A method for signal detection in mobile molecular communication

Abstract

The application discloses a signal detection method in mobile molecular communication, constructs a signal detection neural network model based on Informer, the signal detection neural network model comprises a position coding module, an encoder module, a position coding decoding module and a decoder module, analyzes the autocorrelation coefficient of the molecular number sequence received by the receiving nanomachine in the sample data, determines the input length of the signal detection neural network model, trains the signal detection neural network model, adopts the trained signal detection neural network model, constructs the input sequence of the MMC system to be detected, and predicts the received transmission signal. Through the modification of the model structure, the model training time is reduced; through the fixed input length, the model training and calculation time is reduced, the training difficulty is greatly simplified, and the fixed mode of the input length is designed to maximize the completeness of the input information.

Description

Technical Field

[0001] This application belongs to the field of mobile molecular communication technology, and in particular relates to a signal detection method in mobile molecular communication. Background Technology

[0002] Molecular communication (MC) is a form of communication that uses biomolecules as information carriers. It has significant application prospects in fields such as biomedicine, especially in targeted drug delivery.

[0003] A key characteristic of mobile molecular communication (MMC) systems is the time-varying channel impulse response. To improve the detection capability of receivers in MMC systems, methods such as deep learning have been extensively studied for signal detection. For example, detectors using recurrent neural networks (RNNs), bidirectional RNN detectors, and neural network-based (CNN) detectors have been proposed to optimize signal detection. These methods aim to address the challenges in signal detection, particularly the detection of unknown signals in time-varying MMC systems. In mobile molecular communication (MMC), the channel impulse response varies with time. To improve the detection capability of receivers in MMC systems, numerous signal detection methods have been investigated.

[0004] Existing work mainly focuses on static signal detection (MC), or the training time of detection methods is relatively long. In MMC systems, when the channels are time-varying and the CSI is unknown, improving signal detection capability and reducing the time complexity of offline training is a challenge. Summary of the Invention

[0005] The purpose of this application is to provide a signal detection method in mobile molecular communication to improve signal detection capability and reduce the time complexity of offline training.

[0006] To achieve the above objectives, the technical solution of this application is as follows:

[0007] A signal detection method in mobile molecular communication, comprising:

[0008] A signal detection neural network model based on Informer is constructed. The signal detection neural network model includes a position encoding module, an encoder module, a position encoding decoding module, and a decoder module.

[0009] Generate sample data for training a signal detection neural network model;

[0010] Analyze the autocorrelation coefficients of the molecular number sequences received by the receiving nanomachines in the sample data to determine the input length of the signal detection neural network model;

[0011] After processing the sample data according to the determined length, the signal detection neural network model is trained.

[0012] A trained signal detection neural network model is used to construct the input sequence of the MMC system to be detected, and the received transmitted signal is obtained by prediction.

[0013] Furthermore, the position encoding module performs position encoding according to the following formula:

[0014] , ;

[0015] , ;

[0016] in The dimension representing the position encoding. Indicates the position in the input sequence. This represents the hidden layer dimension of the model. Indicates the even index position in the input sequence. This indicates the odd index position in the input sequence.

[0017] Furthermore, the encoder module includes a multi-head ProbSparse self-attention module, a feedforward neural network, and a normalization layer.

[0018] Furthermore, the decoding module for the position encoding is a feedforward neural network.

[0019] Furthermore, the decoder module is a feedforward neural network.

[0020] Further, the analysis of the sample data includes determining the input length of the signal detection neural network model by analyzing the autocorrelation coefficients of the molecular number sequence received by the receiving nanomachines.

[0021] The number of molecules received by the receiving nanomachines in each time slot is arranged into a time series according to time.

[0022] The difference between the position of each element and the position of the first element is used as the lag number to calculate the autocorrelation coefficient of each element with respect to the first element in the time series.

[0023] The lag number, taken from the convergence of the correlation coefficient, is used as the input length of the signal detection neural network model.

[0024] This application proposes a signal detection method for mobile molecular communication (MMC) using a novel signal sequence detector specifically designed for MMC systems. This detector is built upon an Informer model, with a particular focus on reducing training time. To achieve this, the length of the input sequence is fixed, and the structure of the Informer model is simplified to improve overall efficiency. Performance surpasses traditional deep neural networks (DNNs) and Transformer-based detectors, achieving a lower bit error rate (BER) and shorter training time. This application reduces model training time through modifications to the model structure; significantly simplifies training by fixing the input length to reduce both training and computation time; and designs a method for fixing the input length to maximize the completeness of the input information. Attached Figure Description

[0025] Figure 1 This is a flowchart of the signal detection method in mobile molecular communication according to this application.

[0026] Figure 2 This is a schematic diagram of mobile molecular communication.

[0027] Figure 3 This is a schematic diagram of the signal detection neural network model of this application.

[0028] Figure 4 This is a graph showing the relationship between autocorrelation and lag number.

[0029] Figure 5 The graph shows a comparison of the BER of the two experimental models.

[0030] Figure 6 The graph shows a performance comparison of the two models in terms of training time BER.

[0031] Figure 7 This is a schematic diagram illustrating how the bit error rate changes with the initial distance between two nodes.

[0032] Figure 8 This is a graph comparing the performance of the experimental models. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] The overall idea of ​​this application is to propose an Informer-based neural network model for signal detection in MMC systems. To improve the training efficiency of the Informer-based model, unnecessary components are removed, and a strategy of fixed-length input sequences is adopted. Based on this, a training data method is designed. Experimental results show that the proposed model achieves higher detection accuracy and shorter training time. The proposed model demonstrates better overall performance, providing support for effective detection methods in MMC systems.

[0035] One embodiment of this application, such as Figure 1 As shown, a signal detection method for mobile molecular communication is proposed, including:

[0036] Step S1: Construct a signal detection neural network model based on Informer. The signal detection neural network model includes a position encoding module, an encoder module, a position encoding decoding module, and a decoder module.

[0037] Informer is a Long Short-Term Memory (LSTF) network based on the Transformer architecture, designed to solve the problem of long-sequence time series prediction. It employs an innovative self-attention mechanism, ProbSparse, which simplifies self-attention computation, effectively captures long-term dependencies between sequences, and improves prediction accuracy. It also utilizes self-attention distillation, reducing the input to cascaded layers to effectively handle extremely long input sequences, enhancing the model's ability to process long sequences and making Informer adaptable to a wider range of applications. Informer employs a generative decoder, enabling it to predict the entire long-sequence sequence at once, rather than making predictions incrementally, significantly improving the inference speed for long-sequence prediction and making the Informer model more competitive in practical applications.

[0038] The traditional Informer architecture consists of an embedding module, an encoder module, and a decoder module. The embedding module comprises positional encoding, timestamp encoding, and a one-dimensional convolutional layer. The encoder module mainly consists of probabilistic sparse multi-head attention, a feedforward neural network, layer normalization, and residual connections. The decoder module mainly consists of probabilistic sparse multi-head attention, a feedforward neural network, positional encoding, layer normalization, and residual connections.

[0039] Unlike traditional Informer models, this application uses an Informer-based signal detection neural network model, such as... Figure 3As shown. The signal detection neural network model includes a positional encoding module, an encoder module, a positional encoding decoding module (FNN), and a decoder module. Compared to the traditional Informer structure, the modules of the signal detection neural network model in this application are described as follows:

[0040] 1. Position encoding module: Since the MMC signal detection problem only needs to retain the relative position of the sequence and not the actual time information, the timestamp encoding and one-dimensional convolutional layer are removed, and only the position encoding is retained to simplify the model structure and reduce the computational requirements.

[0041] 2. The encoder module includes a multi-head probSparse self-attention module, a feedforward neural network (FNN), and a normalization layer. Compared to the traditional Informer model, residual connections are removed. Traditional Informer models require stacking multiple attention modules, so residual connections are used to avoid the vanishing gradient problem that occurs when the number of layers is too deep. In MMC signal detection, fewer attention modules are needed to obtain sequence information, so residual connections are removed to simplify the network structure.

[0042] 3. The decoding module for position encoding employs a feedforward neural network, with a newly added FNN network for decoding the position encoding. For the MMC signal detection problem, the input is a long sequence and the positional information is crucial. Using an FNN for decoding the position encoding can help the model capture this positional information more accurately, but it will incur some additional computational costs.

[0043] 4. The decoder module employs a feedforward neural network, replacing the traditional decoder module with a simple FNN network. While this FNN reduces the performance of the Informer model, it improves computational speed. For longer sequences requiring the capture of long-range dependencies, the performance loss becomes more significant. Therefore, this application fixes the input length to ensure the performance loss is not too large, achieving a balance between performance loss and computational speed using the FNN model.

[0044] like Figure 2 As shown, the MMC system comprises a transmitting nanomachine (TN) and a passive receiving nanomachine (RN). The TN and RN move within the channel. The initial distance between the centers of the TN and RN is defined as... In time The distance between them is Molecules, TN, and RN diffuse in the channel at diffusion coefficients. , and The MMC system employs an ON / OFF keying modulation strategy, where the transmitter TN releases its signal when a bit 1 is transmitted. There are [number] molecules, and the molecules are not released when bit 0 is sent in each time slot, with a time slot duration of [duration]. .

[0045] In this embodiment, the signal detection neural network model takes as input the molecular quantity sequence received by the RN. Each data point is positionally encoded and then input into a matrix format.

[0046] In the Informer model, the input needs to be a fixed-length sequence. However, this model does not inherently capture positional information in the input data. To address this issue, the positional codes for even-indexed and odd-indexed positions in the sequence are denoted as follows: and These are manually added to the input, and the position encoding method can be different depending on the actual situation.

[0047] In one specific embodiment, the location code is calculated as follows:

[0048] , ;

[0049] , ;

[0050] in The dimension representing the position encoding. This indicates the position in the input sequence. This refers to the hidden layer dimension of the model, which usually corresponds to the feature dimension of the model.

[0051] The input data, with position encoding, is fed into the encoder module. In the encoder, the input sequence... The input sequence is multiplied by three different weight matrices respectively to make the input sequence The input data is transformed into three different vectors: a query (Q) vector, a key (K) vector, and a value (V) vector. Then, a similarity score is calculated between each element and elements in the sequence that have been filtered by a probability formula to determine their high relevance. Each similarity score is then used to weight its corresponding value (V) vector. This means that each element's score is multiplied by its corresponding value (V) vector, highlighting other elements with higher relevance to the current element. All these weighted value (V) vectors are summed and normalized using a softmax function to obtain the final output for each element. In this way, the output reflects the relevance and importance of different elements in the input data. The specific calculation formula is shown below:

[0052] , ;

[0053] in It is a vector used to determine the attention distribution at the current location. Represents a probability distribution . It is a vector used to measure the correlation between different locations, and It is an output vector obtained by weighted averaging based on the attention distribution. This is the scaling factor. The formula enables the Informer model to better handle data with uncertainty and variability, which can improve its performance in tasks such as time series forecasting.

[0054] The output is decoded by the FNN in the positional encoding decoding module and becomes a sequence. It is then input into the decoder module, i.e., the FNN. After the sequence passes through the FNN and the softmax layer, the dimension changes from 200 dimensions to 2 dimensions, resulting in the final binary output, i.e., the probabilities of 0 and 1. At this point, the result with the higher probability is identified as the output result of the network (0 or 1).

[0055] This application modifies the structure of the Informer model described above to reduce its computational complexity. It fully leverages the advantages of the Informer model in time series forecasting to solve the signal detection problem in MMC systems.

[0056] Step S2: Generate sample data for training the signal detection neural network model.

[0057] The signal detection neural network model of this application predicts the digital signal received by the receiving nanomachine (RN) in the MMC system, given the molecular number sequence received by the receiving nanomachine (RN). Therefore, when generating training samples, the digital signal to be transmitted is used as a label, and the molecular number sequence received by the receiving nanomachine (RN) in the MMC system is used as the input data sequence.

[0058] Specifically, the training dataset From the input data sequence and digital signals Its composition, and its formulaic expression, is as follows:

[0059] .

[0060] in, It is the digital signal transmitted in the j-th sample data, used as a tag. It is the sequence of the number of molecules received by the nanomachine (RN) in the j-th sample data in the MMC system.

[0061] In an MMC system, the probability of receiving a molecule from TN to RN at time t is expressed as:

[0062] ; (4)

[0063] in , Based on time The relative time. Let RN be the radius and its volume be... .

[0064] Then mean It is calculated in the following way:

[0065] , (5)

[0066] in .

[0067] Each time slot is divided into The sampling time period, the first Each sampling time period is represented as ,in .

[0068] In the time slot At the beginning The transmission probability of sending a bit 1 is determined manually by the sender in each time slot; therefore, an assumption of the probability of generating 0 and 1 is needed in the experiment. The experiment was conducted where the probability of the transmitter sending 0 and 1 in each time slot was 0.5, thus setting... .

[0069] In the time slot The number of molecules released from TN and reaching RN follows a Poisson distribution, expressed as:

[0070] , (6;

[0071] in, This represents the Poisson distribution.

[0072] Represented as The mean is given by the following formula:

[0073] , (7;

[0074] The ISI number refers to the number of molecules in the ISI system. Transmitted and received before the current time slot The number of molecules received in the current time slot, It is the first The average number of ISI molecules in a time slot is given by the following formula:

[0075] (8);

[0076] and These represent the cases where bit 0 and bit 1 are transmitted in the j-th time slot TN, respectively. Only one signal (0 or 1) is transmitted in the entire j-th time slot, which are represented as follows: and However, a time slot is further divided into several intervals (e.g., 50), and each interval RN has a number of receiver molecules, for a total of 50 receiver molecules.

[0077] For in and The number of molecules received at the w-th sampling time in the j-th time slot. and Following a Poisson distribution, it can be represented as:

[0078] (9);

[0079] , (10)

[0080] in, Indicates the amount of noise. This indicates the duration of a time slot, for example, set to 0.1s.

[0081] Therefore, after acquiring the digital signal to be transmitted in advance, the sequence of the number of molecules transmitted in each time slot of the MMC system can be obtained from the above formulas (9) and (10). ,in .in This indicates the slot number. In this experiment, the maximum number of slots (M) is set to 30,000.

[0082] Step S3: Analyze the autocorrelation coefficient of the molecular number sequence received by the receiving nanomachine in the sample data to determine the input length of the signal detection neural network model.

[0083] In this embodiment, the input length of the model is determined by analyzing the autocorrelation coefficient of the molecular number sequence received by the receiving nanomachine, and the input of the model is determined according to the algorithm guidance.

[0084] Specifically, the steps include:

[0085] Step 3.1: Arrange the number of molecules received by the receiving nanomachines in each time slot into a time series.

[0086] To determine the input length of the model, the autocorrelation of the molecular number sequence is considered. Let... Let N be a time series of length N in the j-th time slot. This time series may include the number of molecules in multiple time slots.

[0087] Step 3.2: Use the difference between the position of each element and the position of the first element as the lag number, and calculate the autocorrelation coefficient of each element in the time series with respect to the first element.

[0088] Consider the other elements in the sequence relative to the first element. The correlation is defined by the second element in the sequence. Relative to the first element The lag number is 1. Similarly, the elements in the sequence... For the first element The lag number is N.

[0089] Then in the j-th time slot, for The autocorrelation coefficient of an element with lag number m is... Its representation is as follows:

[0090] , (11)

[0091] in It is the average value of the sequence.

[0092] Step 3.3: Take the lag number from which the correlation coefficient tends to converge as the input length of the signal detection neural network model.

[0093] Figure 4 The horizontal axis in the graph represents the number of lags in the time series, and the vertical axis represents the value of the autocorrelation coefficient. (Analysis of the series...) about coefficient diagram ( Figure 4 From this, we can obtain that when the autocorrelation coefficient... The lag is close to zero near 200, meaning that determining a sequence length of 200 for the input of the Informer-based model is reasonable. This is because when the lag value is greater than 200, its impact on detecting the signal in the current time slot is relatively small. Therefore, the input length is determined to be 200, meaning the model receives data from four time slots. Obtaining the length of the input sequence has several advantages. On one hand, it allows capturing the most relevant information for signal detection, avoiding unnecessary computation and improving model efficiency. On the other hand, it helps prevent overfitting, avoids reliance on data with low autocorrelation coefficients, and ensures the model extracts necessary information from the data while eliminating redundant content.

[0094] Combining steps S2 and S3 above, the training sample generation process is as follows:

[0095] Step 4.1: Node TN releases molecules at the beginning of time slot j.

[0096] Node RN from node TN Number of molecules received in the next sampling Determined by equations (9) and (10).

[0097] Step 4.2: In each time slot, node RN performs W detections. Based on step 4.1, the number of molecules detected by node RN in each detection can be obtained.

[0098] The number of molecules received in each detection Arranged in order to form a sequence .

[0099] Step 4.3: Construct the training input sequence.

[0100] Based on the calculated input length, a training input sequence is constructed. When the input length is 200, four time-slot sequences are generated. , , and Arranged in chronological order to form a training input sequence. .

[0101] The number of samples, 50, is fixed, while the input length is determined based on the autocorrelation of the data (e.g., ...). Figure 3As shown in the figure, the specific setting is 200. In this embodiment, the number of time slots arranged is 200 / 50, which means that the molecular number sequences of four time slots are spliced ​​into one input sequence.

[0102] It should be noted that if j-3<0, then fill in 0s to extend the sequence length to 200.

[0103] Finally, the training input sequence With digital signals By concatenating the data, we obtain a training dataset:

[0104] .

[0105] Step S4: After processing the sample data according to the determined length, train the signal detection neural network model.

[0106] In this experiment, 30,000 data points were used for training. The training epochs were 100. Stochastic gradient descent was used to update the model parameters. The model used cross-entropy loss as its loss function.

[0107] Specifically, when using a model to detect input signal sequences At that time, the model will calculate the output sequence. The cross-entropy loss is expressed as:

[0108] , (12)

[0109] in This represents the bit detected in the j-th time slot based on the Informer model. This represents the cross-entropy between the detected bit and the actual bit.

[0110] Step S5: Using the trained signal detection neural network model, construct the input sequence of the MMC system to be detected, and predict the received transmission signal.

[0111] After training the signal detection neural network model, it can be used to detect signals. First, the number of molecules received by the receiving nanomachine is obtained, and then processed according to the input length to obtain the input sequence. Then, the input is fed into the trained signal detection neural network model to detect the received transmission signal.

[0112] Specifically, assume that the transmitting end TN sends a transmission signal [0,1,0], and releases a fixed number of molecules at the beginning of each time slot when transmitting 1. In the first time slot, the receiving end RN detects 50 times to obtain the sequence of received molecule counts. At this point, 150 zeros and The concatenated input is fed into a neural network, which calculates based on the input to determine that the first time slot TN transmits a signal of 0. The RN then performs 50 more detections in the second time slot to obtain the sequence of received molecule counts. splicing 100 zeros and , The input is fed into the neural network, which calculates that the signal transmitted in the second time slot is 1. Similarly, the third time slot is concatenated with 50 zeros and... , , The signal is calculated to be 0 by inputting it into the network. After three calculations, the final signal sequence sent by the transmitter is [0,1,0].

[0113] In this embodiment, for a containing For the test set of data, the bit error rate (BER) based on the Informer model in this application is expressed as:

[0114] , (13)

[0115] in This represents the number of data points correctly detected by the Informer base model.

[0116] This application also provides experimental data comparing the superiority of different models over the Informer-based model. In the experiments, different models were trained on a training dataset generated with a signal-to-noise ratio (SNR) of 40 dB, including a DNN model, a Transformer-based model, and an Informer-based model. The results showed that the BER performance of all three models decreased as environmental noise increased. Furthermore, the Informer-based and Transformer-based models in this application outperformed the DNN model in terms of BER performance. When SNR = 20 dB, the corresponding BER values ​​for the Informer-based model, Transformer-based model, and DNN model were 0.0045, 0.0062, and 0.3245, respectively. In particular, the Informer-based model outperformed the Transformer-based model on most test datasets. Moreover, especially when SNR ≥ 20 dB, the Informer-based model in this application was more stable than the Transformer-based model at different SNR values.

[0117] The experiments also compared the training time of the Informer-based model and the Transformer-based model in this application. The results show that, under the same conditions, the Informer-based model has a lower BER (0.0036), while the Transformer-based model exhibits a higher BER (0.004). Regarding training time, the Informer-based model requires less training time, approximately 22 seconds, while the Transformer-based model requires approximately 84 seconds, significantly longer than the Informer-based model. In conclusion, the detector using the Informer-based model achieves a lower BER in less training time, demonstrating its better overall performance.

[0118] Note that when using the Informer-based model, the bit error rate (BER) varies with the initial distance between the two nodes. When the initial distance between the two nodes is 9 micrometers, the BER is 0.0037, indicating high signal detection accuracy. As the initial distance increases to 10 micrometers, the BER gradually increases, indicating a decrease in detection performance. This performance change can be attributed to factors such as ISI and noise interference, which cause signal attenuation at larger initial distances. It is important to emphasize that the model's training dataset was generated based on training data with an initial distance of 9 micrometers. If the training dataset includes initial distances of 10 micrometers, 11 micrometers, and 12 micrometers, the model's BER at those corresponding initial distances may decrease further.

[0119] Figure 5 This paper presents a comparison of the convergence results during training of the Informer-based model and the Transformer-based model from this application. It can be seen that the Informer-based model is more efficient, characterized by a lower and more stable BER. In contrast, while the Transformer-based model has a lower initial BER, its decrease during training is slow and unstable. These results demonstrate the advantage of the Informer-based model in signal detection.

[0120] Figure 6This paper compares the training time BER of the Informer-based and Transformer-based models presented in this application. The Informer-based model has a lower BER of 0.0036, while the Transformer-based model has a higher BER of 0.004. In terms of training time, the Informer-based model requires less training time, approximately 22 seconds, while the Transformer-based model requires approximately 84 seconds, significantly longer than the Informer-based model. In conclusion, the Informer-based detector achieves a lower bit error rate in a shorter training time, demonstrating better overall performance.

[0121] Figure 7 This illustrates how the bit error rate (BER) changes with the initial distance between two nodes using an Informer-based model. When the initial distance is 9... At a distance of m, the bit error rate is 0.0037, indicating high signal detection accuracy. When the initial distance increases to 10... As m approaches infinity, the bit error rate gradually increases, and the detection performance deteriorates. The observed performance changes can be attributed to factors such as ISI and noise interference, which cause the signal to attenuate at a greater initial distance.

[0122] exist Figure 8 In this study, three models—a DNN model, a Transformer-based model, and an Informer-based model—were trained on a training dataset with an SNR of 40 dB. The results represent the BER performance on different test datasets with different SNR values. The BER performance of these three models decreases as environmental noise increases. Furthermore, the Informer-based and Transformer-based models outperform the DNN model in BER performance. When SNR = 20 dB, the BER values ​​are 0.0045, 0.0062, and 0.3245, corresponding to Informer-based, Transformer-based, and DNN, respectively. In particular, the Informer-based model demonstrates superior performance compared to the Transformer-based model on most test datasets. Additionally, the Informer-based model is more stable than the Transformer-based model at different SNR values.

[0123] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A signal detection method in mobile molecular communication, characterized in that, The signal detection method in the mobile molecular communication includes: A signal detection neural network model based on Informer is constructed. This model includes a position encoding module, an encoder module, a position encoding decoding module, and a decoder module. The position encoding module removes timestamp encoding and one-dimensional convolutional layers, retaining only position encoding. The encoder module includes a multi-head ProbSparse self-attention module, a feedforward neural network, and a normalization layer; compared to the traditional Informer model, residual connections are removed. The position encoding decoding module is a feedforward neural network. The decoder module is also a feedforward neural network. Generate sample data for training a signal detection neural network model; Analyze the autocorrelation coefficients of the molecular number sequences received by the receiving nanomachines in the sample data to determine the input length of the signal detection neural network model; After processing the sample data according to the determined length, the signal detection neural network model is trained. A trained signal detection neural network model is used to construct the input sequence of the MMC system to be detected, and the received transmitted signal is obtained by prediction.

2. The signal detection method in mobile molecular communication as described in claim 1, characterized in that, The location encoding module performs location encoding according to the following formula: ， ； ， ； in The dimension representing the position encoding. Indicates the position in the input sequence. This represents the hidden layer dimension of the model. Indicates the even index position in the input sequence. This indicates the odd index position in the input sequence.

3. The signal detection method in mobile molecular communication as described in claim 1, characterized in that, The analysis of the sample data includes determining the input length of the signal detection neural network model by analyzing the autocorrelation coefficients of the molecular number sequence received by the nanomachines. The number of molecules received by the receiving nanomachines in each time slot is arranged into a time series according to time. The difference between the position of each element and the position of the first element is used as the lag number to calculate the autocorrelation coefficient of each element with respect to the first element in the time series. The lag number, taken from the convergence of the correlation coefficient, is used as the input length of the signal detection neural network model.

Images