A method for achieving faster end-to-end data transmission over a voice channel

Abstract

This invention provides a method for achieving faster end-to-end data transmission via a voice channel. It designs a modulation and demodulation scheme based on chirp signals and deep learning, solving the impact of voice activity detection technology on data transmission and significantly improving data throughput. The method includes: the sender generating a random data certificate and modulating the data-signal using a chirp-based modulation scheme; the modulated chirp signal being transmitted in the voice channel; the receiver demodulating the received signal using a deep learning-based demodulation scheme; the receiver verifying the received data using error correction codes; and both parties using a dedicated data link protocol based on stop / resume mechanisms, time synchronization, and retransmission schemes between the caller and the callee to ensure complete data reception by the receiver.

Description

Technical Field

[0001] This invention relates to the field of data transmission via voice channels, and more particularly to a method for achieving faster end-to-end data transmission via voice channels. Background Technology

[0002] To prevent phone fraud, end-to-end authentication of callers must be provided over traditional (untrusted) telephone networks. Similar to internet websites, Secure Sockets Layer (SSL) certificates ensure the authenticity of each website's identity.

[0003] However, aside from the caller's identifier (ID), modern telephone infrastructure provides no way for the called party to infer the caller's identity. Therefore, it is necessary for the caller to be able to transmit its digital certificate to the called party for authentication. This transmission should be end-to-end, requiring no assistance from the telephone provider, compatible with existing infrastructure, and not reliant on mobile data such as 4G / 5G.

[0004] One possible solution is to use a dial-up modem, available for decades, to transmit data over a telephone line. However, this method is not suitable for mobile phones. This is because the baseband in a smartphone, which connects to a cellular network (2G, 3G, 4G, etc.) and helps convert digital data into radio frequency signals (and vice versa), is a black box to the end user. Without a smartphone vendor or network provider, it is virtually impossible for a user to implement their own data modem on their smartphone. While mobile data offers an alternative solution for transmitting data over a cellular network, it incurs additional financial costs. Research by the Global System for Mobile Communications Association (GSMA) shows that despite living in areas with mobile data coverage, 3.4 billion people still cannot afford mobile internet, making this not a universal solution.

[0005] If a data modem can be established on a cellular network, custom encryption algorithms can be used to encrypt data, thereby improving data security. To address these issues, some academic research has proposed methods for data transmission over voice channels in cellular networks, including transmitting data over unknown voice channels using frequency shift keying (FPS) coding; transmitting voice using a single codebook, incorporating an efficient low-bit-rate voice encoder; a voice compression technique relying on Linear Predictive Coding (LPC); data over voice (DoV) based on short harmonic waveform codebooks; and a strong encrypted authentication protocol inspired by Transport Layer Security (TLS) 1.2 to determine the identity of the entity on the other end of the call (i.e., the caller ID), etc.

[0006] However, according to experiments, their work mostly fails to achieve the data rates claimed in cellular networks. This is generally due to the following reasons: 1) complex network infrastructure distorts signals transmitted from one subsystem to another; 2) some optimization techniques, such as Voice Activity Detection (VAD), tend to reject non-speech-type decoded frames; and 3) signals unlike speech will be severely distorted by the codec. Summary of the Invention

[0007] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for achieving faster end-to-end data transmission through a voice channel, addressing the shortcomings of the prior art.

[0008] To combat signal distortion in complex network infrastructures, this invention proposes a chirp-based modulation / demodulation scheme, as chirp signals have been proven robust to channel noise. To reduce the demodulation error rate, deep learning (DL) technology is used to decode the distorted linear frequency modulated signal. To avoid the effects of voice-over-dampening (VAD), a stop / resume mechanism is proposed, which inserts gaps into the signal. To ensure the integrity and reliability of the receiver's data, a dedicated data link protocol with time synchronization and retransmission schemes between the caller and the callee is proposed.

[0009] The method of the present invention includes the following steps:

[0010] Step 1, Data Modulation: The sender generates a random certificate or uses an existing certificate, and uses a data link protocol based on stop and resume mechanisms and time synchronization, and uses a chirp-based modulation method to modulate the data certificate into an analog signal, which is then transmitted through the voice channel;

[0011] Step 2, Data Demodulation: The receiver receives the signal and uses a deep learning-based demodulation method to demodulate the analog signal into a data certificate;

[0012] Step 3, Error Recovery and Retransmission: The receiver corrects the received data certificate based on the error correction code. If the certificate cannot be completely restored, it sends a retransmission request to the sender. The sender receives the retransmission request and retransmits the certificate.

[0013] Step 1 includes:

[0014] Step 1-1: The sender adjusts the data signal according to the data link protocol with stop, resume mechanism and time synchronization, so that the receiver can determine the exact position of the data in the audio stream (each audio stream contains N data frames). Each data frame includes N chirp signals. The N chirp signals are divided into two or more chirp signal groups. Each chirp signal group is called a symbol group (Data is modulated into an audio stream, which contains several data frames. The audio stream is the data stream).

[0015] Symbol groups will be separated by gaps, and a unique chirp signal will be added at the beginning and end of each data frame as a separator to indicate when the data frame begins and ends.

[0016] Steps 1-2: Modulate the data certificate based on the chirp signal.

[0017] In step 1-1, to detect the exact location of the delimiter at the receiver, the data frame employs a cross-correlation-based method, where the known delimiter signal is associated with the received audio stream within a sliding window: Assume the received audio stream comprises N audio sample points (audio sample points are the time-discretionary representation of a continuous signal; the number of sample points is calculated based on the signal sampling rate and the audio stream duration). Based on the audio sample points, the entire audio stream is represented as {u... i}, i = 1, 2, ..., n; the separator sent by the sender is represented as {v} based on its sampling points. i}, i = 1, 2, ..., m; {u i} and {v i In the}, i represents the i-th sample point of the entire received audio stream and the sender's separator, and n >> m; n takes the value of audio stream duration * signal sampling rate; m takes the value of separator duration * signal sampling rate;

[0018] Use matched filtering from {u i Extract a sliding window of length m from the sample data, and calculate the sample correlation coefficient r using the following formula:

[0019]

[0020] in It is the average of the samples from the sliding window. The meaning is {v i}

[0021] In step 1-1, in order to speed up the calculation time, the approximate position c of the separator is first located by calculating the sample correlation coefficient window by window. Within the range of [cm, c+m], multiple threads use fine-grained correlation with a sliding window size of 1 to locate the accurate position in parallel.

[0022] In step 1-1, the following delimiter position adjustment scheme is formulated: each data frame contains k audio sampling points, and the two delimiters surrounding the data frame are located at sampling point indices d1 and d2. The indices are determined by locating the starting positions of the two delimiters based on the sampling point positions d1 and d5. d2-d1-mk=δ, where m is the length of the delimiter, and δ is the difference between the number of sampling points of the receiver's data frame obtained by the receiver based on d1 and d2 and the number of sampling points of the sender's data frame.

[0023] Adjust d1 and d2 to compensate for δ: First, set some chirp signals in the data frame (e.g., the first 30 chirp signals in each symbol group) to fixed values. In each possible receiver data frame of {(d1, d2-δ), ..., (d1+i, d2-δ+i), ..., (d1+δ, d2)}, select an adjustment that provides the highest accuracy for decoding the selected chirp signal based on the demodulation accuracy of the fixed chirp signal. If δ is found to exceed the threshold (|δ|<80), the receiver discards the data frame and requests retransmission, i = 0, ..., δ.

[0024] Steps 1-2 include: The corresponding time-domain function x(t) of the linear frequency modulated chirp signal is expressed as:

[0025]

[0026] Where c is the frequency modulation frequency, f0 is the starting frequency, and φ0 is the initial phase at time t = 0; based on the corresponding time domain function x(t), the 3-bit information is modulated by changing the frequency, shape, and phase of the linear frequency modulation signal chirp.

[0027] In steps 1-2 (the specific parameter values ​​used in this step are only for illustrative purposes), the step of changing the frequency of the linear frequency modulated signal chirp based on the corresponding time-domain function x(t) specifically includes: encoding bits 0 and 1 using f0 = 300Hz and f0 = 1.9kHz respectively, given a sampling rate of 44.1kHz and a symbol duration of 0.001s, bit 0 is represented by the frequency range [300Hz, 1900Hz], and bit 1 is modulated by the range (1900Hz, 3400Hz).

[0028]

[0029] Where b represents a bit;

[0030] The change in the shape of the linear frequency modulated signal chirp specifically includes: switching the start frequency f0 and the end frequency f1 within the frequency range of [300Hz, 1900Hz] or (1900Hz, 3400Hz] to encode an additional 1 bit of information, thereby changing the shape of the chirp without altering its frequency band.

[0031]

[0032] The change of phase of the linear frequency modulated signal chirp specifically includes: modulating additional bits by using different initial phases without changing the frequency and shape of the linear frequency modulated signal chirp; whether the signal carries 0 or 1 depends on whether the initial phase φ0 is equal to zero.

[0033]

[0034] Step 2 includes:

[0035] Step 2-1, Signal Features: Demodulate using the time domain, frequency domain, and phase angle features of each chirp signal, and extract frequency domain features from three sources. The first source is extracted from all sampling points (the discretization of a continuous signal in time, taking its instantaneous value point by point on the analog signal x(t) at a certain time interval Δt). The second and third sources are extracted from the first half of the sampling points and the second half of the sampling points, respectively.

[0036] Step 2-2, Signal Demodulation: Principal Component Analysis (PCA) is used to reduce the dimensionality of the frequency domain features. Then, a deep learning model ResNet34 is used to demodulate the features generated by PCA. The deep learning model takes a 1×10×10 input and outputs 8 classes.

[0037] Step 3 includes:

[0038] Step 3-1: The receiving end uses Reed-Solomon error correction code (RS) to correct errors in the received data;

[0039] Step 3-2: When the Reed-Solomon error correction code (RS) check fails, a retransmission mechanism will be triggered. When the sender has not sent all data frames, the gap between the two chirp signal groups is used by the receiver to send feedback pulses. Once energy is detected in the gap, the sender will immediately retransmit the symbol group before the gap. The two chirp signal groups refer to two symbol groups, and each symbol group includes N chirp signals.

[0040] In step 3-2, the gap in data frame i+1 is used to identify the corresponding symbol group in data frame i;

[0041] If the receiver fails to receive certain data frames correctly, it will request the sender to retransmit the erroneous frames. Depending on whether the sender has finished sending all data frames, the receiver will choose a different retransmission request mechanism to request retransmission. If the sender has not finished sending all data frames, the receiver will use a pulse signal as a retransmission request signal during the interval between the sender sending the next data frame. If the sender has finished sending all data frames, the receiver will directly send a retransmission request frame to the sender to request retransmission.

[0042] The present invention has the following beneficial effects:

[0043] (1) A dedicated data link protocol with a stop / resume mechanism can effectively solve the problem of voice activity detection technology;

[0044] (2) A dedicated data link protocol with time synchronization and retransmission schemes between the caller and the callee can greatly ensure the integrity of data reception.

[0045] (3) Based on chirp signal modulator, since chirp signal has strong anti-interference ability, it can tolerate noise and signal distortion on cellular voice channel, and greatly ensure the correctness of data transmission process;

[0046] (4) Deep learning-based demodulators can improve the accuracy of signal demodulation by learning various features of the signal.

[0047] (5) The retransmission scheme between the caller and the callee ensures the integrity and correctness of data reception by retransmitting the data frame with the transmission error when the receiver fails to receive the data correctly. Attached Figure Description

[0048] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0049] Figure 1This is a diagram of the architecture of the present invention.

[0050] Figure 2 It is the format of the data frame.

[0051] Figure 3 Examples of eight modulation-related signals are given.

[0052] Figure 4 This is a diagram illustrating retransmission between the caller and the callee. Detailed Implementation

[0053] The technical solution adopted in this invention is a fast data transmission mechanism that provides reliable output over an unknown cellular channel, specifically involving the modulation of chirp signals and a deep learning-based demodulation mechanism, such as... Figure 1 As shown, it includes the following steps:

[0054] Step 1, Data Modulation: The sender modulates the data information using a modulation mechanism based on chirp signals and a dedicated data link protocol with stop, resume, and time synchronization mechanisms.

[0055] Step 2, Data Demodulation: The analog signal is demodulated using a demodulation mechanism based on principal component analysis and deep learning.

[0056] Step 3, Error Recovery and Retransmission: The received data is checked for correctness, and errors are corrected. If completely correct data cannot be obtained after error correction, a retransmission mechanism is used to notify the sender to retransmit the erroneous frames. Finally, all correct frames are combined to recover the original data.

[0057] Furthermore, step 1 includes the following steps:

[0058] Step 1-1: The sender will adjust the data signal according to a dedicated data link protocol with stop / resume mechanisms and time synchronization to minimize the impact of VAD and allow the receiver to determine the exact location of the data within the audio stream. Data frames are as follows: Figure 2 As shown, each data frame consists of several chirp signals, which are divided into several chirp signal groups (each chirp signal group can be called a symbol group). Symbol groups are separated by gaps. A unique chirp is added at the beginning and end of each data frame as a separator to indicate when the data frame begins and ends.

[0059] Furthermore, to detect the exact location of the delimiter at the receiver, the data frame employs a cross-correlation-based method, where the known delimiter signal is correlated with the received audio stream within a sliding window. Let the received signal stream be {u}. i}(i=1,2,…,n), and each u i These are all audio samples. The separator sent by the sender is represented as {v i}(i=1,2,…,m), and n>>m. Use matched filtering from {u i Extract a sliding window of length m from the sample. The sample correlation coefficient r is calculated as follows:

[0060]

[0061] in and These are the sampled mean of the sliding window and {v} i For each r, Welford's one-way algorithm achieves a computational complexity of O(m). As the sliding window moves sample-by-sample from the beginning to the end of the audio stream, the total complexity of computing each r equals O(nm). Large values ​​of r indicate high similarity between the two sequences. The location of the separator can be found at the maximum peak of these coefficients.

[0062] Furthermore, to accelerate computation time, the approximate position c of the separator is first located by calculating coefficients window by window. Within the range of [cm, c+m], the scheme uses fine-grained correlation with a sliding window size of 1 in parallel multi-threaded operation to locate the accurate position.

[0063] Furthermore, due to unpredictable channel noise, cross-correlation may sometimes fail to accurately locate delimiters, resulting in exported data frames that do not have a predefined length. Therefore, a delimiter position adjustment scheme is developed. Assume each data frame contains k audio samples, and two delimiters surrounding the frame are located at sample indices d1 and d2. Ideally, d2 - d1 - m should equal k, where m is the length of the delimiter. However, in reality, d2 - d1 - mk = δ, where δ can be positive or negative, and has a high probability of not being zero. Therefore, d1 and d2 need to be adjusted to compensate for this δ. First, some symbols in the data frame are set to fixed values ​​so that the receiver knows these symbols in advance. Next, an adjustment that provides the highest accuracy for decoding the selected symbols is attempted from each possible receiver data frame of {(d1, d2 - δ), ..., (d1 + i, d2 - δ + i), ..., (d1 + δ, d5)} (i = 0, ..., δ). If δ is found to exceed a threshold, the receiver simply discards the data frame and requests a retransmission.

[0064] Steps 1-2 modulate the data certificate based on the chirp signal. Chirp has strong anti-interference characteristics and is widely used in communications, sonar, radar, and other fields. Linear frequency chirp is sufficient to tolerate noise and signal distortion in cellular voice channels. In linear frequency modulation, the instantaneous frequency f(t) = ct + f0 changes strictly linearly with time t, where c is the modulation frequency and f0 is the starting frequency. The corresponding time-domain function of linear frequency modulation chirp can be expressed as follows:

[0065]

[0066] Where φ0 is the initial phase at time t = 0. Based on the above equation, the 3-bit information is modulated by changing the frequency, shape, and phase of the linear frequency modulated signal chirp.

[0067] 1) Frequency: Bits 0 and 1 are encoded using f0 = 300 Hz and f0 = 1.9 kHz, respectively. Given a sampling rate of 44.1 kHz and a symbol duration of 0.001 s, bit 0 is represented by a frequency range of [300 Hz, 1900 Hz], and bit 1 is modulated by a range of (1900 Hz, 3400 Hz).

[0068]

[0069] 2) Shape: Within the frequency range of [300Hz, 1900Hz] or (1900Hz, 3400Hz], the start frequency f0 and end frequency f1 can be switched to encode an additional 1 bit of information. In other words, the shape of the chirp can be changed without altering its frequency band. For example, 300Hz → 1.9kHz represents bit 0, and 1.9kHz → 300Hz represents bit 1.

[0070]

[0071] 3) Phase: Without changing the frequency and shape of the linear frequency modulated signal chirp, additional bits can be modulated by using different initial phases. Whether the signal carries 0 or 1 depends on whether the initial phase φ0 is equal to zero.

[0072]

[0073] Furthermore, the specific method for signal demodulation in step 2 includes the following steps:

[0074] Step 2-1, Signal Features: Demodulation is performed using the time-domain, frequency-domain, and phase angle features of each chirp signal. To improve recognition accuracy, frequency-domain features are extracted from three sources: the first part is extracted from all sampling points; the second and third parts are extracted from the first and second halves of the sampling points, respectively.

[0075] Step 2-2, Signal Demodulation: First, Principal Component Analysis (PCA) is used to reduce the feature dimension to 100. PCA is a commonly used dimensionality reduction method. It works by performing linear projection and can alleviate overfitting problems. Then, a deep learning model, ResNet34 (ResNet34 is a residual network proposed by Microsoft Labs in 2015, a type of convolutional neural network), is used to demodulate the features generated by PCA. The deep learning model takes a 1×10×10 input (the PCA output will be adjusted to this size) and outputs 8 classes (8 different classes of signals will be modulated using the method in Step 1-1, such as...). Figure 3 (As shown).

[0076] Furthermore, the specific method for error recovery and retransmission in step 3 includes the following steps:

[0077] Step 3-1: The receiving end uses RS to correct errors in the received data.

[0078] Step 3-2: When the RS check fails, a retransmission mechanism will be triggered, such as... Figure 4 As shown. When the sender has not yet transmitted all data frames, the gap between two signal groups is used by the receiver to send a feedback pulse. Once energy is detected in the gap, the sender will immediately retransmit the symbol group before the gap ends. For ease of implementation, the gap in data frame i+1 is used to confirm the corresponding symbol group in data frame i.

[0079] Furthermore, if the receiver cannot successfully receive all frames when the sender has sent all data frames, a second mechanism is triggered after the transmitter sends the last data frame. The receiver sends a separate ACK (Acknowledgment) frame to notify of the lost signal groups in the previous data frames. Such an ACK includes a separator and a gap, and the index of the gap is associated with the symbol group.

[0080] Experimental Results: This invention was tested on networks provided by China's three major cellular operators (China Mobile, China Telecom, and China Unicom). The main performance indicators were accuracy, throughput, and effective throughput. Accuracy refers to the demodulation accuracy in classifying modulation symbols. Throughput is the total number of received bits, including protocol overhead bits and repetition bits per unit time. Effective throughput is the amount of useful information transmitted to the receiver per unit time, which is the size of the transmitted file divided by the time required to transmit the file.

[0081] In this embodiment of the invention, a chirp separator of 0.1s length is used, and each symbol group consists of a chirp of 0.001s length, with a gap duration of 0.5s. Experiments show that the research method of this patent has similarly good performance in all cellular networks. The throughput can reach 1785.12 bits / s, and the average effective throughput can reach 1265.43 bits / s. In environments with good signal strength, the data obtained by deep learning demodulation, after RS ​​error correction, achieves an average accuracy of 100%; while in environments with moderate signal strength, the average accuracy is still 99.56%. All of the above data indicate that this patent has good experimental results.

[0082] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a method for achieving faster end-to-end data transmission via a voice channel, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0083] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0084] This invention provides a method for achieving faster end-to-end data transmission via a voice channel. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for achieving faster end-to-end data transmission via a voice channel, characterized in that, Includes the following steps: Step 1, Data Modulation: The sender generates a random certificate or uses an existing certificate, and uses a data link protocol based on stop and resume mechanisms and time synchronization, and uses a chirp-based modulation method to modulate the data certificate into an analog signal, which is then transmitted through the voice channel; Step 2, Data Demodulation: The receiver receives the signal and uses a deep learning-based demodulation method to demodulate the analog signal into a data certificate; Step 3, Error Recovery and Retransmission: The receiver corrects the received data certificate based on the error correction code. If the certificate cannot be completely restored, a retransmission request is sent to the sender. The sender receives the retransmission request and retransmits the certificate; Step 1 includes: Step 1-1: The sender adjusts the data signal according to the data link protocol with stop, resume mechanism and time synchronization, so that the receiver can determine the exact position of the data in the audio stream. Each audio stream contains N data frames, each data frame includes N chirp signals, and the N chirp signals are divided into two or more chirp signal groups. Each chirp signal group is called a symbol group. Symbol groups will be separated by gaps, and a unique chirp signal will be added at the beginning and end of each data frame as a separator to indicate when the data frame begins and ends. Steps 1-2: Modulate the data certificate based on the chirp signal; In step 1-1, the following delimiter position adjustment scheme is formulated: Each data frame is set to contain... There are 10 audio sample points, and the two delimiters surrounding the data frame are located at the sample point index. and The index is based on the location of the sampling point. and Position the starting points of the two separators. ,in It is the length of the separator. It is the recipient based on and The difference between the number of sampling points in the received data frame and the number of sampling points in the sent data frame; Adjustment and To make up for First, set some chirp signals in the data frame to fixed values, in {( , ), ..., ( , ), ..., ( , In each possible receiver data frame, based on the demodulation accuracy of the fixed chirp signal, an adjustment that provides the highest accuracy for decoding the selected chirp signal is selected. If it is found... If the threshold is exceeded, the receiver discards the data frame and requests a retransmission. .

2. The method according to claim 1, characterized in that, In step 1-1, to detect the exact location of the delimiter at the receiver, the data frame employs a cross-correlation-based method, where the known delimiter signal is correlated with the received audio stream within a sliding window: Assume the received audio stream comprises N audio sample points, and based on these sample points, the entire audio stream is represented as { }, The separator sent by the sender is represented as { based on its sampling points} }, ; and n is the audio stream duration * signal sampling rate; m is the separator duration * signal sampling rate. Use matched filtering from { Extracting length equal to} The sliding window is used to calculate the sample correlation coefficient using the following formula. : ， in It is the average of the samples from the sliding window. { } 3. The method according to claim 2, characterized in that, In step 1-1, the approximate location of the separator is first determined by calculating the sample correlation coefficient window by window. ,exist[ , Within the range of ], multiple threads use fine-grained correlation with a sliding window size of 1 to locate the accurate position in parallel.

4. The method according to claim 3, characterized in that, Steps 1-2 include: the corresponding time-domain function of the linear frequency modulated chirp signal. Represented as: , Where c is the frequency modulation frequency. It is the starting frequency. It is the initial phase at time t=0; based on the corresponding time-domain function. By changing the frequency, shape, and phase of the linear frequency modulated signal chirp, 3 bits of information can be modulated.

5. The method according to claim 4, characterized in that, In steps 1-2, the step based on the corresponding time-domain function Changing the frequency of the linear frequency modulated signal chirp specifically involves: encoding bits 0 and 1 respectively, given a sampling rate and symbol duration, where bit 0 is represented by a frequency range [300Hz, 1900Hz], and bit 1 is modulated by a range (1900Hz, 3400Hz). ； Where b represents a bit; The change in the shape of the linear frequency modulated signal chirp specifically includes: switching the starting frequency within the frequency range of [300Hz, 1900Hz] or (1900Hz, 3400Hz]. and end frequency To encode an additional 1 bit of information, the shape of the chirp is changed without altering its frequency band: ； The change of phase of the linear frequency modulated signal chirp specifically includes: modulating additional bits by using different initial phases without changing the frequency and shape of the linear frequency modulated signal chirp; whether the signal carries 0 or 1 depends on the initial phase. Is it equal to zero? 。 6. The method according to claim 5, characterized in that, Step 2 includes: Step 2-1, Signal Features: Demodulate using the time domain, frequency domain, and phase angle features of each chirp signal, and extract frequency domain features from three sources. The first source is extracted from all sampling points, and the second and third sources are extracted from the first half and the second half of the sampling points, respectively. Step 2-2, Signal Demodulation: Principal Component Analysis (PCA) is used to reduce the dimensionality of the frequency domain features. Then, a deep learning model ResNet34 is used to demodulate the features generated by PCA. The deep learning model takes a 1×10×10 input and outputs 8 classes.

7. The method according to claim 6, characterized in that, Step 3 includes: Step 3-1: The receiving end uses Reed-Solomon error correction code (RS) to correct errors in the received data; Step 3-2: When the Reed-Solomon error correction code (RS) check fails, a retransmission mechanism will be triggered. When the sender has not sent all data frames, the gap between the two chirp signal groups is used by the receiver to send feedback pulses. Once energy is detected in the gap, the sender will immediately retransmit the symbol group before the gap. The two chirp signal groups refer to two symbol groups, and each symbol group includes N chirp signals.

8. The method according to claim 7, characterized in that, In step 3-2, data frames are used. Use the gaps in the data frame to confirm the data frame The corresponding symbol group in; If the receiver fails to receive the data frame correctly, it will request the sender to retransmit the erroneous frame. Depending on whether the sender has sent all data frames, the receiver will choose a different retransmission request mechanism to request retransmission from the sender. If the sender has not sent all data frames, the receiver will use a pulse signal as a retransmission request signal during the interval between the sender sending the next data frame. If the sender has already sent all data frames, the receiver will directly send a retransmission request frame to the sender to request a retransmission.

Images