network) based on -Decoder network.
After data augmentation by combining clean speech and noise, the amplified data is connected to an encoder-decoder network based on U-net with LSTM (Long Short-Term Memory). -Decoder network) performs noise removal (denoising),
Step (B-2) above; Is
A rescue signal recognition model is implemented by performing machine learning based on a deep learning neural network based on a customized training data set optimized for identifying rescue signal keyword information.
In the process of classifying segmented spectogram images at regular time intervals, image input data is classified through image recognition using artificial intelligence machine learning,
Step (B-3) above; Is
By comparing the saved wavelength data and the rescue signal wavelength data, the rescue signal is recognized according to the comparison result, and the wavelength matching rate between the saved wavelength data and the rescue signal is converted into a rescue signal occurrence probability value, and a notification sound or visual object is generated according to the calculated probability value. An artificial intelligence-based automatic identification method for marine rescue signals that outputs and transmits rescue signals to rescue organizations.
delete delete delete delete