Low-power-consumption, full-stack and multifunctional wireless radio frequency sensing system and method

Abstract

The invention provides a low-power-consumption, full-stack and multifunctional wireless radio frequency sensing system and method, and the method comprises the steps: S1, constructing a phase difference-amplitude conversion circuit, and converting a phase difference into a corresponding amplitude; s2, a wake-up circuit is arranged, and when the intensity of the received radio frequency signal is higher than a preset receiving sensitivity threshold value, the circuit immediately outputs a high-level signal; developing a wake-up algorithm and operating the wake-up algorithm by a microprocessor, and determining whether to start amplitude sampling and phase difference operation by taking the level output by the wake-up circuit as input; s3, executing a phase difference calculation method, and calculating the phase difference by the microprocessor according to the amplitude output by the conversion circuit; and building a lightweight deep neural network model, taking the phase difference sequence as input data of the model, and outputting a corresponding sensing result. The device can be integrated to various mobile devices, the original power consumption of the devices is basically not increased, and the mobile devices can be endowed with the transparent radio frequency sensing capability.

Description

Technical Field

[0001] The present invention relates to the field of radio frequency sensing, and in particular, to a low-power, full-stack, multi-functional radio frequency sensing system and method. Background Art

[0002] In the past decade, with the evolution of technology, many devices have been enabled with radio frequency sensing capabilities by integrating dedicated RF sensors (such as millimeter-wave radars) or repurposing communication modules (such as WiFi chips) on mobile devices and Internet of Things (IoT) devices as radio frequency sensors. This progress has brought new application possibilities to many fields, from environmental sensing in smart homes to human body monitoring in smart healthcare, etc., demonstrating the broad prospects of radio frequency sensing technology.

[0003] However, existing radio frequency sensors have significant drawbacks in practical applications, and the power consumption problem is particularly prominent. On the one hand, the power consumption of radio frequency sensors themselves is relatively high, ranging from 100 mW to the watt level. A typical example is Google, which removed the Soli millimeter-wave radar in smartphone products after Pixel 4, and the key factor leading to this decision was its excessive power consumption. On the other hand, these sensors rely on the host device to calculate the final sensing result (such as gesture recognition) based on their output data (such as I / Q samples), but on ordinary host devices (such as smartphones, laptops, mobile robots), this calculation process itself may consume power at the watt level.

[0004] Obviously, it is of great practical significance to develop a radio frequency sensor that is "transparent" to mobile devices and IoT devices. Such an ideal sensor should have full-stack capabilities, sub-milliwatt power consumption, and multi-functionality. Full-stack capabilities mean having the complete hardware and software suite required to complete the entire radio frequency sensing process, that is, from radio frequency signal reception to sensing result calculation. Sub-milliwatt power consumption means that its power consumption is almost negligible, will not burden ordinary host devices, and can even be powered by low-capacity power sources (such as button batteries) or environmental energy harvesting. Multi-functionality requires supporting multiple radio frequency sensing tasks, such as gesture recognition, respiration sensing, positioning, etc., to meet the diverse sensing needs of the host device.

[0005] Although this direction has a bright future, there is currently no RF sensor in the market that simultaneously possesses the above three characteristics. Existing full-stack RF sensors either have a single function and are customized only for one or two specific types of sensing tasks, lacking versatility, or have a power consumption as high as the watt level. In fact, this key limitation of sub-milliwatt power consumption poses a huge challenge to achieving both versatility and full-stack characteristics simultaneously. To achieve versatility, an RF sensor needs to extract the phase difference of the RF signals received by any antenna pair, which is a prerequisite for deriving the angle of arrival (AoA) and time difference of arrival (TDoA) required for common sensing tasks. However, there is currently a lack of sub-milliwatt-level phase difference extraction methods. Existing methods often use superheterodyne, low intermediate frequency, or zero intermediate frequency receivers, whose high-power hardware components (such as voltage-controlled oscillators VCOs and low-noise amplifiers LNAs) and intensive computational processes (such as large-scale FFTs) generally result in a total power consumption in the range of 10 mW to 100 mW. Even if the phase difference can be extracted with a power consumption far lower than 1 mW, it is still extremely difficult to achieve full-stack characteristics within the remaining limited power consumption budget. Existing methods generally use neural network models to calculate sensing results, and these models require at least a system clock frequency of the order of 10 MHz to achieve real-time inference. However, even a low-power microcontroller unit (MCU) operating at this system clock frequency may still have a power consumption in the milliwatt level.

[0006] Through the retrieval of patent documents, it is found that the invention patent with the publication number CN117939605B discloses a low-power vibration signal acquisition device and its sleep wake-up method. By setting an accelerometer, an analog-to-digital converter, a microcontroller, a built-in power supply, and a low-power monitoring circuit, the analog-to-digital converter is connected between the accelerometer and the microcontroller to read the vibration signal from the accelerometer and convert it from an analog signal to a digital signal when the microcontroller is in a non-sleep state. The low-power monitoring circuit is connected between the accelerometer and the microcontroller to continuously monitor the vibration signal collected by the accelerometer when the microcontroller is in a sleep state, and wake up the microcontroller when the vibration signal exceeds the standard, so that the microcontroller starts the analog-to-digital converter and the wireless communication module to perform vibration signal monitoring. The function of this patent is limited to vibration signal monitoring, lacking multi-functional sensing and intelligent processing, without a deep neural network to implement complex sensing tasks, and the application scenario is also relatively single.

[0007] In summary, aiming at the problems of the above-mentioned existing technologies, researching a low-power, full-stack, multi-functional wireless radio frequency sensing system and method has become a key task that needs to be solved urgently at present. Summary of the Invention

[0008] Aiming at the defects in the existing technologies, the purpose of the present invention is to provide a low-power, full-stack, multi-functional wireless radio frequency sensing system and method.

[0009] A low-power, full-stack, multi-functional radio frequency sensing system provided by the present invention includes:

[0010] A phase difference - amplitude conversion module, including a phase difference - amplitude conversion circuit, which is used to convert the phase difference information of Bluetooth broadcast signals received by a pair of receiving antennas into amplitude information;

[0011] A wake-up module, including a wake-up circuit, which is used to output a rising edge level when a radio frequency signal exceeding the receiving sensitivity of the wake-up circuit appears in the environment;

[0012] A processor module, including a microprocessor, which is used to run a wake-up algorithm according to the rising edge level output by the wake-up module and decide when to perform amplitude sampling and phase difference calculation; if started, it controls an analog-to-digital converter to sample the amplitude information converted by the phase difference - amplitude conversion module, calculates the phase difference of the received signal according to the sampled amplitude information, and runs a lightweight neural network with the calculated phase difference sequence as the input to output a sensing result.

[0013] Preferably, the phase difference - amplitude conversion circuit includes a pair of receiving antennas, three Wilkinson power dividers, several capacitors and inductors for impedance matching, and three envelope detectors.

[0014] Preferably, a pair of receiving antennas are respectively connected to a Wilkinson power divider, and each Wilkinson power divider divides the received antenna signal of one path into two paths; one path of the signal output by each Wilkinson power divider is connected to a Wilkinson combiner and output after being combined by the Wilkinson combiner; the other path of the output signal of each Wilkinson power divider and the output signal of the Wilkinson combiner are respectively input into an envelope detector with an impedance matching network in front; each envelope detector detects the input signal and outputs the detected signal amplitude.

[0015] Preferably, the wake-up circuit includes a signal detector, a common-mode amplifier circuit and a comparison circuit. The signal detector includes a pair of Schottky diodes and a capacitor, and when a radio frequency signal exceeding the receiving sensitivity of the signal detector appears in the environment, the signal detector outputs a high level; the common-mode amplifier circuit includes an operational amplifier and three resistors, and the comparison circuit includes a comparator and three resistors.

[0016] Preferably, the microprocessor has an adjustable system clock frequency and supports entering a low-power sleep mode; when the microprocessor has no calculation task, it enters the low-power sleep mode; when the microprocessor performs phase difference calculation or wake-up algorithm, it adopts a first preset clock frequency; when the microprocessor runs a lightweight neural network, it adopts a second preset clock frequency higher than the first preset clock frequency.

[0017] The present invention also provides a low-power, full-stack, multi-functional radio frequency sensing method, which uses the above-mentioned low-power, full-stack, multi-functional radio frequency sensing system, and includes the following steps:

[0018] Step S1, the phase difference-amplitude conversion module converts the phase difference information of the Bluetooth broadcast signal received by a pair of receiving antennas into amplitude information;

[0019] Step S2, taking the output level of the wake-up module as the input, running the wake-up algorithm to determine whether to perform amplitude sampling and phase difference calculation;

[0020] Step S3, if amplitude sampling and phase difference calculation are to be performed, the microprocessor calculates the phase difference according to the amplitude information output by the phase difference-amplitude conversion module; and taking the calculated phase difference sequence as the input, running lightweight neural network inference to output the sensing result.

[0021] Preferably, step S2 includes the following sub-steps:

[0022] Step S2.1, construct a tuple (S 1 , S 2 ), where S 1 represents the state of the microprocessor, and S 2 represents the state of the analog-to-digital converter;

[0023] Step S2.2, when the low-power, full-stack, multi-functional radio frequency sensing system is powered on in an environment with Bluetooth broadcast signals, it enters the (REC, SLP) state. At this time, the microprocessor is activated for recording (REC), and the analog-to-digital converter is controlled to be in the sleep state (SLP); if the wake-up module detects a radio frequency signal within a predetermined duration T = 10(t f + 10ms), the algorithm assigns a unique index i to the detected signal, records its appearance time disappearance time and duration D i , and remains in the (REC, SLP) state;

[0024] When the preset timeout condition is met, the algorithm finds the set D = {i: D i ≈ 400μs}, that is, the set composed of all signal indexes with a duration close to 400μs. Then, the algorithm searches for the set and sets the disappearance time of the nearest Bluetooth broadcast signal to The algorithm then obtains the current time t and sets the countdown time Τ = max(0, T d + t f - t) for the low-power timer of the microprocessor, so that the low-power, full-stack, multi-functional radio frequency sensing system enters the (IACT, SLP) state;

[0025] Step S2.3, in the (IACT, SLP) state, the microprocessor is in the inactive state (IACT), and the analog-to-digital converter is controlled to remain in the sleep state (SLP); when the low-power timer times out and the wake-up circuit detects an RF signal, the low-power, full-stack, multi-functional radio frequency sensing system enters the (IC, SLP) state;

[0026] Step S2.4, in the (IC, SLP) state, the microprocessor is activated for interval check (IC), and the analog-to-digital converter is controlled to remain in the sleep state (SLP); the algorithm assigns an index k to the signal just detected in (IACT, SLP) and records its occurrence time If the interval the above-mentioned low-power, full-stack, multi-functional radio frequency sensing system re-enters the (IACT, SLP) state; otherwise, it enters the (DC, ACT) state;

[0027] Step S2.5, in the (DC, ACT) state, the microprocessor is activated for duration check (DC), and the analog-to-digital converter is activated to sample the amplitude output by the conversion circuit (ACT); the algorithm records the disappearance time of the signal with index k If the signal duration The algorithm resets and obtains the current time t, and sets the countdown time Τ = max(0, T d + t f - t) for the low-power timer of the microprocessor, so that the low-power, full-stack, multi-functional radio frequency sensing system enters the (IACT, SLP) state and waits for the next Bluetooth broadcast signal.

[0028] Preferably, step S3 includes:

[0029] Step S3.1, if amplitude sampling and phase difference calculation are performed, the microprocessor calculates the phase difference according to the amplitude information output by the phase difference - amplitude conversion module;

[0030] Step S3.2, taking the calculated phase difference sequence as the input, running lightweight neural network inference, and outputting the sensing result.

[0031] Preferably, step S3.1 includes: the output amplitudes of the envelope detectors connecting two Wilkinson power dividers are A 1 and A 2 , the output amplitude of the envelope detector connecting the Wilkinson combiner is A 3 , and the phase difference calculated by the formula is:

[0032]

[0033] Among them, p 1 and p 2 are the amplitude attenuation coefficients caused by the Wilkinson power divider, and ψ is the phase delay caused by the Wilkinson power divider.

[0034] Preferably, step S3.2 includes the following sub-steps:

[0035] Step S3.2.1, build a lightweight neural network model. The lightweight neural network model includes several cascaded feature extraction blocks and an inference block, and is used to process the time series input of the RF sensing task;

[0036] Step S3.2.2, build a feature extraction block. The feature extraction block consists of a one-dimensional convolutional neural network layer, a ReLU activation layer, a normalization layer, and a dropout layer in sequence; the dilation rate of the one-dimensional convolutional neural network layer of the k-th feature extraction block is set to 2 k-1 ;

[0037] Step S3.2.3, build an inference block. The inference block includes a max pooling layer, followed by a fully connected layer and a softmax layer. The output size of the max pooling layer is 1 / m of the input size, where m is the kernel size of the max pooling layer;

[0038] Step S3.2.4, apply the post-training quantization technology to the lightweight neural network, and quantize the 32-bit floating-point parameters into 8-bit integers.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The power consumption of the present invention is only in the sub-milliwatt level, realizing the full-process processing from the reception of the radio frequency signal to the calculation and output of the sensing result.

[0041] 2. The present invention supports a variety of different types of sensing tasks and can be integrated into various mobile devices. After integration, it basically does not increase the original power consumption of the device, but can endow the mobile device with "transparent" radio frequency sensing ability, enabling it to achieve low-power sensing functions in multiple scenarios, and providing a transparent RF sensing function for mobile devices and IoT devices.

[0042] 3. Compared with traditional radio frequency sensors, the present invention first integrates the characteristics of low power consumption, full-stack and multi-function. By reasonably constructing each cooperating module, a compact design is formed, reducing the dependence on hardware performance. The overall system is easy to integrate, can be conveniently installed on various devices, and is easy to operate, effectively overcoming the defects of high power consumption, single function, and complex structure in the prior art, bringing new breakthroughs to the application of radio frequency sensing technology. Description of the Drawings

[0043] Other features, objectives, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0044] Figure 1 This is a diagram of a low-power, full-stack, multi-functional radio frequency sensing system in an embodiment of the present invention;

[0045] Figure 2 This is an example diagram of the process flow of a radio frequency sensing method in an embodiment of the present invention. Detailed Embodiments

[0046] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several changes and improvements can still be made. These all fall within the protection scope of the present invention.

[0047] The present invention provides a low-power, full-stack, multi-functional radio frequency sensing system and method, which relates to the technical fields of radio frequency sensing, radio frequency circuits, Bluetooth communication, and deep neural networks. The method includes: Step S1, constructing a phase difference - amplitude conversion circuit to convert the phase difference of Bluetooth broadcast signals received by a pair of receiving antennas into corresponding amplitudes; Step S2, setting up a wake-up circuit, which outputs a high-level signal immediately when the received radio frequency signal strength is higher than its preset received sensitivity threshold; developing a wake-up algorithm and running it on a microprocessor, using the level output by the wake-up circuit as input to determine whether to start amplitude sampling and phase difference calculation; Step S3: executing a phase difference calculation method, where the microprocessor calculates the phase difference based on the amplitudes output by the conversion circuit; building a lightweight deep neural network model, using the phase difference sequence as the input data of the model, and outputting corresponding sensing results. The present invention compresses the 32-bit floating-point parameters of each layer of the model into 8-bit integer parameters to achieve model compression and reduce the inference power consumption of the neural network on the microprocessor.

[0048] Embodiment 1:

[0049] Figure 1 This is a diagram of a low-power, full-stack, multi-functional radio frequency sensing system in an embodiment of the present invention.

[0050] As Figure 1 shown, this embodiment provides a low-power, full-stack, multi-functional radio frequency sensing system, including:

[0051] A phase difference - amplitude conversion module, including a phase difference - amplitude conversion circuit, for converting the phase difference information of Bluetooth broadcast signals received by a pair of receiving antennas into amplitude information.

[0052] Specifically, the phase difference - amplitude conversion circuit includes a pair of receiving antennas, three Wilkinson power dividers, several capacitors and inductors for impedance matching, and three envelope detectors. The pair of receiving antennas are respectively connected to a Wilkinson power divider, and each Wilkinson power divider divides the received antenna signal into two paths; one path of the signal output by each Wilkinson power divider is connected to a Wilkinson combiner and output after being combined by the Wilkinson combiner; the other path of the output signal of each Wilkinson power divider and the output signal of the Wilkinson combiner are respectively input into an envelope detector with an impedance matching network in front; each envelope detector detects the input signal and outputs the detected signal amplitude.

[0053] The wake - up module includes a wake - up circuit for outputting a rising - edge level when a radio - frequency signal exceeding the receiving sensitivity of the wake - up circuit appears in the environment.

[0054] Specifically, the wake - up circuit includes a signal detector, a common - mode amplifier circuit, and a comparison circuit. The signal detector includes a pair of Schottky diodes and a capacitor, and when a radio - frequency signal exceeding the receiving sensitivity of the signal detector appears in the environment, the signal detector outputs a high level; the common - mode amplifier circuit includes an operational amplifier and three resistors, and the comparison circuit includes a comparator and three resistors. The resistors add a hysteresis function to the comparator, improving its robustness against input signal noise.

[0055] The processor module includes a micro - controller unit (MCU) for running a wake - up algorithm according to the rising - edge level output by the wake - up module and determining when to perform amplitude sampling and phase - difference calculation; if started, it controls an analog - to - digital converter (ADC) to sample the amplitude information converted by the phase - difference - amplitude conversion module, calculates the phase - difference of the received signal based on the sampled amplitude information, and runs a lightweight neural network with the calculated phase - difference sequence as the input to output a perception result.

[0056] Specifically, the micro - controller unit (MCU) has an adjustable system clock frequency and supports entering the low - power sleep mode; when the micro - controller unit (MCU) has no calculation tasks, it enters the low - power sleep mode. Although being in standby mode, stop - mode and other lower - power modes can make the MCU save more power, waking up the MCU from these modes also takes longer time. The sleep mode is the lowest - power mode that allows the MCU to be woken up in time, driving the ADC to collect amplitude information and calculate the phase - difference before the Bluetooth signal disappears; when the micro - processor performs phase - difference calculation or the wake - up algorithm, it uses a first preset clock frequency; when the micro - controller unit (MCU) runs the lightweight neural network, it uses a second preset clock frequency higher than the first preset clock frequency to support the MCU to complete neural network inference within one hundred milliseconds.

[0057] In this embodiment, the first preset clock frequency is 2 MHz, and the second preset clock frequency is 80 MHz.

[0058] Embodiment 2:

[0059] Figure 2 It is an example diagram of the wireless radio frequency sensing method flow in the embodiment of the present invention.

[0060] As Figure 2 shown, this embodiment provides a low-power, full-stack, multi-functional wireless radio frequency sensing method, which is implemented on the low-power, full-stack, multi-functional wireless radio frequency sensing system in the above embodiment. That is, those skilled in the art can understand the low-power, full-stack, multi-functional wireless radio frequency sensing method as the operation mode of the low-power, full-stack, multi-functional wireless radio frequency sensing system.

[0061] Specifically, the low-power, full-stack, multi-functional wireless radio frequency sensing method includes the following steps:

[0062] Step S1, the phase difference - amplitude conversion module converts the phase difference information of the Bluetooth broadcast signal received by a pair of receiving antennas into amplitude information.

[0063] Step S2, taking the output level of the wake-up module as the input, running the wake-up algorithm to determine whether to perform amplitude sampling and phase difference operation.

[0064] Specifically, step S2 includes the following sub-steps:

[0065] Step S2.1, construct a tuple (S 1 , S 2 ), where S 1 represents the state of the microcontroller unit (MCU), and S 2 represents the state of the analog-to-digital converter (ADC).

[0066] Step S2.2, when the low-power, full-stack, multi-functional wireless radio frequency sensing system is powered on in an environment with Bluetooth broadcast signals, it enters the (REC, SLP) state. At this time, the microcontroller unit (MCU) is activated for recording (REC), and the analog-to-digital converter (ADC) is controlled to be in the sleep state (SLP); if the wake-up module detects a radio frequency signal within a predetermined duration T = 10(t f + 10 ms), the algorithm assigns a unique index i to the detected signal, records its appearance time disappearance time and duration D i , and remains in the (REC, SLP) state.

[0067] When the preset timeout condition is met, the algorithm finds the set D = {i: D i≈400 μs}, that is, the set composed of all signal indices with a duration close to 400 μs. Then, the algorithm searches for the set and sets the disappearance time of the nearest Bluetooth broadcast signal to The algorithm then obtains the current time t and sets the countdown time Τ = max(0, T d + t f - t) for the low-power timer (LPTIM) of the microprocessor, enabling the low-power, full-stack, multi-functional radio frequency sensing system to enter the (IACT, SLP) state.

[0068] Step S2.3, in the (IACT, SLP) state, the microprocessor is in the inactive state (IACT), controlling the analog-to-digital converter (ADC) to remain in the sleep state (SLP); when the low-power timer (LPTIM) times out and the wake-up circuit detects an RF signal, the low-power, full-stack, multi-functional radio frequency sensing system enters the (IC, SLP) state.

[0069] Step S2.4, in the (IC, SLP) state, the microprocessor is activated for interval checking (IC), controlling the analog-to-digital converter (ADC) to remain in the sleep state (SLP); the algorithm assigns an index k to the signal just detected in (IACT, SLP) and records its occurrence time If the interval the above-mentioned low-power, full-stack, multi-functional radio frequency sensing system re-enters the (IACT, SLP) state; otherwise, it enters the (DC, ACT) state;

[0070] Step S2.5, in the (DC, ACT) state, the microprocessor is activated for duration checking (DC), controlling the analog-to-digital converter (ADC) to be activated to sample the amplitude output by the conversion circuit (ACT); the algorithm records the disappearance time of the signal with index k If the signal duration The algorithm resets and obtains the current time t, and sets the countdown time Τ = max(0, T d + t f - t) for the low-power timer (LPTIM) of the microprocessor, enabling the low-power, full-stack, multi-functional radio frequency sensing system to enter the (IACT, SLP) state and wait for the next Bluetooth broadcast signal.

[0071] Step S3, if amplitude sampling and phase difference calculation are performed, the microcontroller unit (MCU) calculates the phase difference based on the amplitude information output by the phase difference - amplitude conversion module; and using the calculated phase difference sequence as the input, runs lightweight neural network inference to output the sensing result.

[0072] Specifically, step S3 includes:

[0073] Step S3.1, if amplitude sampling and phase difference calculation are performed, the microcontroller (MCU) calculates the phase difference based on the amplitude information output by the phase difference - amplitude conversion module.

[0074] In this embodiment, the output amplitudes of the envelope detectors connecting two Wilkinson power dividers are A 1 and A 2 , and the output amplitude of the envelope detector connecting the Wilkinson combiner is A 3 . The phase difference calculated by the formula is:

[0075]

[0076] where p 1 and p 2 are the amplitude attenuation coefficients caused by the Wilkinson power divider, and ψ is the phase delay caused by the Wilkinson power divider; theoretically, ψ = 0. In practice, due to imperfect impedance matching, the values of p 1 , p 2 , and ψ deviate from the actual values, and their actual values can be obtained through off - line measurement.

[0077] Step S3.2, taking the calculated phase difference sequence as the input, running lightweight neural network inference, and outputting the perception result.

[0078] Specifically, step S3.2 includes the following sub - steps:

[0079] Step S3.2.1, build a lightweight neural network model. The lightweight neural network model includes several cascaded feature extraction blocks and an inference block, and is used to process the time - series input of the RF perception task.

[0080] Step S3.2.2, build a feature extraction block. The feature extraction block consists of a one - dimensional convolutional neural network (1D CNN) layer, a ReLU activation layer, a normalization layer, and a dropout layer in sequence; the 1D CNN layer helps extract time features. Compared with other model structures (such as LSTM or Transformer) for processing time - series input, the 1D CNN has less computational complexity; for the k - th feature extraction block, the dilation rate of its 1D CNN layer is set to 2 k-1, this design enables different feature extraction blocks to extract time features at different scales. Since the dilation rate of the 1D CNN layer grows rapidly in an exponential form, this design can extract global features through a small number of feature extraction blocks; the ReLU activation layer introduces non-linearity to the network, the normalization layer stabilizes the training process by reducing the dependence of the gradient on the input scale, and the Dropout layer randomly sets some model parameters to zero during the training process to prevent the model from overfitting.

[0081] Step S3.2.3, build an inference block. The inference block includes a maxpooling layer, followed by a fully connected layer and a softmax layer. The output of the maxpooling layer retains the main features of the input, and at the same time its size is 1 / m of the input size, where m is the kernel size of the maxpooling layer. This design reduces the computational amount of the fully connected layer to 1 / m of it.

[0082] Step S3.2.4, apply post-training quantization technology to the lightweight neural network to reduce the computational burden on the MCU while ensuring the inference accuracy. Specifically, dynamic quantization technology needs to be used to quantize the above lightweight neural network. This technology offline quantizes the weights of the 1D CNN layer and the fully connected layer before inference, calculates the optimal quantization coefficients for the activation layer at the same time, and dynamically quantizes the activation layer during the inference process. It should be noted that the remaining layers remain unchanged during the quantization process. Dynamic quantization allows adapting the optimal quantization coefficients during inference. The lightweight neural network quantizes 32-bit floating-point parameters into 8-bit integers, which is the highest quantization level with negligible impact on the inference accuracy.

[0083] Those skilled in the art know that in addition to implementing the system and its various devices, modules, and units provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the system and its various devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be regarded as a kind of hardware component, and the devices, modules, and units included therein for implementing various functions can also be regarded as the structures within the hardware component; the devices, modules, and units for implementing various functions can also be regarded as either software modules for implementing the method or structures within the hardware component.

[0084] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined arbitrarily.

Claims

1. A low-power, full-stack, multi-functional wireless radio frequency sensing system, characterized in that: include: A phase difference-amplitude conversion module, including a phase difference-amplitude conversion circuit, for converting phase difference information of a Bluetooth broadcast signal received by a pair of receiving antennas into amplitude information; A wake-up module, comprising a wake-up circuit, configured to output a rising edge level when a radio frequency signal exceeding the receiving sensitivity of the wake-up circuit appears in the environment; The processor module includes a microprocessor, which is used to run the wake-up algorithm according to the rising edge level output by the wake-up module to determine when to perform amplitude sampling and phase difference calculation; if started, the analog-to-digital converter is controlled to sample the amplitude information converted by the phase difference-amplitude conversion module, and the phase difference of the received signal is calculated according to the amplitude information obtained by the sampling, and the lightweight neural network is run with the calculated phase difference sequence as input to output the perception result.

2. The low-power, full-stack, multi-functional wireless radio frequency sensing system according to claim 1, characterized in that: The phase difference-amplitude conversion circuit includes a pair of receiving antennas, three Wilkinson power dividers, a number of capacitors and inductors for impedance matching, and three envelope detectors.

3. The low-power, full-stack, multi-functional wireless radio frequency sensing system according to claim 2, characterized in that: The pair of receiving antennas are respectively connected to a Wilkinson power divider, and each Wilkinson power divider divides one received antenna signal into two paths; one signal output by each Wilkinson power divider is connected to a Wilkinson combiner, and is output after being combined by the Wilkinson combiner; the other output signal of each Wilkinson power divider and the output signal of the Wilkinson combiner are respectively input into an envelope detector with an impedance matching network in front; each envelope detector detects the input signal and outputs the detected signal amplitude.

4. The low-power, full-stack, multi-functional wireless radio frequency sensing system according to claim 1, characterized in that: The wake-up circuit includes a signal detector, a same-direction amplification circuit and a comparison circuit. The signal detector includes a pair of Schottky diodes and a capacitor. When a radio frequency signal exceeding the receiving sensitivity of the signal detector appears in the environment, the signal detector outputs a high level; the same-direction amplification circuit includes an operational amplifier and three resistors, and the comparison circuit includes a comparator and three resistors.

5. The low-power, full-stack, multi-functional wireless radio frequency sensing system according to claim 1, characterized in that: The microprocessor has an adjustable system clock frequency and supports entering a low-power sleep mode; when the microprocessor has no computing tasks, it enters a low-power sleep mode; when the microprocessor performs a phase difference calculation or a wake-up algorithm, it uses a first preset clock frequency; when the microprocessor runs a lightweight neural network, it uses a second preset clock frequency that is higher than the first preset clock frequency.

6. A low-power, full-stack, multi-functional wireless radio frequency sensing method, using the low-power, full-stack, multi-functional wireless radio frequency sensing system according to any one of claims 1 to 5, characterized in that: The steps include: Step S1, a phase difference-amplitude conversion module converts phase difference information of a Bluetooth broadcast signal received by a pair of receiving antennas into amplitude information; Step S2, using the output level of the wake-up module as input, running the wake-up algorithm to determine whether to perform amplitude sampling and phase difference calculation; Step S3, if amplitude sampling and phase difference calculation are performed, the microprocessor calculates the phase difference according to the amplitude information output by the phase difference-amplitude conversion module; and uses the calculated phase difference sequence as input to run lightweight neural network reasoning and output the perception result.

7. The low-power, full-stack, multi-functional wireless radio frequency sensing method according to claim 6, characterized in that: The step S2 includes the following sub-steps: Step S2.1, construct a tuple (S1, S2), where S1 represents the state of the microprocessor and S2 represents the state of the analog-to-digital converter; Step S2.2, when the low-power, full-stack, multi-functional wireless RF sensing system is powered on in an environment with a Bluetooth broadcast signal, it enters the (REC, SLP) state, at which time the microprocessor is activated to record (REC) and controls the analog-to-digital converter to be in a dormant state (SLP); if the wake-up module is awakened within a predetermined duration T = 10 (t f +10ms) the algorithm assigns a unique index i to the detected signal and records the time of occurrence of the signal Disappearing time and duration D i , and remain in the (REC, SLP) state; When the preset timeout condition is met, the algorithm finds the set D = {i:D i ≈400μs}, that is, the set of all signal indices with a duration close to 400μs. Then, the algorithm searches for the set And set the disappearance time of the most recent Bluetooth broadcast signal to The algorithm then obtains the current time t and sets the countdown time Τ = max(0, T d +t f -t), so that the low-power, full-stack, multi-functional wireless radio frequency sensing system enters the (IACT, SLP) state; Step S2.3, in the (IACT, SLP) state, the microprocessor is in an inactive state (IACT), and the control analog-to-digital converter is still in a sleep state (SLP); when the low-power timer times out and the wake-up circuit detects an RF signal, the low-power, full-stack, multi-function wireless RF sensing system enters the (IC, SLP) state; Step S2.4, in the (IC, SLP) state, the microprocessor is activated to perform an interval check (IC), and the control analog-to-digital converter is still in the sleep state (SLP); the algorithm assigns an index k to the signal just detected in (IACT, SLP) and records the occurrence time of the signal If the interval The low-power, full-stack, multi-functional wireless radio frequency sensing system re-enters the (IACT, SLP) state; otherwise, it enters the (DC, ACT) state; Step S2.5, in the (DC, ACT) state, the microprocessor is activated to perform a duration check (DC), and the analog-to-digital converter is activated to sample the amplitude of the conversion circuit output (ACT); the algorithm records the disappearance time of the signal with index k If the signal duration Algorithm reset And get the current time t, and set the countdown time Τ = max(0,T d +t f -t), so that the low-power, full-stack, multi-functional wireless radio frequency sensing system enters the (IACT, SLP) state and waits for the next Bluetooth broadcast signal.

8. The low-power, full-stack, multi-functional wireless radio frequency sensing method according to claim 6, characterized in that: The step S3 comprises: Step S3.1, if amplitude sampling and phase difference calculation are performed, the microprocessor calculates the phase difference according to the amplitude information output by the phase difference-amplitude conversion module; Step S3.2, using the calculated phase difference sequence as input, runs lightweight neural network inference and outputs the perception result.

9. The low-power, full-stack, multi-functional wireless radio frequency sensing method according to claim 8, characterized in that: The step S3.1 includes: the output amplitudes of the envelope detectors connected to the two Wilkinson power dividers are A1 and A2 respectively, the output amplitude of the envelope detector connected to the Wilkinson combiner is A3, and the phase difference calculated by the formula is: Wherein, p1 and p2 are the amplitude attenuation coefficients caused by the Wilkinson power divider, and ψ is the phase delay caused by the Wilkinson power divider.

10. The low-power, full-stack, multi-functional wireless radio frequency sensing method according to claim 8, characterized in that: The step S3.2 includes the following sub-steps: Step S3.2.1, building a lightweight neural network model, wherein the lightweight neural network model includes several cascaded feature extraction blocks and an inference block, which is used to process the time series input of the RF sensing task; Step S3.2.2, construct a feature extraction block, which is composed of a one-dimensional convolutional neural network layer, a ReLU activation layer, a normalization layer, and a dropout layer in sequence; the expansion rate of the one-dimensional convolutional neural network layer of the kth feature extraction block is set to 2 k-1 ; Step S3.2.3, build an inference block, the inference block includes a maximum pooling layer, followed by a fully connected layer and a softmax layer, the output size of the maximum pooling layer is 1 / m of the input size, where m is the kernel size of the maximum pooling layer; Step S3.2.4, applying post-training quantization technology to the lightweight neural network to quantize the 32-bit floating-point parameters into 8-bit integers.

Images