Iot wireless communication mixed signal detection method and system

Abstract

The application discloses a kind of Internet of Things wireless communication hybrid signal detection method and system, realize adaptive signal detection under different channel conditions, give consideration to the low complexity advantage of traditional method and the high robustness of deep learning, improve the demodulation success rate of LoRa system in interference, low signal-to-noise ratio and other harsh environments, avoid single detector performance bottleneck through dynamic routing mechanism, improve overall detection accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of signal detection, and more particularly to a method and system for detecting mixed signals in Internet of Things (IoT) wireless communication. Background Technology

[0002] With the surge in the number of Internet of Things (IoT) devices, the unlicensed Industrial, Scientific and Medical (ISM) bands are becoming increasingly congested. Many devices are using random access protocols (such as ALOHA-based LoRaWAN), which do not require base station coordination of resource allocation. While this simplifies the protocol stack and reduces power consumption, it leads to severe co-channel overlap transmission. This uncoordinated access causes significant colored interference, especially LoRa-to-LoRa interference, in high-density scenarios such as cities. When multiple devices use the same spreading factor (SF), performance degrades sharply, causing packet collisions and throughput loss.

[0003] Traditional methods such as matched filters and energy detection assume that the channel is linear and the noise is Gaussian distributed, which cannot effectively handle non-Gaussian, structured co-channel interference, CFO and sampling frequency offset (SFO) caused by oscillator mismatch, and symbol boundary misalignment caused by time synchronization error. In high interference scenarios, the interference signal can mask the spectral peak of the target symbol, resulting in incoherent detection bit error rate.

[0004] In summary, there is a need for a method and system for detecting mixed signals in IoT wireless communication to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for detecting mixed signals in Internet of Things (IoT) wireless communication, aiming to solve the aforementioned problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting mixed signals in Internet of Things (IoT) wireless communication, comprising the following steps:

[0007] Step S1: Signal modeling and dataset generation. Using the LoRa modulation scheme, the spreading factor and signal bandwidth are set, a random symbol sequence vector is created, signal samples are generated, Gaussian white noise is superimposed to obtain complex baseband I / Q signal samples, the complex baseband I / Q signal samples are normalized, and labels are assigned according to the true value of each transmitted signal to construct a labeled dataset.

[0008] Step S2: Multimodal signal preprocessing, normalized signal samples generate different forms of input, dynamically select and output the preprocessed data of the corresponding mode;

[0009] Step S3: Construct two parallel detection branches. By selecting the position corresponding to the spectral peak of the de-linearly frequency-modulated signal as the demodulated symbol, an incoherent detector is constructed. A deep learning detector is constructed by using a pre-trained CNN to classify the pre-processed signal.

[0010] Step S4: Intelligent switching mechanism: acquire received LoRa symbols, perform curve-type frequency modulation operation and FFT transformation, extract amplitude spectrum, and select detector;

[0011] Step S5: Dynamic routing and symbol decision. Input the frequency domain characteristics of the received signal into the switching discriminator to determine the category of the current channel conditions. Based on the discrimination result, dynamically route the signal to the corresponding detection branch, trigger the selected branch to perform symbol demodulation operation, output the final symbol decision result, and complete the signal detection.

[0012] Optionally, the dataset generation in step S1 is performed in the following manner:

[0013] Step A1: Set LoRa modulation parameters, select the spreading factor, set the signal bandwidth, and determine the sampling frequency;

[0014] Step A2: Generate a random symbol sequence, create a random integer vector of length n, and generate a LoRa symbol index sequence representing the symbols to be sent;

[0015] Step A3: Modulate each symbol into the corresponding linear frequency modulation spread spectrum complex baseband waveform, add Gaussian white noise, divide the synthesized received signal by the target signal power, and output the normalized complex baseband I / Q waveform;

[0016] Step A4: Extract a segment of the length of a single LoRa symbol from the received waveform to correspond to a complete, time-aligned LoRa symbol. Bind the original transmitted value to each received symbol sample. Repeat the above process to construct a structured labeled dataset.

[0017] Optionally, the multimodal signal preprocessing in step S2 is performed in the following manner:

[0018] Step B1: Separate the real and imaginary parts of the normalized complex basis band LoRa symbol samples, and stack I and Q according to the channel dimension to form a real value matrix;

[0019] Step B2: Generate FFT modes, calculate the discrete Fourier transform for the received symbols, take the amplitude spectrum and normalize it to form a one-dimensional vector;

[0020] Step B3: Generate STFT mode, configure STFT parameters, perform frame windowing on the received symbols, calculate short-time Fourier transform frame by frame to obtain complex time-frequency matrix, take the squared amplitude to generate spectrum diagram, and form a two-dimensional matrix;

[0021] Step B4: Dynamically select and output the corresponding modality. Based on deployment requirements or experimental configuration, specify the target modality, receive a normalized symbol, and generate the data of the selected modality in parallel or on demand.

[0022] Optionally, the incoherent detector is constructed in step S3 in the following way:

[0023] Step C1: Perform curve-type frequency modulation, construct a standard LoRa downlink frequency modulation signal, generate a reference inverse linear frequency modulation signal, and multiply the received signal by the conjugate of the reference inverse linear frequency modulation signal;

[0024] Step C2: Frequency domain energy detection, perform discrete Fourier transform on the delinearized frequency modulated signal, and calculate the power spectral density;

[0025] Step C3: Sign determination, find the location of the maximum power spectrum, and output the index as the final demodulation result.

[0026] Optionally, the deep learning detector is constructed in step S3 in the following way:

[0027] Step D1: Select the input mode and preprocess the signal. Perform a discrete Fourier transform on the original received signal, take the amplitude and normalize it.

[0028] Step D2: Pre-train the CNN model, configure the input layer, stack multiple convolutional modules in sequence, connect a fully connected layer after the convolutional layer, feed the output into the Softmax activation function, and complete the training using the LoRa dataset;

[0029] Step D3: Input the preprocessed amplitude into the trained CNN model, obtain the output probability distribution, and use the class with the highest probability as the estimated symbol.

[0030] Optionally, the detector selection in step S4 is performed in the following manner:

[0031] Step E1: Receive a complete LoRa baseband symbol sample, perform a fast Welsh transition on the received symbol, and calculate the amplitude spectrum;

[0032] Step E2: Compare the maximum value of the amplitude spectrum obtained from the amplitude spectrum with a preset threshold; if the maximum value of the amplitude spectrum is greater than the preset threshold, it is determined that there is strong interference and a deep learning detector is selected; otherwise, an incoherent detector is selected.

[0033] Optionally, the dynamic routing and symbol decision in step S5 are performed in the following manner:

[0034] Step F1: Based on the judgment result, send a routing command to the main control logic;

[0035] Step F2: Dynamically route to the corresponding detection branch. If it is an incoherent detector branch, generate a reference downlink frequency modulation signal, calculate the conjugate multiplication of the received signal and the reference inverse linear frequency modulation signal, perform a Fourier transform, and take the position of the maximum power spectrum as the symbol estimate.

[0036] If it is a deep learning detector branch, the FFT amplitude is calculated for the received signal, the Softmax output is obtained through the CNN model, and the class with the highest probability is taken as the symbol estimate;

[0037] Step F3: Output the final symbol decision and complete the detection. Output an integer symbol estimate and pass it to the upper-layer protocol stack for data recovery and complete signal detection.

[0038] An IoT wireless communication mixed signal detection system, employing the aforementioned IoT wireless communication mixed signal detection method, includes a signal modeling and dataset generation module, a multimodal signal preprocessing module, an incoherent detector module, a deep learning detector module, an intelligent switching discrimination module, and a dynamic routing and symbol decision module;

[0039] The signal modeling and dataset generation module is used to set LoRa modulation parameters, generate random LoRa symbol sequences and modulate them into complex baseband I / Q waveforms, superimpose Gaussian white noise to simulate the real channel environment, normalize the signal, and bind real labels to each sample to build a structured labeled training and testing dataset.

[0040] The multimodal signal preprocessing module is used to convert the normalized complex baseband signal into multiple input modes to support the needs of different detectors. It can dynamically select and output data of a specified mode according to system configuration or deployment requirements.

[0041] The incoherent detector module is used to achieve low-complexity LoRa demodulation based on traditional signal processing methods, generate a reference inverse linear frequency modulation signal, multiply the received signal with the conjugate of the reference signal to complete the delinear frequency modulation, perform FFT on the result, calculate the power spectrum, and select the peak position of the power spectrum as the demodulation symbol.

[0042] The deep learning detector module is used to classify pre-processed signals using a pre-trained convolutional neural network. The input is usually an FFT amplitude spectrum or other modal features. The CNN extracts frequency domain / time frequency features through multiple convolutions, and the output is processed by Softmax to obtain the probability distribution of symbol classes. The one with the highest probability is selected as the symbol estimate.

[0043] The intelligent switching discrimination module is used to evaluate the channel quality of the currently received signal in real time, calculate the FFT amplitude spectrum of the received LoRa symbol, compare the maximum value of the amplitude spectrum with the preset threshold, and perform branch selection.

[0044] The dynamic routing and symbol decision module receives instructions from the switching decision module, routes the signal to the selected detection branch, triggers the corresponding detector to perform symbol demodulation, outputs the final integer symbol estimate, and passes the result to the upper-layer protocol stack to complete the data recovery process.

[0045] The beneficial effects of this invention are:

[0046] 1. In this invention, adaptive signal detection under different channel conditions is realized, which combines the advantages of low complexity of traditional methods with the high robustness of deep learning, improves the demodulation success rate of LoRa system in harsh environments such as interference and low signal-to-noise ratio, avoids the performance bottleneck of single detector through dynamic routing mechanism, and improves the overall detection accuracy and reliability.

[0047] 2. In this invention, the maximum value of the FFT amplitude spectrum is compared with a threshold to determine whether to enable the deep learning detector, thereby achieving rapid online assessment of channel quality. When there is strong narrowband interference (causing spectral spikes), it automatically switches to the deep learning branch with strong anti-interference capability, avoiding the use of high-overhead deep models in simple scenarios, saving computation and energy consumption. The switching logic is simple and efficient, and easy to implement in hardware.

[0048] 3. In this invention, the aforementioned methods are integrated into a complete IoT wireless communication hybrid signal detection system, forming a software and hardware collaborative solution that can be engineered and implemented. It is applicable to LoRa gateways, edge nodes or terminal devices, improves communication reliability in actual deployment, and has both high performance and low power consumption characteristics. It meets the IoT's requirements for wide coverage, low power consumption and high connectivity, and provides an architectural reference for supporting other modulation methods in the future. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0050] Figure 2 This is a schematic diagram of step S1 of the present invention.

[0051] Figure 3 This is a schematic diagram of step S2 of the present invention.

[0052] Figure 4 This is a schematic diagram of the process for constructing an incoherent detector in step S3 of the present invention.

[0053] Figure 5 This is a schematic diagram of the process of constructing a deep learning detector in step S3 of the present invention.

[0054] Figure 6 This is a schematic diagram of step S4 of the present invention.

[0055] Figure 7This is a schematic diagram of step S5 of the present invention.

[0056] Figure 8 This is a schematic diagram of a system structure according to the present invention. Detailed Implementation

[0057] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] like Figures 1 to 7 As shown, an IoT wireless communication mixed-signal detection method includes the following:

[0059] Step S1: Signal modeling and dataset generation. Using the LoRa modulation scheme, the spreading factor and signal bandwidth are set, a random symbol sequence vector is created, signal samples are generated, Gaussian white noise is superimposed to obtain complex baseband I / Q signal samples, the complex baseband I / Q signal samples are normalized, and labels are assigned according to the true value of each transmitted signal to construct a labeled dataset.

[0060] Specifically:

[0061] Set LoRa modulation parameters: Select the spreading factor (SF), for example, SF = 7; set the signal bandwidth (B), for example, B = 125 kHz; determine the sampling rate, such as fs = 1 MHz for high-precision simulation, which can be resampled to the Nyquist rate of 125 kHz later.

[0062] Generate a random symbol sequence: Create a random integer vector M = {m1, m2, ..., mn} of length n, where each mi ∈ {0, 1, ..., 2}. SF -1}; This sequence represents the LoRa symbol index to be sent.

[0063] Modulation generates the target LoRa baseband signal:

[0064] Each symbol mi is modulated into a corresponding linear frequency modulated spread spectrum complex baseband waveform using an open-source LoRa simulator (such as MATLAB); the duration of each symbol is Ts = (2 SF ) / B seconds, corresponding to N=2 SF One sampling point.

[0065] Introducing channel impairments: Adding additive white Gaussian noise to form complex Gaussian noise n(t)~CN(0,σ2); controlling the noise power σ2 according to the target SINR, such as uniform distribution in [-15,15]dB.

[0066] This step can also optionally superimpose co-channel interference signals: generate another independent LoRa interference signal, which can use the same or different SF, such as SF=7 or SF=8; apply a random phase offset фI∈[0°,360°] and a random time delay τ to the interference signal; set the interference power pI according to the preset INR (e.g., uniformly distributed in [-30,30]dB); and add the interference signal to the target signal.

[0067] Normalized received signal: Divide the synthesized received signal by the target signal power (set ps=1) to ensure that the target signal amplitude is consistent; output normalized complex baseband I / Q waveform r(t)=xs(t)+xI(t)+n(t).

[0068] Pruning and synchronizing symbols: Assuming the receiver has completed preamble detection and coarse synchronization, a segment of a single LoRa symbol length is extracted from the received waveform; ensuring that each sample corresponds to a complete, time-aligned LoRa symbol.

[0069] Assigning supervision labels: Each received symbol sample is assigned its original transmitted value mi as a category label; the total number of labels is 2. SF For example, when SF=7, there are 128 classes.

[0070] Build the complete dataset:

[0071] Repeat the above process to generate a large-scale dataset, such as 200,000 to 2,000,000 samples; ensure that SINR and INR are evenly distributed within the specified range, covering noise-limited and interference-limited scenarios; divide the dataset proportionally, such as 90% training set and 10% test set.

[0072] Step S2: Multimodal signal preprocessing, normalized signal samples generate different forms of input, dynamically select and output the preprocessed data of the corresponding mode;

[0073] Specifically:

[0074] Separate the real and imaginary parts: extract the in-phase component of the signal: I=Re{r(t)}; extract the quadrature component of the signal: Q=Im{r(t)}.

[0075] Concatenate into a two-dimensional array: Stack I and Q along the channel dimension to form a real-valued matrix of shape [2×N]; output this matrix as the input to IQ-CNN.

[0076] Performing a Fast Fourier Transform (FFT): Calculating the Discrete Fourier Transform for the received symbol r(t):

[0077] R(f) = FFT{r(t)}.

[0078] Take the amplitude spectrum and normalize it: Calculate the amplitude: |R(f)|; it can be further divided by the length for energy normalization; form a one-dimensional vector, output XFFT=|R(f)|∈RN; as the input of FFT-CNN.

[0079] Configure STFT parameters: Window function: Hamming window g(.); Window length W (e.g., 64 or 128 points);

[0080] Frame shift step size L (e.g., 50% overlap); FFT points are usually equal to window length.

[0081] Calculate the short-time Fourier transform: window r(t) frame by frame and perform FFT frame by frame; obtain the complex time-frequency matrix X(ω,p), where p is the time frame index.

[0082] The squared amplitude is used to generate a spectrogram: S(p,f)=|X(ω,p)|2; a two-dimensional matrix XSTFT∈RF×T (e.g., 65×8) is formed and used as the input of STFT-CNN, which is regarded as an "image".

[0083] Specify the target modality based on deployment requirements or experimental configuration:

[0084] If you want to balance low complexity and high accuracy, choose FFT mode.

[0085] If studying the characteristics of the original waveform → select the I / Q mode;

[0086] If you want to explore the advantages of time-frequency structure, then choose STFT mode.

[0087] In the HybNet system, the main detection branch uses a fixed modality (such as FFT-CNN); the switching discriminator (selector CNN) uses FFT amplitude (i.e., frequency domain modality) as its input to ensure lightweight and consistency; it does not rely on STFT or I / Q, reducing the complexity of the switching logic.

[0088] The preprocessing pipeline is implemented by receiving a normalized symbol and generating data for the selected modality in parallel or on demand; the results are then fed into the corresponding CNN model for inference.

[0089] Step S3: Construct two parallel detection branches. By selecting the position corresponding to the spectral peak of the de-linearly frequency-modulated signal as the demodulated symbol, an incoherent detector is constructed. A deep learning detector is constructed by using a pre-trained CNN to classify the pre-processed signal.

[0090] Construct an incoherent detector in the following way:

[0091] Step C1: Perform curve-type frequency modulation, construct a standard LoRa downlink frequency modulation signal, generate a reference inverse linear frequency modulation signal, and multiply the received signal by the conjugate of the reference inverse linear frequency modulation signal;

[0092] Step C2: Frequency domain energy detection, perform discrete Fourier transform on the delinearized frequency modulated signal, and calculate the power spectral density;

[0093] Step C3: Sign determination, find the location of the maximum power spectrum, and output the index as the final demodulation result.

[0094] Construct a deep learning detector in the following ways:

[0095] Step D1: Select the input mode and preprocess the signal. Perform a discrete Fourier transform on the original received signal, take the amplitude and normalize it.

[0096] Step D2: Pre-train the CNN model, configure the input layer, stack multiple convolutional modules in sequence, connect a fully connected layer after the convolutional layer, feed the output into the Softmax activation function, and complete the training using the LoRa dataset;

[0097] Step D3: Input the preprocessed amplitude into the trained CNN model, obtain the output probability distribution, and use the class with the highest probability as the estimated symbol.

[0098] Step S4: Intelligent switching mechanism: acquire received LoRa symbols, perform curve-type frequency modulation operation and FFT transformation, extract amplitude spectrum, and select detector;

[0099] The detector is selected in the following ways:

[0100] Step E1: Receive a complete LoRa baseband symbol sample, perform a fast Welsh transition on the received symbol, and calculate the amplitude spectrum;

[0101] Step E2: Compare the maximum value of the amplitude spectrum obtained from the amplitude spectrum with a preset threshold; if the maximum value of the amplitude spectrum is greater than the preset threshold, it is determined that there is strong interference and a deep learning detector is selected; otherwise, an incoherent detector is selected.

[0102] Step S5: Dynamic routing and symbol decision. Input the frequency domain characteristics of the received signal into the switching discriminator to determine the category of the current channel conditions. Based on the discrimination result, dynamically route the signal to the corresponding detection branch, trigger the selected branch to perform symbol demodulation operation, output the final symbol decision result, and complete the signal detection.

[0103] Through the following methods:

[0104] Step F1: Based on the judgment result, send a routing command to the main control logic;

[0105] Step F2: Dynamically route to the corresponding detection branch. If it is an incoherent detector branch, generate a reference downlink frequency modulation signal, calculate the conjugate multiplication of the received signal and the reference inverse linear frequency modulation signal, perform a Fourier transform, and take the position of the maximum power spectrum as the symbol estimate.

[0106] If it is a deep learning detector branch, the FFT amplitude is calculated for the received signal, the Softmax output is obtained through the CNN model, and the class with the highest probability is taken as the symbol estimate;

[0107] Step F3: Output the final symbol decision and complete the detection. Output an integer symbol estimate and pass it to the upper-layer protocol stack for data recovery and complete signal detection.

[0108] The method of this invention realizes adaptive signal detection under different channel conditions, combining the advantages of low complexity of traditional methods with the high robustness of deep learning, improving the demodulation success rate of LoRa system in harsh environments such as interference and low signal-to-noise ratio, and avoiding the performance bottleneck of a single detector through dynamic routing mechanism, thereby improving the overall detection accuracy and reliability.

[0109] Based on the comparison between the maximum value of the FFT amplitude spectrum and the threshold, it is determined whether to enable the deep learning detector, so as to realize the rapid online assessment of channel quality. When there is strong narrowband interference (causing spectral spikes), it automatically switches to the deep learning branch with strong anti-interference ability, avoiding the use of high-overhead deep models in simple scenarios, saving computation and energy consumption. The switching logic is simple and efficient and easy to implement in hardware.

[0110] like Figure 8 As shown, an IoT wireless communication mixed signal detection system employs the IoT wireless communication mixed signal detection method, including a signal modeling and dataset generation module, a multimodal signal preprocessing module, an incoherent detector module, a deep learning detector module, an intelligent switching discrimination module, and a dynamic routing and symbol decision module;

[0111] The signal modeling and dataset generation module is used to set LoRa modulation parameters, generate random LoRa symbol sequences and modulate them into complex baseband I / Q waveforms, superimpose Gaussian white noise to simulate the real channel environment, normalize the signal, and bind real labels to each sample to build a structured labeled training and testing dataset.

[0112] The multimodal signal preprocessing module is used to convert the normalized complex baseband signal into multiple input modes to support the needs of different detectors. It can dynamically select and output data of a specified mode according to system configuration or deployment requirements.

[0113] The incoherent detector module is used to achieve low-complexity LoRa demodulation based on traditional signal processing methods, generate a reference inverse linear frequency modulation signal, multiply the received signal with the conjugate of the reference signal to complete the delinear frequency modulation, perform FFT on the result, calculate the power spectrum, and select the peak position of the power spectrum as the demodulation symbol.

[0114] The deep learning detector module is used to classify pre-processed signals using a pre-trained convolutional neural network. The input is usually an FFT amplitude spectrum or other modal features. The CNN extracts frequency domain / time frequency features through multiple convolutions, and the output is processed by Softmax to obtain the probability distribution of symbol classes. The one with the highest probability is selected as the symbol estimate.

[0115] The intelligent switching discrimination module is used to evaluate the channel quality of the currently received signal in real time, calculate the FFT amplitude spectrum of the received LoRa symbol, compare the maximum value of the amplitude spectrum with the preset threshold, and perform branch selection.

[0116] The dynamic routing and symbol decision module receives instructions from the switching decision module, routes the signal to the selected detection branch, triggers the corresponding detector to perform symbol demodulation, outputs the final integer symbol estimate, and passes the result to the upper-layer protocol stack to complete the data recovery process.

[0117] Integrating the aforementioned methods into a complete IoT wireless communication hybrid signal detection system forms an engineerable hardware and software collaborative solution suitable for LoRa gateways, edge nodes, or terminal devices. This improves communication reliability in actual deployments, combines high performance with low power consumption, meets the IoT's requirements for wide coverage, low power consumption, and high connectivity, and provides an architectural reference for supporting other modulation methods in the future.

[0118] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting mixed signals in Internet of Things (IoT) wireless communication, characterized in that, Includes the following steps: Step S1: Signal modeling and dataset generation. Using the LoRa modulation scheme, the spreading factor and signal bandwidth are set, a random symbol sequence vector is created, signal samples are generated, Gaussian white noise is superimposed to obtain complex baseband I / Q signal samples, the complex baseband I / Q signal samples are normalized, and labels are assigned according to the true value of each transmitted signal to construct a labeled dataset. Step S2: Multimodal signal preprocessing, normalized signal samples generate different forms of input, dynamically select and output the preprocessed data of the corresponding mode; Step S3: Construct two parallel detection branches. By selecting the position corresponding to the spectral peak of the de-linearly frequency-modulated signal as the demodulated symbol, an incoherent detector is constructed. A deep learning detector is constructed by using a pre-trained CNN to classify the pre-processed signal. Step S4: Intelligent switching mechanism: acquire received LoRa symbols, perform curve-type frequency modulation operation and FFT transformation, extract amplitude spectrum, and select detector; Step S5: Dynamic routing and symbol decision. Input the frequency domain characteristics of the received signal into the switching discriminator to determine the category of the current channel conditions. Based on the discrimination result, dynamically route the signal to the corresponding detection branch, trigger the selected branch to perform symbol demodulation operation, output the final symbol decision result, and complete the signal detection.

2. The IoT wireless communication mixed signal detection method according to claim 1, characterized in that, The dataset in step S1 is generated in the following way: Step A1: Set LoRa modulation parameters, select the spreading factor, set the signal bandwidth, and determine the sampling frequency; Step A2: Generate a random symbol sequence, create a random integer vector of length n, and generate a LoRa symbol index sequence representing the symbols to be sent; Step A3: Modulate each symbol into the corresponding linear frequency modulation spread spectrum complex baseband waveform, add Gaussian white noise, divide the synthesized received signal by the target signal power, and output the normalized complex baseband I / Q waveform; Step A4: Extract a segment of the length of a single LoRa symbol from the received waveform to correspond to a complete, time-aligned LoRa symbol. Bind the original transmitted value to each received symbol sample. Repeat the above process to construct a structured labeled dataset.

3. The IoT wireless communication mixed signal detection method according to claim 2, characterized in that, The multimodal signal preprocessing in step S2 is performed in the following manner: Step B1: Separate the real and imaginary parts of the normalized complex basis band LoRa symbol samples, and stack I and Q according to the channel dimension to form a real value matrix; Step B2: Generate FFT modes, calculate the discrete Fourier transform for the received symbols, take the amplitude spectrum and normalize it to form a one-dimensional vector; Step B3: Generate STFT mode, configure STFT parameters, perform frame windowing on the received symbols, calculate short-time Fourier transform frame by frame to obtain complex time-frequency matrix, take the squared amplitude to generate spectrum diagram, and form a two-dimensional matrix; Step B4: Dynamically select and output the corresponding modality. Based on deployment requirements or experimental configuration, specify the target modality, receive a normalized symbol, and generate the data of the selected modality in parallel or on demand.

4. The IoT wireless communication mixed signal detection method according to claim 1, characterized in that, In step S3, the incoherent detector is constructed in the following way: Step C1: Perform curve-type frequency modulation, construct a standard LoRa downlink frequency modulation signal, generate a reference inverse linear frequency modulation signal, and multiply the received signal by the conjugate of the reference inverse linear frequency modulation signal; Step C2: Frequency domain energy detection, perform discrete Fourier transform on the delinearized frequency modulated signal, and calculate the power spectral density; Step C3: Sign determination, find the location of the maximum power spectrum, and output the index as the final demodulation result.

5. The IoT wireless communication mixed signal detection method according to claim 1, characterized in that, In step S3, the deep learning detector is constructed in the following way: Step D1: Select the input mode and preprocess the signal. Perform a discrete Fourier transform on the original received signal, take the amplitude and normalize it. Step D2: Pre-train the CNN model, configure the input layer, stack multiple convolutional modules in sequence, connect a fully connected layer after the convolutional layer, feed the output into the Softmax activation function, and complete the training using the LoRa dataset; Step D3: Input the preprocessed amplitude into the trained CNN model, obtain the output probability distribution, and use the class with the highest probability as the estimated symbol.

6. The IoT wireless communication mixed signal detection method according to claim 1, characterized in that, The detector selection in step S4 is performed in the following manner: Step E1: Receive a complete LoRa baseband symbol sample, perform a fast Welsh transition on the received symbol, and calculate the amplitude spectrum; Step E2: Compare the maximum value of the amplitude spectrum obtained from the amplitude spectrum with a preset threshold to make a judgment; If the maximum value of the amplitude spectrum is greater than the preset threshold, it is determined that there is strong interference, and a deep learning detector is selected; otherwise, an incoherent detector is selected.

7. The IoT wireless communication mixed signal detection method according to claim 1, characterized in that, The dynamic routing and symbol decision in step S5 are performed in the following way: Step F1: Based on the judgment result, send a routing command to the main control logic; Step F2: Dynamically route to the corresponding detection branch. If it is an incoherent detector branch, generate a reference downlink frequency modulation signal, calculate the conjugate multiplication of the received signal and the reference inverse linear frequency modulation signal, perform a Fourier transform, and take the position of the maximum power spectrum as the symbol estimate. If it is a deep learning detector branch, the FFT amplitude is calculated for the received signal, the Softmax output is obtained through the CNN model, and the class with the highest probability is taken as the symbol estimate; Step F3: Output the final symbol decision and complete the detection. Output an integer symbol estimate and pass it to the upper-layer protocol stack for data recovery and complete signal detection.

8. An IoT wireless communication mixed-signal detection system, employing the IoT wireless communication mixed-signal detection method as described in any one of claims 1-7, characterized in that, It includes a signal modeling and dataset generation module, a multimodal signal preprocessing module, an incoherent detector module, a deep learning detector module, an intelligent switching discrimination module, and a dynamic routing and symbol decision module; The signal modeling and dataset generation module is used to set LoRa modulation parameters, generate random LoRa symbol sequences and modulate them into complex baseband I / Q waveforms, superimpose Gaussian white noise to simulate the real channel environment, normalize the signal, and bind real labels to each sample to build a structured labeled training and testing dataset. The multimodal signal preprocessing module is used to convert the normalized complex baseband signal into multiple input modes to support the needs of different detectors. It can dynamically select and output data of a specified mode according to system configuration or deployment requirements. The incoherent detector module is used to achieve low-complexity LoRa demodulation based on traditional signal processing methods, generate a reference inverse linear frequency modulation signal, multiply the received signal with the conjugate of the reference signal to complete the delinear frequency modulation, perform FFT on the result, calculate the power spectrum, and select the peak position of the power spectrum as the demodulation symbol. The deep learning detector module is used to classify pre-processed signals using a pre-trained convolutional neural network. The input is usually an FFT amplitude spectrum or other modal features. The CNN extracts frequency domain / time frequency features through multiple convolutions, and the output is processed by Softmax to obtain the probability distribution of symbol classes. The one with the highest probability is selected as the symbol estimate. The intelligent switching discrimination module is used to evaluate the channel quality of the currently received signal in real time, calculate the FFT amplitude spectrum of the received LoRa symbol, compare the maximum value of the amplitude spectrum with the preset threshold, and perform branch selection. The dynamic routing and symbol decision module receives instructions from the switching decision module, routes the signal to the selected detection branch, triggers the corresponding detector to perform symbol demodulation, outputs the final integer symbol estimate, and passes the result to the upper-layer protocol stack to complete the data recovery process.

Images