Non-contact liquid component recognition model training method, recognition method, system and device

Abstract

The present invention provides a non-contact liquid component identification model training method, identification method, system and device. In the identification process, a millimeter wave radar is used to send a frequency sweep signal from low to high to the liquid within a set distance range, and the obtained intermediate frequency signal is divided into a plurality of frequency bands with different starting frequencies and ending frequencies and fast Fourier transform is performed respectively to obtain the reflection characteristics of the liquid target within the set range to the millimeter wave signals of multiple frequency bands. By processing the multi-band reflection characteristics through a customized neural network, the reflection characteristics related to the liquid to be detected are extracted, and the liquid components are classified and identified, which can improve the recognition accuracy of the liquid components and realize more fine-grained liquid identification. Among them, the neural network pays attention to the reflection characteristics presented by the distance unit where the liquid to be identified is located in the reflection characteristics, and can realize accurate recognition results when the liquid to be identified is placed at random positions and random angles within the set distance range.

Description

Technical Field

[0001] The present invention relates to the field of liquid sensing technology, and in particular to a non-contact liquid component recognition model training method, recognition method, system and device. Background Art

[0002] Liquid sensing technology refers to a type of technology used to detect, measure and monitor the properties, states and changes of liquids. These liquid sensing technologies usually use sensors or detectors to sense the physical or chemical properties of liquids. With the frequent occurrence of food and beverage safety cases in recent years and the needs of cases in specific places, liquid sensing technology has been widely used in production and life. This technology analyzes the characteristics of liquids through instruments to identify the type and composition of liquids. For example, it can detect small amounts of alcohol that may be contained in beverages for users who are allergic to alcohol, or identify the content of liquid components in different liquors to identify counterfeit and inferior products.

[0003] Traditional liquid detection methods usually rely on expensive and large professional equipment in the laboratory. They are generally based on the absorption and scattering characteristics of different liquids to light of different frequencies, and use spectrometers to sense liquid components such as alcohol. However, due to the high cost of professional instruments and the complexity of operation, these methods are difficult to be widely used in people's daily lives.

[0004] This year, a method for using wireless sensing technology to sense liquid components has emerged. By analyzing the changes that occur when wireless signals are reflected or penetrated in liquids, the characteristics of liquids can be analyzed. Although wireless sensing can achieve non-contact and lossless sensing, that is, identifying the liquid components in a container without opening the container or damaging the liquid sample, there are still problems such as limited recognition range, insufficient recognition granularity, and unstable recognition effect. Summary of the invention

[0005] In view of this, the embodiments of the present invention provide a non-contact liquid component identification model training method, identification method, system and device to eliminate or improve one or more defects existing in the prior art, and solve the problems of limited recognition range, insufficient recognition granularity, and high dependence on specific detection positions and angles of existing liquid wireless sensing technologies.

[0006] One aspect of the present invention provides a non-contact liquid component recognition model training method, the method comprising the following steps:

[0007] The millimeter wave radar uses a frequency-increasing sweep signal to detect the sample liquid within a set distance range, and mixes the received signal and the reflected signal to obtain an intermediate frequency signal; the intermediate frequency signal is divided into multiple frequency bands, and fast Fourier transform is performed on each frequency band to obtain a frequency domain signal, and each frequency in the frequency domain signal corresponding to each frequency band is associated with a distance unit within the set distance range; the received signal strength and phase in the frequency domain signal corresponding to each frequency band are extracted and constructed as reflection parameters for classification and identification of the sample liquid;

[0008] Constructing a training sample set, wherein the training sample set includes a plurality of samples, each sample includes the reflection parameter collected and processed by the millimeter wave radar for a single sample liquid within the set distance range, and adding a category of a corresponding sample liquid component as a label;

[0009] Acquire an initial neural network model for liquid component identification, the initial neural network model comprising a multi-distance unit feature extraction module and a liquid component classification module; the multi-distance unit feature extraction module uses a shared convolution kernel with the same learnable parameters to extract the reflection features of the sample liquid corresponding to the reflection parameters at different distance units in a translation-invariant manner; the liquid component classification module comprises a multi-layer first fully connected layer, and the liquid component classification module flattens the reflection features into one dimension, inputs the multi-layer first fully connected layer, and outputs a component identification result corresponding to the sample liquid;

[0010] The initial neural network model is trained using the training sample set, and parameters of the initial neural network model are updated based on a cross entropy loss function to obtain a liquid component recognition model.

[0011] In some embodiments, the sample liquid is placed arbitrarily within the set distance range, the set distance range is 20 cm to 2 m, and the sample liquid is provided in multiples according to different components and concentrations.

[0012] In some embodiments, the multi-distance unit feature extraction module includes a continuous two-branch module, a first activation function layer, an attention module and a temporary back-off mechanism layer; wherein the two-branch module includes a parallel main branch and a residual branch, the main branch is composed of a continuous first one-dimensional convolution layer, a first regularization layer, a second activation function layer, a second one-dimensional convolution layer and a second regularization layer, the convolution kernel of the first one-dimensional convolution layer is 1×1, and the convolution kernel of the second one-dimensional convolution layer is 1×3; the residual branch includes a continuous third one-dimensional convolution layer and a third regularization layer, and the convolution kernel of the third one-dimensional convolution layer is 1×1.

[0013] In some embodiments, the attention module averages the input feature parameters at each distance unit dimension to obtain the average response of each feature channel, and obtains the weight of each feature channel through two layers of the second fully connected layer, and finally multiplies it with the original input feature parameters and outputs it.

[0014] In some embodiments, the first fully connected layer is a three-layer stacked structure.

[0015] In some embodiments, the calculation formula of the cross entropy loss function is:

[0016]

[0017]

[0018] Where N is the number of samples, x y represents the probability that the i-th sample liquid belongs to the true category y, x c represents the probability that the i-th sample liquid belongs to category c.

[0019] On the other hand, the present invention also provides a non-contact liquid component identification method, comprising:

[0020] Based on the millimeter wave radar, a sweep frequency signal is emitted to the liquid to be tested within a set distance range, and the received signal and the reflected signal are mixed to obtain an intermediate frequency signal; the intermediate frequency signal is divided into multiple frequency bands, and fast Fourier transform is performed on each frequency band to obtain a frequency domain signal, and the received signal strength and phase in the frequency domain signal corresponding to each frequency band are extracted to construct the reflection parameter;

[0021] The reflection parameters are input into the liquid component identification model in the non-contact liquid component identification model training method, and the component identification result of the liquid to be tested is output.

[0022] On the other hand, the present invention also provides a non-contact liquid component identification system, comprising:

[0023] Millimeter wave radar is used to transmit a sweep frequency signal to the liquid to be tested within a set distance range, and mix the received signal and the reflected signal to obtain an intermediate frequency signal;

[0024] The processor is used to execute the non-contact liquid component identification method to output the component identification result of the liquid to be tested.

[0025] In some embodiments, the processor is a personal mobile terminal device, and the personal mobile terminal device is connected to the millimeter wave radar via USB or WIFI; the personal mobile terminal device is a smart phone or a tablet computer.

[0026] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program implements the steps of the above method when executed by a processor.

[0027] The beneficial effects of the present invention are at least:

[0028] The liquid component identification model training method, identification method, system and device of the present invention use millimeter wave radar to send a frequency sweep signal from low to high to the liquid within the set distance range during the detection process, divide the intermediate frequency signal into multiple frequency bands with different starting frequencies and ending frequencies and perform fast Fourier transform respectively, obtain the reflection parameters of the millimeter wave signals of multiple frequency bands within the set distance range, and capture the reflection characteristics of different distance units based on the signal in the frequency domain. The reflection characteristics related to the liquid to be detected by the multi-band reflection parameters are collected through a customized neural network, and the liquid is classified and identified to improve the recognition accuracy of the liquid components and achieve more fine-grained liquid identification. Among them, the neural network pays attention to the reflection characteristics presented by the distance unit where the liquid to be identified is located in the reflection characteristics, and can achieve accurate recognition results when the liquid to be identified is placed at random positions and random angles within the set distance range. The liquid component identification system of the present invention has high recognition accuracy and can identify 0.2% of liquid component differences; the device is small in size and supports the random placement of the target position to be detected, and is easy to deploy; it is expected to be widely used in a variety of daily life scenarios: such as genuine and fake wine identification, blood sugar concentration change monitoring, food safety detection, etc.

[0029] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention may be achieved and obtained by the structures specifically indicated in the specification and the accompanying drawings.

[0030] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of the present application, and do not constitute a limitation of the present invention. In the drawings:

[0032] Figure 1 The figure is a flow chart of a non-contact liquid composition identification method according to an embodiment of the present invention.

[0033] Figure 2 A schematic diagram of extracting multi-frequency reflection features in a non-contact liquid component identification method according to an embodiment of the present invention.

[0034] Figure 3 This is a diagram of the neural network structure used in the non-contact liquid component identification method described in one embodiment of the present invention.

[0035] Figure 4 for Figure 3 Schematic diagram of the feature extraction method with translation invariance used in the neural network.

[0036] Figure 5 Bit Figure 3 Schematic diagram of the attention module structure in the neural network. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0038] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0039] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0040] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0041] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0042] At present, there are solutions for liquid composition identification using smartphones, LEDs, RFID, Wi-Fi devices, and millimeter wave radars as low-cost liquid composition sensors. However, these existing methods have obvious limitations in practical application. Smartphone-based methods require users to use special containers, thereby relying on specific ripples generated on the surface of the liquid when it oscillates for liquid identification. The requirement for special containers limits the availability of this method. LED-based methods require the sensor to be immersed in the liquid for sensing, so they are not suitable for sealed containers. Non-contact and non-invasive liquid composition sensing using wireless signals such as RFID, UWB, and Wi-Fi signals usually requires the container to be placed in a fixed area to sense the liquid target, resulting in poor mobility and complex deployment in different scenarios. Wireless sensing methods are not only relatively rough in terms of liquid composition sensing accuracy, but are also easily affected by changes in target position, such as changes in distance and sensing angle, making it difficult to perform stable and refined liquid composition identification in real life. In daily use, the displacement and angular rotation of liquid containers are inevitable. These changes affect the reflected signal of the millimeter wave and are confused with the signal changes caused by changes in liquid composition, reducing the recognition accuracy.

[0043] This application utilizes the millimeter-wave reflection of liquid to achieve refined perception and recognition of extremely subtle differences in liquid composition, overcome the degradation of recognition performance caused by changes in the distance and perception angle between the millimeter-wave radar and the target liquid in actual use, and combines a customized neural network to capture the characteristics of liquid reflection signals at random positions and random angles to achieve stable, accurate and fine-grained liquid component recognition.

[0044] Specifically, the present invention provides a non-contact liquid component recognition model training method, which includes the following steps S101-S104:

[0045] Step S101: The sample liquid within a set distance range is detected by using a millimeter wave radar with a sweeping frequency signal with increasing frequency, and the received signal and the reflected signal are mixed to obtain an intermediate frequency signal; the intermediate frequency signal is divided into multiple frequency bands, and fast Fourier transform is performed on each of them to obtain frequency domain signals, and each frequency band corresponds to a distance unit within a set distance range associated with each frequency in the frequency domain signal; the received signal strength and phase in the frequency domain signal corresponding to each frequency band are extracted, and constructed as reflection parameters for classification and identification of the sample liquid.

[0046] Step S102: construct a training sample set, which includes multiple samples. Each sample contains reflection parameters collected and processed by millimeter wave radar for a single sample liquid within a set distance range, and adds the category of the corresponding sample liquid component as a label.

[0047] Step S103: Obtain an initial neural network model for liquid component identification, the initial neural network model includes a multi-distance unit feature extraction module and a liquid component classification module; the multi-distance unit feature extraction module uses a shared convolution kernel with the same learnable parameters to extract the reflection characteristics of the sample liquid corresponding to the reflection parameters at different distance units in a translation-invariant manner; the liquid component classification module includes multiple first fully connected layers, and the liquid component classification module flattens the reflection characteristics into one dimension, inputs the multiple first fully connected layers and outputs the component identification results of the corresponding sample liquid.

[0048] Step S104: using the training sample set to train the initial neural network model, and updating the parameters of the initial neural network model based on the cross entropy loss function to obtain a liquid component recognition model.

[0049] In step S101, first prepare sample data for training the model. Specifically, this embodiment uses FMCW millimeter wave radar to transmit millimeter waves to sample bodies within a limited spatial range. The same liquid has different dielectric properties and reflection properties for millimeter waves of different frequencies. Therefore, this embodiment uses millimeter wave radar to transmit swept frequency signals to provide diversified frequency millimeter wave signals for detection. A swept frequency signal (Sweep Signal) is a signal that continuously changes frequency, and its frequency changes within a certain frequency range with a certain step size over a period of time. Within a period of a swept frequency signal, the frequency of the signal (Tx) transmitted by the millimeter wave radar transmitting antenna gradually increases, and the signal reflection signal (Rx) received by the receiving antenna is mixed with the Tx signal to generate an intermediate frequency signal. The intermediate frequency signal is divided into multiple frequency bands to explore the reflection characteristics of the sample liquid at different frequencies. Further, the intermediate frequency signal in the time domain is converted to the frequency domain using a fast Fourier transform, and each frequency component corresponds to the reflection signal of each distance unit.

[0050] It should be noted that range bin is a term used to represent target range resolution in radar or other sensor systems. A range range is divided into discrete intervals, each of which represents a range bin. These range bins can be of equal size or of unequal distances defined according to application requirements. Each range bin contains target information within a specific range, which enables the radar system to distinguish and locate targets at distance. The size of the range bin determines the range resolution of the radar system, that is, the system's ability to distinguish between two targets at close distances.

[0051] Therefore, by transmitting a frequency sweep signal through the millimeter-wave radar and performing fast Fourier transform in frequency bands, the reflection characteristics of the sample liquid to multiple frequencies can be collected within a multi-distance unit range, thus ensuring the spatial stability and fine-grainedness of liquid component identification from a data level.

[0052] In some embodiments, the sample liquid is placed arbitrarily within the set distance range, the set distance range is 20 cm to 2 m, and multiple distances are set according to different components and concentrations.

[0053] In step S102, the training sample set is input with frequency domain reflection parameters of multiple distances and multiple frequency bands, and the components of the sample liquid are used as labels. The sample liquid can be classified not only according to different components, but also according to the proportion and concentration of the components.

[0054] In step S103, a neural network is constructed to mine the reflection features embodied by the reflection parameters in terms of distance and frequency band, and a classification task is performed. Based on the feature extraction through convolution, the multi-distance unit feature extraction module uses a shared convolution kernel and slides on the feature maps of different distance units through the attention module, so that the reflection signal features of the same liquid target at different distance units can be extracted in a translation-invariant manner. Such a design allows the same learning parameters to be used at different positions to capture the commonality of the target, focusing on the most distinguishing channels, thereby enhancing the model's detection and recognition capabilities for the target, reducing the dependence on specific distances and angles during the detection process, and ensuring that as long as the sample liquid is within the set distance range, no matter how it is placed, it can accurately focus on the reflection features presented by the sample liquid.

[0055] In some embodiments, the multi-distance unit feature extraction module includes a continuous two-branch module, a first activation function layer, an attention module and a temporary back-off mechanism layer; wherein the two-branch module includes a parallel main branch and a residual branch, the main branch consists of a continuous first one-dimensional convolution layer, a first regularization layer, a second activation function layer, a second one-dimensional convolution layer and a second regularization layer, the convolution kernel of the first one-dimensional convolution layer is 1×1, and the convolution kernel of the second one-dimensional convolution layer is 1×3; the residual branch includes a continuous third one-dimensional convolution layer and a third regularization layer, and the convolution kernel of the third one-dimensional convolution layer is 1×1.

[0056] It should be noted that the first, second, and third mentioned in the present application are not limitations on ordinal numbers, but should be understood as a distinction between similar structures at different positions.

[0057] In some embodiments, the attention module averages the input feature parameters in each distance unit dimension to obtain the average response of each feature channel, obtains the weight of each feature channel through two layers of the second fully connected layer, and finally multiplies the original input feature parameters and outputs them. This method can strengthen the neural network's attention to the feature channels that are most capable of representing liquid components.

[0058] In the liquid composition classification module, the first fully connected layer is a three-layer stacked structure. The final output is the probability that the sample liquid belongs to each category.

[0059] In step S104, in the classification task, the parameters of the model are updated using the cross entropy loss function. For example, ten liquids with alcohol contents of 0.1, 0.2, 0.3, ..., 1.0 are distinguished, and the sample liquids are randomly placed within a range of 25 cm to 60 cm from the millimeter wave radar.

[0060] One transmitting antenna of the millimeter-wave radar is used to transmit signals, and four antennas are used to receive signals, sampling 64 points in one frequency sweep cycle (chirp). First, the 64 sampling points are divided into 8 millimeter-wave signals with different starting and ending frequencies using a sliding window with a length of 42 and a step size of 3. Subsequently, the signal of each segment is converted into a reflection signal of a different distance unit using Fourier transform, and the resolution of the corresponding distance unit is 3.81 cm. Furthermore, in each frequency segment, the signal strength and phase features of the reflection signal of the distance unit corresponding to the liquid target placement range (25-60 cm) are extracted respectively. At this distance resolution, 10 distance units are obtained, and each distance unit has 64 features: 4 antennas, 8 frequency bands corresponding to the signal strength and phase. The data features with a dimension of 10*64 are input into the neural network to generate the corresponding probabilities of 10 types of liquids, and the type with the highest probability is selected as the predicted liquid component. During training, the loss function used is the cross entropy function. The calculation formula of the cross entropy loss function is:

[0061]

[0062]

[0063] Where N is the number of samples, x y represents the probability that the i-th sample liquid belongs to the true category y, x c represents the probability that the i-th sample liquid belongs to category c.

[0064] Finally, the liquid component recognition model trained in steps S101 to S104 can achieve high-precision recognition of liquid fine-grainedness, random position and angle based on the intermediate frequency signal collected by the millimeter-wave radar when emitting a sweep frequency signal within a set distance range.

[0065] On the other hand, the present invention also provides a non-contact liquid component identification method, comprising steps S201-S202:

[0066] Step S201: Based on the millimeter wave radar, a sweep frequency signal is emitted to the liquid to be tested within a set distance range, and the received signal and the reflected signal are mixed to obtain an intermediate frequency signal; the intermediate frequency signal is divided into multiple frequency bands, and fast Fourier transform is performed on each frequency band to obtain frequency domain signals, and the received signal strength and phase in the frequency domain signal corresponding to each frequency band are extracted to construct the reflection parameters.

[0067] Step S202: inputting the reflection parameters into the liquid component recognition model in the non-contact liquid component recognition model training method described in steps S101 to S104, and outputting the component recognition result of the liquid to be tested.

[0068] On the other hand, the present invention also provides a non-contact liquid component identification system, comprising:

[0069] Millimeter wave radar, used to transmit a sweep frequency signal to the liquid to be tested within a set distance range, and mix the received signal with the reflected signal to obtain an intermediate frequency signal;

[0070] The processor is used to execute the non-contact liquid component identification method described in steps S201 to S202 to output the component identification result of the liquid to be tested.

[0071] In some embodiments, the processor is a personal mobile terminal device, and the personal mobile terminal device is connected to the millimeter wave radar via USB or WIFI; the personal mobile terminal device is a smart phone, a tablet computer, etc.

[0072] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program implements the steps of the above method when executed by a processor.

[0073] The present invention is described below in conjunction with specific embodiments:

[0074] This embodiment provides a non-contact liquid component identification method, and designs a device consisting of a smartphone and a millimeter wave radar. The device transmits millimeter wave signals to the liquid target through the millimeter wave radar connected to the smartphone, receives and analyzes the millimeter wave signals reflected by the liquid, and thus realizes the identification of liquid components. It can perceive the subtle differences in liquid components in a fine-grained manner, such as a difference in alcohol content of only 0.2 degrees. At the same time, it can overcome the serious interference of changes in the distance of the liquid target and the changes in the perception angle on the recognition performance, and stably identify the liquid components under any displacement and rotation of the liquid target.

[0075] The FMCW millimeter-wave radar connected to a smartphone can emit and sense the millimeter-wave signal reflected by an object. The signal strength (RSS) of the reflected signal will vary due to differences in the target liquid properties (such as different liquid dielectric constants). Therefore, the smartphone is used to process and analyze the liquid reflection signal received by the millimeter-wave radar, and the liquid composition information is obtained based on the millimeter-wave reflection differences caused by different liquid compositions.

[0076] Specifically, refer to Figure 1 , this embodiment includes the following two parts:

[0077] 1) Signal feature extraction based on multi-frequency channels

[0078] This embodiment sends a millimeter wave signal to the target liquid through an FMCW millimeter wave radar connected to a smartphone and collects the millimeter wave signal reflected by the liquid. Next, the diverse frequency characteristics of the millimeter wave signal emitted by the FMCW millimeter wave radar are utilized, and the intermediate frequency signal sampled in an FMCW sweep signal (chirp) is divided into multiple perception frequency bands with different starting frequencies and ending frequencies. Then, a fast Fourier transform is performed on the signal of each frequency band, which is converted into a frequency domain signal, and the reflection characteristics of the specific area where the liquid is located are extracted. Finally, these features are combined to generate rich reflection characteristics of liquid targets for millimeter wave signals of multiple frequency bands. This feature contains unique reflection characteristics caused by different liquid components.

[0079] The existing technology relies on the reflection characteristics of liquid targets for single-frequency wireless signals to extract relevant features. However, this method makes it difficult to learn more fine-grained liquid component features. Inspired by the phenomenon observed in the experiment, the same liquid has different dielectric properties and reflection characteristics for millimeter waves of different frequencies. Therefore, this embodiment uses the millimeter wave signal with diversified frequencies emitted by the FMCW millimeter wave radar to obtain the reflection characteristics of liquid components for millimeter waves of different frequencies, thereby generating more discriminating fine-grained liquid component features.

[0080] like Figure 2As shown, the frequencies of the transmitting antenna and the receiving antenna gradually increase within a sweep signal cycle. This embodiment first divides the signal samples sampled within a sweep signal cycle into multiple frequency bands, each frequency band having a different starting frequency and ending frequency. For the sampled signal points within each frequency band, the time domain signal is converted into a frequency domain signal by applying a fast Fourier transform. Then, this embodiment extracts the received signal strength (RSS) of the area where the liquid is located in each frequency band to obtain the reflection characteristics of the liquid target for millimeter waves of different frequencies. Through this method, the multi-frequency signal characteristics of the FMCW millimeter wave radar itself are fully utilized. Only the sampled data obtained within a sweep signal cycle can simultaneously obtain the reflection characteristics of multiple different frequency bands without adding additional transmit and receive signal overhead. This embodiment then uses the obtained set of multi-frequency features as input information for the neural network to distinguish different liquid components.

[0081] 2) Building a neural network for liquid composition recognition

[0082] This embodiment designs a customized neural network to extract reflection features related to liquid components from multi-frequency reflection features obtained at different positions, and overcome signal interference caused by changes in the position of the liquid target and changes in the perception angle. In this neural network, this embodiment designs a feature extraction module with translational invariance, which can automatically extract the component features of liquid targets placed in different positions. These features will then be sent to the liquid component classification module to distinguish different liquid components. This embodiment trains the neural network by collecting data on liquid targets at different placement positions and rotation angles, so that the neural network can learn how to remove signal interference caused by changes in position and perception angle from a variety of data samples, thereby achieving reliable and stable liquid component identification.

[0083] like Figure 3 As shown, the neural network designed in this embodiment has two modules: a multi-range bin feature extraction module and a liquid component classification module. First, the multi-frequency reflection information of each of the multiple range bins is obtained as the input of the neural network, so that it has the ability to perceive the multiple range bins to cope with the different possible placement positions of the liquid target; the reflection information includes the signal strength and phase characteristics received by the multiple receiving antennas of the millimeter wave radar, so as to provide the reflection information and position information of the target liquid respectively.

[0084] Figure 3The multi-distance unit feature extraction module has a 6-layer structure, each of which consists of a two-branch module (consisting of a main branch and a residual branch), an activation function (ReLU), an attention module, and a dropout layer (Dropout). The main branch in the two-branch module consists of a one-dimensional convolution (convolution kernel size is 1x 1), a regularization layer (BatchNorm), an activation function, a one-dimensional convolution (convolution kernel size is 1x 3), and the residual branch consists of an additional one-dimensional convolution (convolution kernel size is 1x 1) and a regularization layer. After the two-branch structure, we use the attention module to focus on the most discriminative channels in the multi-channel feature map.

[0085] like Figure 4 As shown, this embodiment uses a shared convolution kernel with the same learnable parameters to slide on the feature maps of different distance units to extract the reflection signal features of the same liquid target at different distance units in a translation-invariant manner. This method reduces the dependence of the feature extraction process on the position of the liquid target (i.e., the distance unit where it is located) and achieves stable and reliable liquid recognition.

[0086] The structure of the attention module is as follows Figure 5 As shown, first, the average is calculated at different distance unit dimensions to obtain the average response of each feature channel, and then the weights of each channel are obtained through two fully connected layers, and finally multiplied with the input features to strengthen the neural network's attention to the feature channels that are most capable of representing liquid components.

[0087] At the end of the neural network, this embodiment uses a liquid component classification module, which first flattens the two-dimensional features (features under multiple distance and frequency channels) into one dimension, then processes the features through three stacked fully connected layers, and outputs the corresponding probability score for each liquid component.

[0088] The neural network is trained and its parameters are updated based on the classification task, and the trained model is used to perform liquid component identification.

[0089] The method provided in this embodiment breaks through the limitation of the recognition granularity of the existing wireless sensing method, and can stably and reliably achieve high-accuracy fine-grained liquid component recognition in different environments. The technical solution of the present invention can distinguish very similar liquids, such as a difference of 0.2 degrees in alcohol concentration, and overcomes the dependence on the placement of the target liquid. As a non-contact liquid recognition solution, this technology can be applied to a variety of life scenarios to achieve non-destructive liquid component analysis, such as detecting trace alcohol that may be contained in beverages, counterfeit and shoddy wines with highly deceptive properties, etc.

[0090] This embodiment enhances the ability to perceive subtle differences in liquid composition: the existing technology can only support the recognition of 1° alcohol content difference, while this embodiment can support the recognition of more subtle liquid differences, such as 0.2° alcohol content difference. In addition, the existing technology only supports a liquid target position difference of about 3cm, while this embodiment can stably perform liquid recognition under the random displacement and rotation of the liquid target of tens of centimeters.

[0091] Corresponding to the above method, the present invention also provides an apparatus / system, which includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system implements the steps of the method described above.

[0092] The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned edge computing server deployment method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0093] In summary, the liquid component identification model training method, identification method, system and device of the present invention use millimeter wave radar to send a frequency sweep signal from low to high to the liquid within a set distance range during the detection process, divide the received intermediate frequency signal into multiple frequency bands with different starting frequencies and ending frequencies and perform fast Fourier transform respectively, and obtain the reflection parameters of millimeter wave signals of multiple frequency bands within the set distance range, so as to capture the reflection characteristics of different distance units based on the signal in the frequency domain. By collecting the reflection characteristics related to the liquid to be detected through a customized neural network, and performing liquid classification and identification, more fine-grained detection can be achieved and the detection capability can be improved. Among them, the neural network uses the attention module to pay attention to the reflection characteristics presented by the distance unit where the liquid to be detected is located in the reflection characteristics, which can realize the accurate detection of the liquid to be detected at random positions and random angles within the set distance range.

[0094] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0095] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0096] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A non-contact liquid component recognition model training method, characterized in that: The method comprises the following steps: The millimeter wave radar uses a frequency-increasing sweep signal to detect the sample liquid within a set distance range, and mixes the received signal and the reflected signal to obtain an intermediate frequency signal; the intermediate frequency signal is divided into multiple frequency bands, and fast Fourier transform is performed on each frequency band to obtain a frequency domain signal, and each frequency in the frequency domain signal corresponding to each frequency band is associated with a distance unit within the set distance range; the received signal strength and phase in the frequency domain signal corresponding to each frequency band are extracted and constructed as reflection parameters for classification and identification of the sample liquid; Constructing a training sample set, wherein the training sample set includes a plurality of samples, each sample includes the reflection parameter collected and processed by the millimeter wave radar for a single sample liquid within the set distance range, and adding a category of a corresponding sample liquid component as a label; An initial neural network model for liquid component identification is obtained, wherein the initial neural network model includes a multi-distance unit feature extraction module and a liquid component classification module; the multi-distance unit feature extraction module uses a shared convolution kernel with the same learnable parameters to extract the reflection features of the sample liquid corresponding to the reflection parameters at different distance units in a translation-invariant manner; the liquid component classification module includes a plurality of first fully connected layers, and after the liquid component classification module flattens the reflection features into one dimension, inputs the plurality of first fully connected layers and outputs the component identification results of the corresponding sample liquid; The initial neural network model is trained using the training sample set, and parameters of the initial neural network model are updated based on a cross entropy loss function to obtain a liquid component recognition model; Among them, the multi-distance unit feature extraction module includes a continuous two-branch module, a first activation function layer, an attention module and a temporary retreat mechanism layer; wherein the two-branch module includes a parallel main branch and a residual branch, the main branch consists of a continuous first one-dimensional convolution layer, a first regularization layer, a second activation function layer, a second one-dimensional convolution layer and a second regularization layer, the convolution kernel of the first one-dimensional convolution layer is 1×1, and the convolution kernel of the second one-dimensional convolution layer is 1×3; the residual branch includes a continuous third one-dimensional convolution layer and a third regularization layer, and the convolution kernel of the third one-dimensional convolution layer is 1×1.

2. The non-contact liquid component recognition model training method according to claim 1, characterized in that: The sample liquid is placed arbitrarily within the set distance range, and multiple samples are provided according to different components and concentrations.

3. The non-contact liquid component identification model training method according to claim 1, characterized in that: The attention module averages the input feature parameters in each distance unit dimension to obtain the average response of each feature channel, and obtains the weight of each feature channel through two layers of the second fully connected layer, and finally multiplies it with the original input feature parameters and outputs it.

4. The non-contact liquid component recognition model training method according to claim 1, characterized in that: The calculation formula of the cross entropy loss function is: Where N is the number of samples, x y represents the probability that the i-th sample liquid belongs to the true category y, x c represents the probability that the i-th sample liquid belongs to category c.

5. A non-contact liquid component identification method, characterized in that: include: Based on the millimeter wave radar, a sweep frequency signal is transmitted to the liquid to be tested within a set distance range, and the received signal and the reflected signal are mixed to obtain an intermediate frequency signal; Divide the intermediate frequency signal into multiple frequency bands, perform fast Fourier transform on each frequency band to obtain frequency domain signals, extract the received signal strength and phase in the frequency domain signal corresponding to each frequency band to construct the reflection parameters; The reflection parameters are input into the liquid component recognition model in the non-contact liquid component recognition model training method according to any one of claims 1 to 4, and the component recognition result of the liquid to be tested is output.

6. A non-contact liquid component identification system, characterized in that: include: Millimeter wave radar, used to transmit a sweep frequency signal to the liquid to be tested within a set distance range, and mix the received signal and the reflected signal to obtain an intermediate frequency signal; A processor is used to execute the non-contact liquid component identification method as claimed in claim 5 to output the component identification result of the liquid to be tested.

7. The non-contact liquid component identification system according to claim 6, characterized in that: The processor is a personal mobile terminal device, and the personal mobile terminal device is connected to the millimeter wave radar via USB or WIFI; the personal mobile terminal device is a smart phone or a tablet computer.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images