A deep learning-based electrocardiogram signal sparse encryption transmission and recovery method

Abstract

The application discloses a kind of electrocardiogram signal sparse encryption transmission and recovery method based on deep learning, it is related to electrocardio signal processing field.The method includes: in sending end, by electrocardio collection terminal, original sensing signal is collected, and original sensing signal is compressed and encrypted by using randomly generated key and encryption measurement matrix that electrocardiogram acquisition device is collected;In receiving end, by key and encryption measurement matrix, channel noise and key interference component are stripped, and the measurement vector containing sparse feature is obtained;The internal sparse feature of original sensing signal is learned by the deep neural network that is learned in advance, so that sensing signal is accurately decrypted and recovered.The method provided by the application can effectively improve the security performance and spectral efficiency of the system under the condition that the computing resources and energy supply are limited, and can be applied to the remote electrocardiogram monitoring scene of home wear and implantable medical device.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram (ECG) signal processing, and in particular to a method for sparse encrypted transmission and recovery of ECG signals based on deep learning. Background Technology

[0002] In recent years, with the rapid development of information and communication technologies, the Internet of Things (IoT) has been widely applied in the healthcare field. To monitor patients' health in real time, electrocardiogram (ECG) signal acquisition and transmission modules are commonly integrated into many wearable and implantable medical devices. However, due to the open nature of wireless channels, ECG signal transmission faces the risk of eavesdropping attacks. Furthermore, IoT medical devices typically have limited computing resources and energy supplies, placing higher demands on the computational complexity of encryption algorithms and the spectral efficiency of the system. Therefore, it is necessary to design an efficient and secure ECG signal transmission method suitable for resource-constrained conditions.

[0003] Existing methods for transmitting electrocardiogram (ECG) signals each have their own focus. For example, Chinese patent CN105100272A proposes an intelligent file transmission method and system for ECG machines to improve file transmission efficiency; Chinese patent CN106214144A proposes a remote ECG monitoring system to improve the wireless reliability of remote ECG acquisition; and Chinese patent CN113080985A proposes a wireless transmission ECG machine to improve the stability of signal transmission and the convenience of operation. However, none of the above methods consider a confidentiality mechanism, which may lead to the leakage of patient privacy information.

[0004] To address this, this invention proposes a sparse encryption transmission and recovery method for electrocardiogram (ECG) signals based on deep learning. This method utilizes deep neural networks to learn the sparse characteristics of ECG signals and employs a low-complexity algorithm to achieve signal compression and encryption, thereby improving spectral efficiency and security performance. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient privacy protection, high encryption complexity, low spectral efficiency, and incompatibility with low-resource medical devices in existing technologies. It provides a deep learning-based sparse encryption transmission and recovery method for electrocardiogram (ECG) signals. This method utilizes the sparsity of ECG signals for compression and encryption, and employs a constructed deep neural network to learn the sparse characteristics of ECG signals, further improving the recovery accuracy. This method can effectively improve the system's security performance and spectral efficiency under conditions of limited computing resources and energy supply.

[0006] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0007] A deep learning-based method for sparse encrypted transmission and recovery of electrocardiogram (ECG) signals includes the following steps:

[0008] S1. Within one sensing interval of the medical device, the electrocardiogram (ECG) acquisition device acquires data of a length of... raw sensor signal The length is raw sensor signal It possesses inherent sparse distribution characteristics in the discrete cosine sparse domain, providing a priori basis for subsequent compressed sensing measurements and deep learning sparse feature learning.

[0009] S2. At the sending end, generate a length of... key and compared it with the original sensing signal. Encrypted sensor signals are obtained by superposition. ;

[0010] S3, generated size is Encryption measurement matrix The encrypted sensing signal is compressed and encrypted to obtain a length of... Compressed encrypted ciphertext signal Using wireless channels to compress and encrypt ciphertext signals Transmitted to the receiving end;

[0011] S4. At the receiving end, let the received signal be... ,in, Noise in the wireless channel; using a key and encrypted measurement matrix Eliminate received signal The key components in the vector are used to obtain the sparse measurement vector. ;

[0012] S5. Construct a deep neural network, calculate the training dataset using the sensor signals collected by the electrocardiogram acquisition device, train the weights of the deep neural network so that the network can learn the inherent sparse features of the original sensor signals; the network takes the noisy sparse measurement vector as input and the sparse representation vector of the original signal as output, autonomously fits the sparse prior and suppresses channel noise interference.

[0013] S6. Transfer the sparse measurement vector The input is fed into the trained deep neural network to perform one round of feedforward computation to obtain the predicted sparse representation vector. ;

[0014] S7. Using sparse representation matrices Obtain the reconstructed sensing signal, i.e. It completes the sparse encrypted transmission and recovery of electrocardiogram signals.

[0015] In step S1, the length is raw sensor signal It is a signal that has sparse characteristics on a certain sparse representation domain, and can be derived from... sparse representation matrix Represented as ;

[0016] The sparse representation matrix Let be the discrete cosine transform matrix, and its first... Item for ,in, .

[0017] In step S2, the key The elements in the follow a uniform distribution, and their th The uniform distribution of the elements is represented as follows: ,in, It is a key Amplitude control parameters, and These are the original sensing signals. The minimum and maximum amplitudes.

[0018] In step S3, the encrypted measurement matrix Let be a Gaussian random matrix, and the terms in the matrix follow a normal distribution, wherein the nth term is... Item for .

[0019] In step S5, the deep neural network consists of multiple layers with the same structure, but the weights of each layer are not the same; the training dataset It is calculated from the original sensor signal acquired by the actual electrocardiogram acquisition device, wherein, This is a sparse representation vector of the samples. For sparse measurement vectors of the samples, The number of samples in the training dataset.

[0020] In step S5, the deep neural network is trained layer by layer, and the training process is as follows:

[0021] S51, when the number of layers of the deep neural network is When, initialize the first The network weights of the layers are and ,in, The observation matrix is ​​formed by... Give;

[0022] S52, forward The weights of the first layer of the network remain fixed, and the backpropagation algorithm and Adam optimizer are used to optimize the second layer. Layer network weights To minimize the first Layer loss function ;

[0023] S53, if the Layer loss function This will increase the number of network layers. Otherwise, return to step S51; otherwise, training ends.

[0024] S54, The final number of network layers is output as follows: The trained network weights .

[0025] In step S52, the first step in the training process The loss function of the layer is ,in, When the sample sparse measurement vector is Network weights are At that time, the network predicts the sparse representation vector of the sample.

[0026] Compared with the prior art, the present invention has the following outstanding technical effects and advantages:

[0027] 1. Lightweight encryption adapted to low-resource devices: It adopts key element-by-element superposition and Gaussian measurement matrix compression encryption, which has a computational complexity far lower than traditional symmetric / asymmetric encryption algorithms. It does not require high computing power and power consumption, and is adapted to the computing and energy-constrained hardware of wearable and implantable ECG medical devices.

[0028] 2. Sparse compression improves spectrum efficiency: By utilizing the natural sparseness of the discrete cosine domain of ECG signals, compressed transmission is achieved, reducing the amount of data transmitted over the wireless channel, reducing spectrum resource usage, and improving the ability to monitor multiple devices concurrently.

[0029] 3. End-to-end privacy and anti-eavesdropping: The pre-shared key synchronization mechanism at the transmitting and receiving ends, combined with double encryption of the measurement matrix, ensures that the wireless channel only transmits compressed ciphertext signals, eliminating the risk of leakage of the original plaintext ECG and effectively resisting channel eavesdropping attacks, thus protecting patient medical privacy.

[0030] 4. Deep learning-driven high-precision reconstruction: It adopts a deep neural network trained layer by layer to autonomously learn sparse prior features of ECG signals and suppress wireless channel noise interference. Compared with traditional sparse reconstruction algorithms, the signal-to-noise ratio and waveform matching degree of signal reconstruction are significantly improved.

[0031] 5. High versatility and engineering feasibility: It can be directly embedded into existing ECG acquisition terminals and remote monitoring platforms without significant changes to the hardware architecture, and is suitable for various remote medical and home health monitoring scenarios. Attached Figure Description

[0032] Figure 1 This is a diagram of a deep neural network structure according to an embodiment of the present invention.

[0033] Figure 2 The images show the reconstruction results of three sensing signals from the MIT-BIH arrhythmia database. (a) shows the reconstruction result of the ECG signal from record number 100, with a percentage root mean square error (PRD) of 0.833%; (b) shows the reconstruction result of the ECG signal from record number 102, with a percentage root mean square error (PRD) of 0.908%; and (c) shows the reconstruction result of the ECG signal from record number 220, with a percentage root mean square error (PRD) of 0.468%.

[0034] Figure 3 The sparse representation of the reconstruction results of the three sensing signals in the DCT domain is shown. Among them, (a) is the sparse reconstruction result of the ECG signal of record number 100 in the DCT domain, with a normalized mean square error (NMSE) of -30.4dB; (b) is the sparse reconstruction result of the ECG signal of record number 102 in the DCT domain, with a normalized mean square error (NMSE) of -34.9dB; and (c) is the sparse reconstruction result of the ECG signal of record number 220 in the DCT domain, with a normalized mean square error (NMSE) of -34.6dB.

[0035] Figure 4 The proposed solution (SLER) and the comparative solution are shown to demonstrate the PRD performance of the proposed solution (SLER) and the comparative solution at different compression ratios.

[0036] Figure 5 The proposed scheme (SLER) and the comparative scheme represent the successful reconstruction probabilities at different compression rates. Detailed Implementation

[0037] To provide a clearer understanding of the technical content of this invention, the following embodiments are provided for detailed description. Technical steps and parameter settings not described in detail can be implemented using conventional techniques in the art.

[0038] The deep neural network structure according to the embodiments of the present invention is as follows: Figure 1 As shown, this deep neural network can effectively improve the recovery accuracy of electrocardiogram (ECG) signals. A deep learning-based sparse encrypted transmission and recovery method for ECG signals includes the following steps:

[0039] S1. Within one sensing interval of the medical device, the electrocardiogram acquisition device acquires a raw sensing signal with a length of 500. The original sensing signal A signal that has sparse characteristics in a certain sparse representation domain, and can be derived from... sparse representation matrix Represented as The sparse representation matrix Let be the discrete cosine transform matrix, and its first... Item for ,in, .

[0040] S2. At the sending end, generate a key with a length of 500. and compared it with the original sensing signal. Encrypted sensor signals are obtained by superposition. The key The elements in the follow a uniform distribution, and their th The uniform distribution of the elements is represented as follows: ,in, It is the key Amplitude control parameters, and The original sensing signals are respectively The minimum and maximum amplitudes.

[0041] S3, generated size is Encryption measurement matrix The encrypted sensor signal is then compressed and encrypted to obtain a compressed encrypted ciphertext signal with a length of 250. Using wireless channels to compress and encrypt ciphertext signals Transmitted to the receiving end; the encrypted measurement matrix Let be a Gaussian random matrix, where terms follow a normal distribution, and its nth term is... Item for .

[0042] S4. At the receiving end, let the received signal be... ,in, Noise in the wireless channel; using the key and the encrypted measurement matrix The received signal is eliminated through matrix operations. From the key components, a sparse measurement vector is obtained, i.e. .

[0043] S5, Construction as follows Figure 1 The deep neural network uses sensor signals acquired by an electrocardiogram (ECG) acquisition device to calculate a training dataset, and trains the weights of the deep neural network so that the network can learn the inherent sparse characteristics of the original sensor signals.

[0044] First, construct a... Figure 1 A deep neural network in which each layer has the same structure, but the weights of each layer are not the same; the deep neural network's... The following calculations are performed sequentially within the layer. , , , , ;in, For the deep neural network's first... The pseudo-measurement vector of the layer, For the deep neural network's first... The sparse representation vector recovered by the layer network For the deep neural network's first... Learnable network weights of layers For the deep neural network's first... Residual measurement error of the layer, The residual measurement error standard deviation It is a contraction function. The learnable shrinkage threshold of the shrinkage function. The pseudo-measurement vector The One element, For the deep neural network's first... Layer Onsager modification.

[0045] Secondly, a training dataset is calculated using the sensor signals acquired by the electrocardiogram (ECG) acquisition device. This dataset is then used to train the weights of the deep neural network, enabling it to learn the inherent sparse characteristics of the original sensor signals. It is calculated from the original sensor signal acquired by the actual electrocardiogram acquisition device, wherein, This is a sparse representation vector of the samples. The sample sparse measurement vectors are given, and the number of samples in the training dataset is given. The deep neural network is trained layer by layer, and its training process is as follows:

[0046] S51, when the number of layers of the deep neural network is When, initialize the first The network weights of the layers are and ,in, The observation matrix is ​​formed by... Given, among which, To encrypt the measurement matrix, It is a sparse representation matrix.

[0047] S52, forward The weights of the first layer of the network remain fixed, and the backpropagation algorithm and Adam optimizer are used to optimize the second layer. Layer network weights To minimize the loss function The learning rate of the Adam optimizer is set to... The first step in the training process The loss function of the layer is ,in, The sample sparse measurement vector is The network weights are At that time, the network predicts the sparse representation vector of the sample.

[0048] S53, if the Layer loss function This will increase the number of network layers. If the condition is met, return to step S51; otherwise, training ends.

[0049] S54, The final number of network layers is output as follows: The trained network weights .

[0050] S6. Transfer the sparse measurement vector The weights input into the trained network are In a deep neural network, a feedforward computation is performed to obtain the predicted sparse representation vector. .

[0051] S7. Using sparse representation matrices Obtain the reconstructed sensing signal, i.e. It completes encrypted transmission and high-precision recovery of electrocardiogram signals.

[0052] The reconstruction results of some typical electrocardiogram sensing signals from the MIT-BIH arrhythmia database using the proposed scheme are as follows: Figure 2 As shown, the PRDs corresponding to the ECG signals of records 100, 102, and 220 are 0.833%, 0.908%, and 0.468%, respectively. Although the signal quality is slightly reduced due to wireless transmission, the reconstructed signals highly match the original signal waveforms, and the reconstruction accuracy still reaches an "excellent" level in the field of ECG signal analysis. Furthermore, Figure 3 This demonstrates the sparse reconstruction performance of the above signal in the Discrete Cosine Transform (DCT) domain. For example... Figure 3 As shown, the sparse recovery NMSEs of the three ECG signals are -30.4dB, -34.9dB, and -34.6dB, respectively. The results indicate that the proposed scheme can effectively preserve the sparsity and integrity of the sensing signal while achieving compressed and encrypted transmission.

[0053] Figure 4 The PRD performance of various sparse reconstruction schemes under different compression ratios is demonstrated. For example... Figure 4As shown, the proposed scheme (the scheme indicated by SLER in the figure) exhibits significantly lower reconstruction errors at different compression ratios compared to traditional compressed sensing algorithms (OMP, SP) and traditional iterative sparse recovery algorithms (AMP). For example, the SLER scheme achieves lower reconstruction errors at compression ratios of [missing information - likely a specific compression ratio]. It can achieve a PRD similar to that of SP and OMP algorithms at a compression rate of 70%. Furthermore, it will increase the probability of successful reconstruction (i.e., The probability of is used as an evaluation metric for sparse recovery performance in the DCT domain. Figure 5 The diagram illustrates the successful reconstruction probabilities of various sparse recovery schemes under different compression ratios. Figure 5 It can be seen that, Even with a low compression rate (i.e., the original sensed signal is compressed to only one-quarter of its original size), the SLER scheme proposed in this invention still achieves a successful reconstruction probability of no less than 0.9, significantly higher than other comparative schemes. Simulation results show that by effectively learning and fully utilizing the sparsity characteristics of the sensed signal, the SLER scheme proposed in this invention can achieve high recovery accuracy and spectral efficiency in medical IoT signal transmission and recovery, even when the original sensed signal is compressed at an extremely low compression rate. Therefore, the scheme of this invention can significantly reduce the amount of data transmitted in wireless communication between medical IoT devices and post-processing receivers, thereby effectively saving transmission energy consumption and spectral costs.

[0054] In summary, this invention, relying on a technical architecture combining sparse compression encryption and deep learning reconstruction, effectively solves the problems of insufficient privacy in existing ECG signal transmission, limited computing power and power consumption of devices, and poor signal recovery accuracy. The solution has a clear structure, is easy to operate, and has excellent anti-interference capabilities. It can be widely adapted to remote data transmission scenarios of various wearable and implantable ECG monitoring devices, and has good practical application value and industrialization prospects.

[0055] The above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Conventional modifications, equivalent substitutions, and structural fine-tunings made by those skilled in the art based on the technical concept of the present invention should all fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for sparse encrypted transmission and recovery of electrocardiogram (ECG) signals based on deep learning, characterized in that, Includes the following steps: S1. Within one sensing interval of the medical device, the electrocardiogram (ECG) acquisition device acquires data of a length of... raw sensor signal ; S2. At the sending end, generate a length of... key and compared it with the original sensing signal. Encrypted sensor signals are obtained by superposition. ; S3, generated size is Encryption measurement matrix The encrypted sensing signal is compressed and encrypted to obtain a length of... Compressed encrypted ciphertext signal Using wireless channels to compress and encrypt ciphertext signals Transmitted to the receiving end; S4. At the receiving end, let the received signal be... ,in, Noise in the wireless channel; using the key and the encrypted measurement matrix Eliminate received signal The key components in the data are used to obtain the sparse measurement vector. ,Right now ; S5. Construct a deep neural network. Calculate the training dataset using the original sensor signals collected by the electrocardiogram acquisition device. Train the weights of the deep neural network so that the network can autonomously learn the inherent sparse characteristics of the original sensor signals. During training, noise distribution characteristics are learned synchronously to achieve noise suppression; S6. Transfer the sparse measurement vector The input is fed into a pre-trained deep neural network, where a feedforward computation is performed to obtain the predicted sparse representation vector. ; S7. Using sparse representation matrices Obtain the reconstructed sensing signal, i.e. It completes the sparse encrypted transmission and recovery of electrocardiogram signals.

2. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 1, characterized in that... In step S1, the length is raw sensor signal For a signal that has sparse features in a certain sparse representation domain, and is composed of a sparse representation domain that matches the signal dimension. 3D sparse representation matrix Represented as ,in, It is a sparse coefficient vector.

3. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 2, characterized in that... The sparse representation matrix Let be the discrete cosine transform matrix, and its i.e., Item for ,in, The normalized coefficients of the discrete cosine transform matrix, m is the row number and n is the column number.

4. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 1, characterized in that... In step S2, the key The elements in the follow a uniform distribution, and their th element... The uniform distribution of the elements is represented as follows: ,in, It is the key Amplitude control parameters, and The original sensing signals are respectively The minimum and maximum amplitudes.

5. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 1, characterized in that... In step S3, the encrypted measurement matrix Let be a Gaussian random matrix, meaning that its terms follow a normal distribution, and its i-th term... Item for .

6. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 1, characterized in that... In step S5, the deep neural network consists of multiple layers with the same structure, but the weights of each layer are not the same; the training dataset It is calculated from the original sensor signal acquired by the actual electrocardiogram acquisition device, wherein, This is a sparse representation vector of the samples. For sparse measurement vectors of the samples, The number of samples in the training dataset.

7. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 1, characterized in that... In step S5, the deep neural network is trained layer by layer, and the training process is as follows: S51, when the number of layers of the deep neural network is When, initialize the first The network weights of the layers are and ,in, The observation matrix is ​​formed by... Give; S52, forward The weights of the first layer network remain fixed, and the backpropagation algorithm and Adam optimizer are used to optimize the second layer network. Layer network weights To minimize the loss function ; S53, if the Layer loss function This will increase the number of network layers. Otherwise, return to step S51; otherwise, training ends. S54, The final number of network layers is: The trained network weights .

8. The method for sparse encrypted transmission and recovery of electrocardiogram signals based on deep learning as described in claim 7, wherein in the training process of the first... The layer loss function is ,in, When the sample sparse measurement vector is Network weights are At that time, the network predicts the sparse representation vector of the sample.

Images