A container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion

Abstract

This invention belongs to the field of liquid level monitoring technology, specifically relating to a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion. The passive RFID tags of this invention do not require direct contact with the liquid surface; liquid level sensing can be achieved by using a general-purpose RFID reader to collect data. Multiple spatial sampling locations, multiple operating frequencies, and RSSI and phase values ​​of multiple RFID tags are fused to obtain a holographic energy map. Through multi-frequency superposition and coherent projection processing, the influence of environmental noise and multipath interference can be effectively suppressed. A multi-scale attention dual-branch feature fusion network based on multi-frequency RFID holograms is employed to comprehensively extract and fuse the features of the original RFID signal and the holographic image, resulting in richer feature information contained in the deep learning network used as the prediction model, enabling accurate prediction of continuous liquid level height.

Description

Technical Field

[0001] This invention belongs to the field of liquid level monitoring technology, specifically relating to a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion. Background Technology

[0002] Liquid level monitoring has wide applications in industrial production and daily life, especially in scenarios such as chemical storage tanks, cooling systems, pharmaceutical containers, and food storage equipment. Real-time and accurate knowledge of the liquid level within containers is crucial for process control and ensuring production safety. Traditional liquid level detection methods typically rely on specialized sensors, such as capacitive sensors, optical laser sensors, or ultrasonic sensors. These specialized sensors usually require contact with the liquid, resulting in high costs, difficult deployment, and complex maintenance, hindering their large-scale adoption in practical applications.

[0003] With the development of radio frequency identification (RFID) technology, it has been gradually applied to the field of environmental sensing due to its advantages such as non-contact characteristics, passive properties, low cost and easy deployment. Studies have shown that RFID signals are highly susceptible to the dielectric properties of the surrounding environment. When the liquid level in the container changes, the liquid medium will change the electromagnetic environment near the tag, causing changes in the tag antenna impedance, backscatter signal strength and phase response.

[0004] Existing RFID-based liquid level detection methods mostly rely on received signal strength indicators or phase characteristics of a single frequency point as detection standards. These methods are easily affected by environmental noise, multipath effects, and random interference. They are also sensitive to changes in container position and deployment environment, making it difficult to effectively distinguish between liquid level changes and signal fluctuations caused by external environmental disturbances. This makes it difficult to perform continuous liquid level detection, resulting in insufficient detection accuracy and robustness. Summary of the Invention

[0005] The purpose of this invention is to provide a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion, which has the advantages of no direct contact, accurate prediction, low cost, easy maintenance, and strong anti-interference ability.

[0006] The specific technical solution adopted by this invention is as follows:

[0007] A container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion includes the following steps:

[0008] RFID tags are placed on the surface of the container to be tested; the RFID antenna is controlled at multiple spatial sampling positions to collect the signal strength RSSI value and phase value of each RFID tag at different liquid levels.

[0009] The acquired RSSI and phase values ​​are preprocessed and complex signals are constructed. The original hologram is then drawn using a coherent projection algorithm.

[0010] The original holograms drawn at different liquid levels were used as training data and input into a preset deep learning network to train the network.

[0011] During the real-time monitoring phase, the RSSI and phase values ​​of RFID tags are collected at multiple spatial sampling locations using a reader, and a real-time hologram is drawn based on the RSSI and phase values.

[0012] The real-time hologram and the feature vector of the real-time acquired RFID signal are input into the trained deep learning network, which then outputs the corresponding real-time liquid level prediction result.

[0013] Preferably, the step of affixing the RFID tag to the surface of the container to be tested includes:

[0014] Construct a spatial coding structure for multiple RFID tags distributed along the direction of liquid level change;

[0015] Several RFID tags are set on the surface of the container to be tested. The spatial coding structure of each multi-RFID tag covers the entire height of the liquid level change in the container, and each RFID tag is numbered sequentially along the direction of liquid level change.

[0016] Preferably, the preset deep learning network is a multi-scale attention dual-branch feature fusion network based on multi-frequency RFID holograms;

[0017] The multi-scale attention dual-branch feature fusion network includes an MLP branch and a holographic feature extraction branch.

[0018] Preferably, the MLP branch is used to process the original RFID signal feature vector composed of the received signal strength and phase information corresponding to different scanning positions, different operating frequencies and different RFID tags;

[0019] The hologram feature extraction branch is used to process the original hologram; the hologram feature extraction branch includes an initial convolutional layer, a multi-scale convolutional module, a residual attention module, and a global average pooling layer;

[0020] The features extracted from the two branches are fused and used for liquid level prediction.

[0021] Preferably, the multi-scale convolution module adopts a parallel structure, including a first convolution branch for extracting local texture features in the hologram, a second convolution branch for extracting mesoscale energy distribution features, and a third convolution branch for extracting large-scale interference fringes and global spatial distribution features.

[0022] Preferably, the stride of the first, second, and third convolutional branches of the multi-scale convolutional module is 1, and the spatial dimensions of the output feature maps are kept consistent through corresponding padding methods. The feature maps output by the first, second, and third convolutional branches are spliced ​​in the channel dimension, and channel compression and fusion are performed through 1×1 convolution to obtain a multi-scale fused feature map.

[0023] Preferably, the residual attention module includes a main branch for extracting features and a shortcut branch for preserving input information;

[0024] The output features of the main branch and the output features of the shortcut branch are added element-wise at the residual connection to form the residual fusion feature, and then the residual fusion feature is input into the attention unit.

[0025] The attention unit is used to calculate spatial weights on the input feature map to obtain the corresponding spatial attention weight map, and then weighted and fused the weight map with the input feature map.

[0026] Preferably, the step of preprocessing the acquired RSSI values ​​and phase values ​​to construct a complex signal, and then using a coherent projection algorithm to draw the original hologram includes:

[0027] The RSSI values ​​of each spatial location, each operating frequency, and each tag corresponding to the power domain are converted into linear power values, and combined with the corresponding phase values ​​to construct complex signals represented by complex numbers.

[0028] Based on the coherent projection algorithm, the propagation distance from each spatial sampling position to each pixel in the imaging plane is calculated, the corresponding propagation phase compensation factor is constructed, and the complex signal and the propagation phase compensation factor are coherently superimposed to reconstruct the holographic reconstructed field strength at each frequency.

[0029] The holographic reconstructed field strength at each frequency is superimposed at multiple frequencies to generate the corresponding holographic energy map.

[0030] Preferably, the grayscale distribution features of the holographic energy map are used to characterize the changes in the electromagnetic environment at different liquid levels, so that the trained deep learning network can output the liquid level in the container based on the grayscale distribution features.

[0031] Preferably, the training samples for training the preset deep learning network include the original hologram, the original RFID signal feature vector, and the corresponding real liquid level height label;

[0032] The trained deep learning network takes real-time holograms and real-time acquired RFID signal feature vectors as inputs and continuous liquid level height as output.

[0033] Beneficial effects

[0034] This invention employs a multi-frequency coherent projection algorithm to construct a hologram, fusing RSSI values ​​and phase values ​​from multiple spatial sampling locations, multiple operating frequencies, and multiple RFID tags to obtain a holographic energy map containing rich spatial information. Through multi-frequency superposition and coherent projection processing, the influence of environmental noise and multipath interference can be effectively suppressed. A multi-scale attention dual-branch feature fusion network based on multi-frequency RFID holograms is used to comprehensively extract and fuse the features of the original RFID signal and the holographic image, making the feature information contained in the deep learning network used as the prediction model richer, and enabling accurate prediction of continuous liquid level height.

[0035] The passive RFID tag of this invention does not require direct contact with the liquid surface, making it suitable for environments with high requirements for liquid hygiene, such as the pharmaceutical and food industries. It can also use commercial passive RFID tags, and liquid level sensing can be achieved by collecting data using general RFID readers. The deployment and maintenance costs are low, no special cleaning and maintenance procedures are required, and the cost of replacing tags is low, which reduces the complexity of deployment and the overall cost, making it suitable for large-scale promotion and application.

[0036] The holographic feature extraction branch of this invention incorporates a multi-scale convolution module and employs a parallel structure, including branches for extracting local texture features, mesoscale energy distribution features, large-scale interference fringes, and global spatial distribution features. This enables the capture of multi-level liquid level-related information from the microscopic to the macroscopic levels within the hologram. The residual attention module enhances feature regions related to liquid level changes through a spatial attention mechanism, highlighting energy focusing areas, interference fringe change areas, and feature regions near the liquid surface, while suppressing background noise interference. Holographic imaging is performed using multi-frequency signals. By fusing and utilizing multi-frequency information, the potential blind spots or signal fading issues at a single frequency are reduced, improving the robustness and reliability of liquid level detection in complex electromagnetic environments. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0038] Figure 2 This is a schematic diagram of the spatial coding structure of multiple RFID tags distributed along the liquid level direction in a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0039] Figure 3 This is a schematic diagram illustrating the data acquisition method during the network training and real-time monitoring phases of a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0040] Figure 4 The hologram is constructed using the collected data for the container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0041] Figure 5 This is a diagram of a multi-scale attention dual-branch feature fusion network structure for a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0042] Figure 6 This is a prediction result diagram after adjusting the test angle of a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention.

[0043] Figure 7 This is a prediction result diagram of the container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion proposed in this invention after adjusting the liquid medium inside the container. Detailed Implementation

[0044] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. The following embodiments are used to illustrate the present invention, but should not be used to limit the scope of protection of the present invention. The conditions in the embodiments can be further adjusted according to specific conditions. Simple improvements to the method of the present invention under the premise of the concept of the present invention are all within the scope of protection claimed by the present invention.

[0045] Please refer to Figures 1-7 This embodiment provides a container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion, including the following steps:

[0046] S1. Deployment and data collection of RFID tags;

[0047] Please refer to Figures 2-3 RFID tags are placed on the surface of the container to be tested; the RFID antenna is controlled at multiple spatial sampling positions to collect the signal strength RSSI value and phase value of each RFID tag at different liquid levels.

[0048] Construct a multi-RFID tag spatial coding structure distributed along the direction of liquid level change; set several RFID tags on the surface of the container to be tested, cover the entire liquid level change height of the container with each multi-RFID tag spatial coding structure, and number each RFID tag sequentially along the direction of liquid level change.

[0049] Please refer to Figure 3The RFID reader collects data from each numbered RFID tag, and the RFID reader transmits data to the computer via data communication. Water is poured into the container under test from zero to full, and the measuring antenna of the RFID reader is placed in multiple different spatial positions. The measuring antenna can receive a strong tag reflection signal at each of these spatial positions.

[0050] The true liquid level in the container is obtained by using a ruler or standard level gauge pre-installed next to the container. The signal strength RSSI value and phase φ value at different true liquid level heights are detected and recorded by measuring the antenna at different spatial positions and multiple measurement frequencies selected in the 900-930MHz range.

[0051] S2, Construction of the original hologram;

[0052] The acquired RSSI and phase values ​​are preprocessed and complex signals are constructed. The original hologram is then drawn using a coherent projection algorithm.

[0053] Please refer to Figure 3 RFID reader in Scanning is performed at various spatial locations, and the surface of the container under test is attached to... A commercial RFID tag, and an RFID reader working in sequence. Each measurement frequency; at each spatial location At this location, the reader checks each RFID tag. At each measurement frequency Get a set of RSSI values and phase value ;

[0054] The acquired RSSI and phase values ​​are converted into linear power values, and then the functional relationship of these values ​​as complex signals in complex number representation is as follows:

[0055]

[0056]

[0057] In the formula, and The spatial location of the RFID reader At the point where the frequency is measured The first one collected below RSSI value and phase value of each RFID tag; For RFID readers in spatial location At the point where the frequency is measured The first one collected below The linear value of the receiving power of an RFID tag. For RFID readers in spatial location At the point where the frequency is measured The first one collected below The complex signal of an RFID tag, It is the symbol for imaginary numbers;

[0058] After obtaining the complex signal, a coherent projection algorithm is used to reconstruct the complex signal onto a two-dimensional imaging plane, defining the region of interest as... A rectangular grid on a plane, discretized into... There are 100 pixels, and the coordinates of each pixel are 100. ;

[0059] For a given measurement frequency Calculate spatial location To the Euclidean distance between pixels To obtain spatial location To the The propagation distance between pixels;

[0060] Calculate the spatial location of the RFID antenna To the The distance between pixels Spatial location The antenna coordinates are written as Calculation yields :

[0061]

[0062] Construct the propagation phase compensation factor:

[0063]

[0064] In the formula, Spatial location To the Propagation phase compensation factor between pixels; It is the symbol for imaginary numbers; For spatial location To the Euclidean distance between pixels; For measuring frequency The corresponding signal wavelength, , The speed of light;

[0065] At the measurement frequency Below, by examining all spatial locations The pixel is obtained by coherently superimposing the complex signal of the RFID tag with the propagation compensation factor. The holographic reconstruction field strength at the location is:

[0066]

[0067] In the formula, For pixels Holographic reconstruction field strength at the location; For RFID readers in spatial location At the point where the frequency is measured The first one collected below The complex signal of an RFID tag; For spatial location To the Propagation phase compensation factor between pixels;

[0068] For each measurement frequency By superimposing the holographic reconstruction field strength at multiple frequencies, the value of each pixel can be obtained. Holographic energy value after multi-frequency superposition Thus, a holographic energy map can be drawn;

[0069]

[0070] In the formula, For each pixel Holographic energy resulting from the superposition of multiple frequencies; For pixels Holographic reconstruction field strength at the location;

[0071] Please refer to Figure 4 At several typical liquid levels (0, 60, 120, 180, 240 mm), the acquired RSSI and phase values ​​were preprocessed to construct complex signals, and the original holograms were drawn using a coherent projection algorithm. The individual pixels obtained from the above processing were then analyzed. Holographic energy value Logarithmic compression is performed to reduce the dynamic range of energy amplitude between different pixels, and normalization is performed to linearly map the compressed energy value to a gray value of 0 to 255, where the larger the energy value, the higher the gray value, thus obtaining a gray-scale hologram; the brightness changes in the original hologram reflect the strength differences of RFID backscatter energy distribution at different spatial locations.

[0072] Specifically, when the liquid level in the container under test changes, the liquid medium alters the electromagnetic environment near the passive RFID tag, causing changes in the RFID tag antenna impedance, backscattered signal strength, and phase response. This manifests as changes in the energy focusing area, interference fringe morphology, and local grayscale distribution in the original hologram. By using different holographic grayscale distribution characteristics corresponding to different liquid levels, effective training data for deep learning networks can be provided.

[0073] S3, Deep Learning Network Construction and Training;

[0074] The original holograms drawn at different liquid levels are used as training data and input into a pre-defined deep learning network for training. Specifically, the training samples for training the pre-defined deep learning network include the original RFID holograms, original RFID signal feature vectors, and corresponding real liquid level labels obtained in the above steps. The obtained original RFID holograms, original RFID signal feature vectors, and corresponding real liquid level labels form a dataset with real liquid level labels. The dataset is divided into a training set, a validation set, and a test set. The training set is used to learn the model parameters of the deep learning network, the validation set is used to select the model and adjust the parameters of the deep learning network, and the test set is used to evaluate the accuracy of the deep learning network. Thus, the trained deep learning network is obtained.

[0075] The deep learning network takes multi-frequency RFID holograms and original RFID signal feature vectors as inputs, continuous liquid level heights as outputs, and trains the deep learning network using the error between the actual liquid level height and the predicted liquid level height.

[0076] The original RFID signal feature vector includes RSSI values ​​at different spatial locations, for different RFID tags, and at different measurement frequencies under the actual liquid level height tag corresponding to the original RFID hologram. and phase value ;

[0077] RFID reader at Scanning is performed at various spatial locations, and the surface of the container under test is attached to... A commercial RFID tag, and an RFID reader working in sequence. Each measurement frequency; at each spatial location At this location, the reader checks each RFID tag. At each measurement frequency Get a set of RSSI values and phase value ;

[0078] Specifically, the multi-frequency RFID holograms of the test samples and the feature vectors of the original RFID signals are input into the trained deep learning network for liquid level prediction to obtain the predicted liquid level height. The predicted liquid level height is then compared with the corresponding real liquid level height, and one or more of the following indicators are used to evaluate the prediction accuracy of the deep learning network: root mean square error, mean absolute error, maximum error, or relative error.

[0079] If the error between the predicted liquid level and the actual liquid level meets the preset accuracy requirement, the deep learning network is considered to be effective. The preset accuracy requirement can be set to the point that when the root mean square error RMSE ≤ 5 mm or the relative error ≤ 3%, the deep learning network is considered to meet the liquid level prediction accuracy requirement, and the training of the deep learning network ends.

[0080] If the error between the predicted liquid level and the actual liquid level does not meet the preset accuracy requirements, the deep learning network parameters are adjusted, including one or more of the following parameters: number of network layers, number of convolutional kernels, learning rate, batch size, and number of training rounds. The training samples are then reused for iterative training until the evaluation metrics meet the preset accuracy requirements.

[0081] The preset deep learning network is a multi-scale attention dual-branch feature fusion network based on multi-frequency RFID holograms; the multi-scale attention dual-branch feature fusion network includes an MLP branch and a hologram feature extraction branch;

[0082] The MLP branch consists of an input layer, a multilayer perceptron branch, and an output layer. Adjacent layers establish a mapping relationship through a nonlinear activation function. The multilayer perceptron branch performs a layer-by-layer nonlinear transformation on the input raw RFID signal features to extract a high-dimensional feature representation that can reflect the liquid level change pattern. The output layer of the MLP branch outputs the raw RFID signal feature vector, which is represented as a deep feature representation of the raw signal features in 128 dimensions.

[0083] The input to the input layer is the original RFID signal feature vector, which consists of the received signal strength and phase information corresponding to different scanning positions, different operating frequencies, and different RFID tags. Specifically, it includes the RSSI values ​​of different spatial positions, different RFID tags, and different measurement frequencies under the real liquid level height tag corresponding to the original RFID hologram. Phase value and measurement frequency ;

[0084] The hologram feature extraction branch is used to process the original hologram; the hologram feature extraction branch includes an initial convolutional layer, a multi-scale convolutional module, a residual attention module, and a global average pooling layer;

[0085] The initial convolutional layer uses a 3×3 kernel with 16 output channels, a stride of 2, and padding of 2. After convolution, a batch normalization layer and a ReLU activation function are connected in sequence.

[0086] The multi-scale convolution module adopts a parallel structure, including a first convolution branch for extracting local texture features in the hologram, a second convolution branch for extracting mesoscale energy distribution features, and a third convolution branch for extracting large-scale interference fringes and global spatial distribution features.

[0087] The first, second, and third convolutional branches of the multi-scale convolutional module use 3×3, 5×5, and 7×7 convolutional kernels, respectively. The third convolutional branch can also use dilated convolution. The stride of the first, second, and third convolutional branches of the multi-scale convolutional module is 1. Zero padding is used to fill the boundaries of the input feature maps to ensure that the spatial size of the output feature maps of each convolutional branch remains consistent. Specifically, the 3×3 convolutional kernel corresponds to a padding width of 1 pixel, the 5×5 convolutional kernel corresponds to a padding width of 2 pixels, and the 7×7 convolutional kernel corresponds to a padding width of 3 pixels.

[0088] The feature maps output from the first, second, and third convolutional branches are concatenated along the channel dimension, and then channel compression and fusion are performed through 1×1 convolution to obtain a multi-scale fused feature map.

[0089] The residual attention module includes a main branch for feature extraction and a shortcut branch for preserving input information. The output features of the main branch and the output features of the shortcut branch are added element-wise at the residual connection to form a residual fusion feature, which is then input into the attention unit. The attention unit is used to calculate the spatial weights of the input feature map to obtain the corresponding spatial attention weight map, and then weighted and fused with the input feature map to highlight the energy focusing region, the interference fringe change region, and the feature region near the liquid surface, thereby suppressing the influence of background noise region on the liquid level prediction result.

[0090] The attention unit is used to generate a spatial attention weight map based on the input feature map. Specifically, the attention unit performs average pooling and max pooling on the input feature map along the channel direction, concatenates the two feature maps, inputs them into the convolutional layer, and maps them through the Sigmoid function to obtain the spatial attention weight map. The spatial attention weight map is then multiplied element-wise with the input feature map to obtain the enhanced feature map. This increases the weights of the energy focusing region, the interference fringe variation region, and the feature region near the liquid surface in the hologram, while reducing the influence of background noise regions on the liquid level prediction results.

[0091] The main branch consists of a 3×3 convolutional layer, a first batch of normalized layers, a ReLU activation function, a 3×3 convolutional layer, and a second batch of normalized layers.

[0092] The shortcut branch connects the multi-scale fused feature map, which serves as the input feature, to the residual connection. When the number of channels or spatial size of the input feature is the same as that of the output feature of the main branch, an identity mapping is used; when the number of channels or spatial size of the two are different, 1×1 convolution and batch normalization layer are used for dimension matching.

[0093] The main branch output and the features extracted by the shortcut branch are fused together for liquid level prediction. Specifically, the output of the attention unit is the sum of the main branch output and the shortcut branch output, and is obtained after processing by the ReLU activation function.

[0094] The two branch networks can integrate features from different dimensions, enabling the pre-defined deep learning network to capture the feature information contained in the hologram more accurately, making the entire network more resistant to interference, improving the accuracy of liquid level prediction, and reducing the possibility of human error and mechanical failure.

[0095] Furthermore, the original hologram feature extraction branch can use various feature extraction networks. When local texture needs to be extracted, a CNN network can be used; when global image relationships need to be processed, a Transformer network can be used; and when relationships between nodes need to be processed, a GNN network can be used. This satisfies various practical engineering requirements, improves the compatibility of deep learning networks, and facilitates further expansion of deep learning networks.

[0096] The 128-dimensional original signal features output from the MLP branch and the 128-dimensional image features output from the holographic feature extraction branch are fused and stitched together to form a 256-dimensional fused feature vector. The fused features are then input into a fully connected layer to output the predicted liquid level height.

[0097] S4. Real-time liquid level monitoring;

[0098] During the real-time monitoring phase, at multiple spatial sampling locations with the same measurement antenna positions as in the S2 step testing phase, the RSSI and phase values ​​of the RFID tags are collected by the reader, and a real-time hologram is drawn based on the RSSI and phase values ​​in the same manner as in the S2 step.

[0099] S5, Liquid Level Prediction;

[0100] The grayscale distribution features in the acquired real-time hologram are used to characterize the changes in the electromagnetic environment under different liquid levels. The real-time hologram is input into a trained deep learning network. Based on the grayscale distribution features, the trained deep learning network outputs the corresponding real-time liquid level prediction results, and monitors and outputs the liquid level height in the container.

[0101] RFID signals were collected at multiple known real liquid level heights and multi-frequency RFID holograms were constructed. The predicted liquid level height output by the trained deep learning network was compared with the real liquid level height, and the prediction accuracy and root mean square error index were used to evaluate the prediction accuracy.

[0102] Among them, prediction accuracy Defined as the ratio of the number of correctly predicted samples to the total number of test samples:

[0103]

[0104] in, To predict the number of correct samples, The total number of test samples; when the error between the predicted liquid level height and the actual liquid level height does not exceed the preset error threshold, the sample is considered to have made a correct prediction. The preset error threshold is usually set to 3~4mm.

[0105] Typically, when the root mean square error (RMSE) is ≤ 5 mm, or the prediction accuracy is... When the accuracy is ≥97% and the relative error is ≤3%, the prediction accuracy is considered to meet the standard. For anti-interference ability and robustness, repeated tests can be conducted under different antenna angles and different liquid media to compare the liquid level prediction error under various working conditions in order to verify the stability of the method under environmental changes.

[0106] For details, please refer to Figure 6 and Figure 7 Based on the test results, the rotation angles of the label and container were controlled to be 0°, 30°, 60° and 90° respectively, and the liquid in the container was changed to water, saline, alcohol and edible oil respectively. The above method was used to predict the liquid level and obtain the prediction accuracy and root mean square error.

[0107] Test results show that, even with changes in test conditions, the prediction accuracy of the deep learning network used as the liquid level prediction model can still be maintained above 98%, and the root mean square error can be kept below 3.22 mm.

[0108] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion, characterized in that, Includes the following steps: RFID tags are placed on the surface of the container to be tested; the RFID antenna is controlled at multiple spatial sampling positions to collect the signal strength RSSI value and phase value of each RFID tag at different liquid levels. The acquired RSSI and phase values ​​are preprocessed and complex signals are constructed. The original hologram is then drawn using a coherent projection algorithm. The original holograms drawn at different liquid levels were used as training data and input into a preset deep learning network to train the network. During the real-time monitoring phase, the RSSI and phase values ​​of RFID tags are collected at multiple spatial sampling locations using a reader, and a real-time hologram is drawn based on the RSSI and phase values. The real-time hologram and the feature vector of the real-time acquired RFID signal are input into the trained deep learning network, which then outputs the corresponding real-time liquid level prediction result.

2. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 1, characterized in that, The step of attaching the RFID tag to the surface of the container to be tested includes: Construct a spatial coding structure for multiple RFID tags distributed along the direction of liquid level change; Several RFID tags are set on the surface of the container to be tested. The spatial coding structure of each multi-RFID tag covers the entire height of the liquid level change in the container, and each RFID tag is numbered sequentially along the direction of liquid level change.

3. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 1, characterized in that, The preset deep learning network is a multi-scale attention dual-branch feature fusion network based on multi-frequency RFID holograms; The multi-scale attention dual-branch feature fusion network includes an MLP branch and a holographic feature extraction branch.

4. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 3, characterized in that: The MLP branch is used to process the original RFID signal feature vector, which consists of the received signal strength and phase information corresponding to different scanning positions, different operating frequencies, and different RFID tags. The hologram feature extraction branch is used to process the original hologram; the hologram feature extraction branch includes an initial convolutional layer, a multi-scale convolutional module, a residual attention module, and a global average pooling layer; The features extracted from the two branches are fused and used for liquid level prediction.

5. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 4, characterized in that: The multi-scale convolution module adopts a parallel structure, including a first convolution branch for extracting local texture features in the hologram, a second convolution branch for extracting mesoscale energy distribution features, and a third convolution branch for extracting large-scale interference fringes and global spatial distribution features.

6. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 5, characterized in that: The stride of the first, second, and third convolutional branches of the multi-scale convolutional module is 1, and the spatial size of the output feature map is kept consistent through corresponding padding methods. The feature maps output by the first, second, and third convolutional branches are spliced ​​in the channel dimension, and channel compression and fusion are performed through 1×1 convolution to obtain a multi-scale fused feature map.

7. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 4, characterized in that: The residual attention module includes a main branch for extracting features and a shortcut branch for preserving input information; The output features of the main branch and the output features of the shortcut branch are added element-wise at the residual connection to form the residual fusion feature, and then the residual fusion feature is input into the attention unit. The attention unit is used to calculate spatial weights on the input feature map to obtain the corresponding spatial attention weight map, and then weighted and fused the weight map with the input feature map.

8. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 1, characterized in that, The process of preprocessing the acquired RSSI and phase values ​​to construct a complex signal, and then using a coherent projection algorithm to draw the original hologram includes: The RSSI values ​​of each spatial location, each operating frequency, and each tag corresponding to the power domain are converted into linear power values, and combined with the corresponding phase values ​​to construct complex signals represented by complex numbers. Based on the coherent projection algorithm, the propagation distance from each spatial sampling position to each pixel in the imaging plane is calculated, the corresponding propagation phase compensation factor is constructed, and the complex signal and the propagation phase compensation factor are coherently superimposed to reconstruct the holographic reconstructed field strength at each frequency. The holographic reconstructed field strength at each frequency is superimposed at multiple frequencies to generate the corresponding holographic energy map.

9. A container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 8, characterized in that: The grayscale distribution features of the holographic energy map are used to characterize the changes in the electromagnetic environment at different liquid levels, so that the trained deep learning network can output the liquid level in the container based on the grayscale distribution features.

10. The container liquid level detection method based on multi-frequency RFID holographic imaging and feature fusion according to claim 1, characterized in that: The training samples for training the pre-defined deep learning network include the original hologram, the original RFID signal feature vector, and the corresponding real liquid level height label. The trained deep learning network takes real-time holograms and real-time acquired RFID signal feature vectors as inputs and continuous liquid level height as output.

Images