Artificial intelligence-based anti-interference method for surface acoustic wave passive wireless sensing system

Abstract

The application discloses an anti-interference method of a surface acoustic wave passive wireless sensing system based on artificial intelligence. Firstly, radio frequency signal intensity data of various surface acoustic wave echo signals including noise, normal response and interference under different frequencies are acquired, and different labels are given to the radio frequency signal intensity data. The radio frequency signal intensity data is sequentially preprocessed, standardized, one-hot encoded and an additional dimension is added, so that the data becomes three-dimensional. A one-dimensional CNN neural network classification model of a training attention mechanism is constructed. The model comprises a one-dimensional convolution layer used for extracting time-frequency features of the radio frequency signal intensity data, and the size of the convolution kernel is 1, the padding is 1 and the step is 1, and an attention mechanism module used for weighted average of the time-frequency features extracted by the one-dimensional convolution layer. The one-dimensional CNN neural network model after training is used to output the category of a new surface acoustic wave echo signal. When the anti-interference is realized, the time domain and frequency domain information of the echo signal can be reserved, the loss of information is reduced, and the signal reconstruction quality and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and in particular to an anti-interference method for a surface acoustic wave passive wireless sensing system based on artificial intelligence. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Surface acoustic waves (SAWs) are elastic sound waves excited by piezoelectric materials and propagating along the surface of a solid. Because their sound wave transmission characteristics (such as sound velocity and resonant frequency) are sensitive to the surrounding environment, and because devices can be integrated with antennas, passive wireless multi-parameter sensing systems can be realized without an external power supply and with strong environmental adaptability. A SAW sensing system consists of a radio frequency reader (RF reader...). Figure 1 The system consists of two parts: a readout unit and a SAW sensing unit. The sensing unit comprises a sensor and a sensing antenna. The sensor further includes piezoelectric materials, interdigital transducers (IDTs), and reflective gratings. The working principle of the SAW passive wireless system is similar to that of radar. First, the reader transmits a radio frequency interrogation signal, and the sensing unit receives the signal through the sensing antenna. Figure 1 The sensor receives an interrogation signal via an IDT (Inductively Coupled Transmission Device), excites the surface acoustic wave (SAW) transmission resonance, and returns a radio frequency response signal containing information about the surrounding environment (such as temperature). The reader then receives and analyzes the response signal to obtain information about the environment around the sensor. Passive wireless sensing technology based on SAW resonators has been widely used in industrial environmental monitoring due to its advantages such as simple structure, no power supply required, no wiring required, and remote reading capability. For example, it is used to monitor the temperature, pressure, and stress state of mechanical equipment in real time and to detect abnormal conditions promptly. In lithium-ion battery energy storage systems, prolonged operation leads to performance degradation, manifested as capacity decay and increased internal resistance. These degradation factors directly affect the battery's lifespan and safety. By using SAW-based passive wireless sensing technology, real-time monitoring of the temperature and strain of lithium-ion batteries can be achieved, allowing for timely detection of battery abnormalities and proactive measures. This not only extends battery life but also improves system safety and reliability. Despite the continuous development of surface acoustic wave (SAW) passive wireless sensing systems in areas such as high temperature resistance and multi-parameter sensing, the challenge of improving the anti-interference capability of SAW device echo signals to enhance detection accuracy and improve system performance and reliability in complex environments and scenarios such as man-made electromagnetic interference remains unsolved.

[0004] The main anti-interference method of the existing SAW passive wireless sensing system is mainly to optimize the design of the SAW sensor and optimize the data detection algorithm. In terms of the physical structure of the sensor, the resistance of the sensor to specific interference is improved by designing special surface acoustic wave structures, such as improving the piezoelectric substrate, the interdigital transducer, the reflective grating array, and the antenna or using special materials such as metal organic frameworks. In terms of optimizing the detection algorithm, repeated queries and improved autocorrelation algorithms are used to remove white noise to improve the anti-interference ability, or autocorrelation algorithms and Fourier transform are used to eliminate the influence of initial phase on frequency estimation results to enhance the anti-interference ability. However, they all mainly use traditional signal processing algorithms such as autocorrelation algorithms, and rely on specific assumptions, and have low processing ability for spectral interference from other wireless devices or environment and for non-linear distortion. There is also a signal matrix denoising method through singular value decomposition technology to obtain better anti-interference performance. This method does not depend on specific assumptions, but has high computational complexity for large-scale data, low data processing speed and efficiency. In addition, there is a method of using compressed sensing to directly improve the detection accuracy. This method requires less data and has low computational complexity, which improves the data processing speed and efficiency, but cannot resist electromagnetic interference. There is also an anti-interference algorithm based on adaptive Kalman filtering, but this method depends on the statistical properties of the measurement noise, and cannot classify noise, response and interference, and can only remove incorrect data. Finally, there is a method of using ICA to classify interference signals and response signals. The ICA method has ambiguity in scale and order, that is, the amplitude and order of the separated source signals are uncertain, and has high requirements for non-Gaussian signals. Using neural network technology can better solve the above problems, but the existing genetic algorithm optimized BP neural network (Back Propagation Neural Networks, BPNN) using artificial intelligence algorithm is used to fit and analyze the measurement data of the SAW sensor to verify the accuracy of the data, and the original physical echo signal of the SAW sensor is not analyzed. When encountering abnormal measurement data, only discarding processing can be performed, and effective real signals cannot be measured in an interference environment. SUMMARY

[0005] To solve the technical problems in the background art, the present application proposes a surface acoustic wave passive wireless sensing system anti-interference method based on artificial intelligence, which can effectively distinguish signals and interference, has high robustness, practicality, universality and flexibility, and minimizes information loss.

[0006] The object of the present application is achieved by the following technical solution:

[0007] A surface acoustic wave passive wireless sensing system anti-interference method based on artificial intelligence, comprising the following steps:

[0008] Step one: Obtain the radio frequency signal intensity data of various surface acoustic wave echo signals including noise, normal response and interference at different frequencies, and assign different labels to the surface acoustic wave echo signals;

[0009] Step two: first, pre-process the radio frequency signal intensity data of various surface acoustic wave echo signals so that all data lengths are consistent; then, standardize the data and perform one-hot encoding on the standardized data; finally, add an extra channel dimension to the two-dimensional data to make the data three-dimensional, i.e., the dimensions of the three-dimensional data are sample number, feature number, and channel number, where the channel number is 1;

[0010] Step three: build a one-dimensional CNN neural network classification model with attention mechanism, which includes one one-dimensional convolution layer and an attention mechanism module; the one-dimensional convolution layer is used to extract the time-frequency features of the radio frequency signal intensity data, with a convolution kernel size of 1, padding of 1, and stride of 1, to enhance the nonlinear ability of feature extraction; the attention mechanism module performs weighted averaging on the time-frequency features extracted by the one-dimensional convolution layer;

[0011] Train the one-dimensional CNN neural network classification model using the data obtained in step two;

[0012] Step four: collect new surface acoustic wave echo signals, process them in the same way as in step two, and input them into the trained one-dimensional CNN neural network model, which outputs the category of the new surface acoustic wave echo signals.

[0013] Further, the label of the surface acoustic wave echo signal at different frequencies including only noise is 1, the label of the surface acoustic wave echo signal containing noise and normal response is 0, and the label of the surface acoustic wave echo signal containing only noise and interference is 2.

[0014] Further, in step two, the radio frequency signal intensity data of various surface acoustic wave echo signals is pre-processed to make all data lengths consistent, which specifically includes:

[0015] For missing data, prefer to use mode filling, and use median filling if mode does not exist;

[0016] For data whose length exceeds the preset threshold, perform truncation processing.

[0017] Further, the one-dimensional CNN neural network classification model includes one one-dimensional convolution layer and an attention mechanism module, and further includes a second one-dimensional convolution layer, a max pooling layer, a first dropout layer, a third one-dimensional convolution layer, a second dropout layer, a flattening layer, a first fully connected layer, a third dropout layer, and a second fully connected layer.

[0018] The max pooling layer is used to reduce the feature dimension while retaining the most important information;

[0019] The dropout layer one, the dropout layer two and the dropout layer three are used to randomly turn off a part of neurons in the training process to prevent model overfitting;

[0020] The flattening layer is used to convert the multi-dimensional time-frequency features into a one-dimensional feature vector;

[0021] The fully connected layer one and the fully connected layer two are used for advanced representation and classification of the surface acoustic wave echo signal.

[0022] An artificial intelligence-based surface acoustic wave passive wireless sensing system, comprising a surface acoustic wave sensing unit, a reader, a radio frequency cable, and an upper computer;

[0023] The surface acoustic wave sensing unit is used to receive an inquiry signal and transmit a surface acoustic wave signal with environmental information;

[0024] The reader is used to receive the signal received by the surface acoustic wave sensing unit and convert it into processable data;

[0025] The radio frequency cable is used to realize signal transmission between the hardware of the reader and the reader antenna;

[0026] The upper computer deploys a one-dimensional CNN neural network classification model of the trained attention mechanism, which is used to realize an artificial intelligence-based surface acoustic wave passive wireless sensing system anti-interference method.

[0027] Further, the upper computer comprises a data acquisition module, a data processing module, a data display module, a model updating module, and an intelligent power detection module;

[0028] The data acquisition module is used to cooperate with the surface acoustic wave sensing unit to realize the acquisition of the surface acoustic wave signal with environmental information;

[0029] The data processing module is used to preprocess the radio frequency signal intensity data of various surface acoustic wave echo signals to make all data lengths consistent; then standardize the data and perform one-hot encoding on the standardized data; finally, add an additional channel dimension to the two-dimensional data to make the data three-dimensional, i.e., the dimensions of the three-dimensional data are sample number, feature number, and channel number, wherein the channel number is 1;

[0030] The intelligent power detection module deploys a one-dimensional CNN neural network classification model of the trained attention mechanism to analyze the signal received by the SAW sensing unit and identify different types of noise, normal response, and interference signals based on the learned time domain features;

[0031] The data display module visualizes the model performance by using a confusion matrix, an ROC curve and a t-SNE method;

[0032] The model updating module is used for inputting the processed surface acoustic wave echo signals under a new environment into the trained one-dimensional CNN neural network classification model, retraining the one-dimensional CNN neural network classification model, and updating the parameters of the trained one-dimensional CNN neural network classification model to adapt to the environmental changes and new interference in long-term operation.

[0033] The beneficial effects of the present application are as follows:

[0034] Compared with the existing method of improving the output signal-to-noise ratio and the anti-interference ability by using autocorrelation algorithm, compressed sensing and other traditional algorithms, the one-dimensional CNN neural network classification model with attention mechanism can more intelligently identify and avoid signals in a complex interference environment, and has higher anti-interference ability; compared with the previous method which needs to manually design and select features, the present method can automatically learn the features of the echo signal, reduces the manual intervention and error, improves the processing efficiency and accuracy of the signal, and has better performance in real-time processing of large-scale data; in addition, compared with the existing method of directly removing noise or interference to realize anti-interference, the present application can also retain the time domain and frequency domain information of the echo signal, reduce the loss of information, and improve the reconstruction quality and reliability of the signal. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a schematic diagram of the existing SAW passive wireless sensing system.

[0036] Figure 2 is a flow chart of the anti-interference method of the surface acoustic wave passive wireless sensing system based on artificial intelligence according to the embodiment of the present application.

[0037] Figure 3 is a whole system diagram according to the embodiment of the present application.

[0038] Figure 4 is a schematic diagram of different types of signals according to the embodiment of the present application; wherein, Figure 4 a in the above formula (1) represents the surface acoustic wave echo signals of different frequencies including only noise, Figure 4 b in the above formula (2) represents the surface acoustic wave echo signals containing noise and normal response, Figure 4 c in the above formula (3) represents the surface acoustic wave echo signals containing only noise and interference.

[0039] Figure 5 is a schematic diagram of a multi-layer neural network model structure according to the embodiment of the present application.

[0040] Figure 6is a model loss function and model accuracy diagram shown by an embodiment of the present application.

[0041] Figure 7 is a model confusion matrix heat map shown by an embodiment of the present application.

[0042] Figure 8 is a model multi-classification ROC curve diagram shown by an embodiment of the present application.

[0043] Figure 9 is a t-SNE visualization diagram of three different layers in the model at the 99th iteration of training shown by an embodiment of the present application, wherein Figure 9 a in the above is a visualization of the first convolutional layer, Figure 9 b in the above is a visualization of the third convolutional layer, Figure 9 c in the above is a visualization of the last fully connected layer.

[0044] Figure 10 is a classification test result of the time domain signal algorithm after multiple sampling under a single frequency shown by an embodiment of the present application, wherein, Figure 10 a in the above indicates a noise test classification result, Figure 10 b in the above indicates a response test classification result, Figure 10 c in the above indicates an interference test classification result.

[0045] Figure 11 is a frequency domain sweep test result of the algorithm shown by an embodiment of the present application, wherein, Figure 11 a in the above indicates a noise sweep test result, Figure 11 b in the above indicates an interference sweep, Figure 11 c in the above indicates a response under an interference environment. DETAILED DESCRIPTION

[0046] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0047] It should be noted that the following detailed description is illustrative only and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application pertains.

[0048] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be further understood that the terms "comprise" and / or "include" as used herein indicate the presence of the features, steps, operations, devices, components and / or combinations thereof.

[0049] It should be noted that the flow diagrams and block diagrams in the drawings are representations of the architectures, functions, and operations of possible implementations of methods and systems according to various embodiments of the present disclosure. It should be noted that each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which includes one or more executable instructions for implementing the specified logical functions (s) in the various embodiments. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the drawings. For example, two blocks noted in succession can in fact be executed substantially concurrently or sometimes in reverse order, depending on the functionality involved. It should also be noted that each block in the flow diagrams and / or block diagrams and combinations of blocks in the flow diagrams and / or block diagrams can be implemented by hardware-based systems that perform specified functions or operations, or can be implemented using a combination of dedicated hardware and computer instructions.

[0050] As introduced in the background, in the research of anti-interference technology of SAW passive wireless sensor system, the main focus is on improving the physical structure of the sensor and optimizing the signal processing algorithm. In terms of the physical structure of the sensor, by designing special SAW structures, such as improving the piezoelectric substrate, interdigital transducer, reflective grating array and antenna or using special materials such as metal organic framework, the resistance of the sensor to specific interference is improved. In terms of signal processing algorithm, researchers try to use autocorrelation algorithm, compressed sensing and other traditional algorithms to improve the output signal-to-noise ratio, and thus improve the anti-interference ability of the system. These methods improve the anti-interference performance of the system to some extent, but still have some shortcomings.

[0051] 1. The improvement of the physical structure of the sensor usually increases the manufacturing cost and complexity of the sensor, and this method has limited adaptability to complex and variable interference environment.

[0052] 2. Although traditional signal processing algorithms can improve the signal-to-noise ratio to some extent, these methods usually need manual design and selection of features, and the processing effect on complex and nonlinear interference signals is limited, and the performance is limited when processing large-scale data in real time.

[0053] 3. Traditional signal processing algorithms generally directly remove noise or interference, and cannot retain more time domain and frequency domain information of the echo signal, resulting in more information loss, and the reconstruction quality and reliability of the signal need to be improved.

[0054] 4. The adaptability and flexibility of traditional anti-interference methods are poor, and it is difficult to adapt to the characteristics of different SAW devices and echo signals, and the application range and scene adaptability are limited.

[0055] To address this, the present invention provides an artificial intelligence anti-interference method for a surface acoustic wave passive wireless sensing system capable of distinguishing noise, response, and interference. This method includes the design of an artificial intelligence algorithm based on a one-dimensional convolutional neural network (1D-CNN) with an attention mechanism. The design comprises the following steps: signal acquisition, data partitioning and calibration, data preprocessing, model design, model evaluation, model optimization, signal classification and recognition, and algorithm deployment and application. The 1D-CNN neural network model learns the characteristics of noise, normal response, and interference signals by training echo signals of different frequencies in the dataset, achieving accurate signal classification. Therefore, during frequency domain sweeping, it can not only measure effective real signals in interference environments but also preserve the original noise or interference information.

[0056] The present invention addresses the above-mentioned problems using the following embodiments.

[0057] Example 1

[0058] like Figure 2 As shown, this embodiment provides an anti-interference method for a surface acoustic wave passive wireless sensing system based on artificial intelligence. This embodiment uses the application of this method to a surface acoustic wave passive wireless temperature sensing system as an example for illustration.

[0059] An anti-interference method for a passive wireless sensing system based on artificial intelligence using surface acoustic waves includes the following steps:

[0060] Step 1: Signal Acquisition.

[0061] The radio frequency signal intensity data of various surface acoustic wave echo signals, including noise, normal response and interference, are acquired at different frequencies, and different labels are assigned to the surface acoustic wave echo signals.

[0062] Firstly, through, as... Figure 3 The SAW sensor unit shown works with a reader to acquire signals. First, the reader, along with an RF cable and antenna, transmits an RF interrogation signal. The SAW sensor unit receives the interrogation signal through its sensing antenna, excites the surface acoustic wave propagation resonance via an IDT, and returns an RF echo signal containing ambient environmental information (such as temperature). The reader then receives the echo signal, thus acquiring the signal as shown. Figure 4 The image shown only includes surface acoustic wave echo signals at different frequencies of noise. Figure 4 (a) contains surface acoustic wave echo signals with noise and normal response. Figure 4 (b) contains only surface acoustic wave echo signals with noise and interference. Figure 4c) in the noise signal is the echo signal at the non-SAW resonance frequency; the normal response is the echo signal at the SAW resonance frequency; the interference signal is simulated by the radio frequency signal generator and connected to the antenna to emit and collect by the reader. During the collection process, the reader uses the power detection method to perform frequency domain sweep, thereby obtaining the radio frequency signal strength indicator (RSSI) data of the echo at each frequency, which is stored in the.csv table format for subsequent reading and processing. In special application scenarios, different interference factors such as high temperature, high humidity and electromagnetic interference also need to be considered, and specific wireless and passive signal collection schemes are adopted to adapt to these extreme conditions.

[0063] Data division and calibration:

[0064] The collected data is accurately labeled using the pandas library tool of python Figure 4 The different types of RSSI table data at different frequencies are accurately labeled as shown in the figure. The surface acoustic wave echo signal including only noise is assigned a label 1, the surface acoustic wave echo signal containing noise and normal response is assigned a label 0, and the surface acoustic wave echo signal containing only noise and interference is assigned a label 2. The data set is divided into a training set, a test set and a validation set, wherein a certain proportion of data is used as a training set for model learning and training, the remaining data is partly used as a test set for verifying the generalization ability of the model and the performance in actual application, and the other part is used as a validation set for adjusting the hyperparameters of the model and comparing different models during the training process.

[0065] Step two: data preprocessing.

[0066] First, the radio frequency signal strength data of various surface acoustic wave echo signals is preprocessed to make all data lengths consistent; then the data is standardized and the standardized data is one-hot encoded; finally, an additional channel dimension is added to the two-dimensional data to make the data three-dimensional. Specifically, the dimensions of the three-dimensional data are sample number, feature number and channel number, and the channel number is 1.

[0067] To ensure that all data lengths are consistent to facilitate standardized data input, the present application fills the mode of the data missing value, and in the case where the mode does not exist, the median is used for filling to avoid the influence of extreme values at the beginning of the data; for data whose length exceeds the preset threshold, truncation processing is performed. Next, after all data lengths are uniform, the present application performs data standardization, that is, the mean and standard deviation are calculated, and then the Standardization transformation is performed, where x is the feature value, μ is the mean of the feature, σ is the standard deviation of the feature, z is the standardized value, and the final distribution has a unit variance and 0 mean feature. Then, the data is One-Hot encoded to help the model better understand that different categories are mutually exclusive. Finally, in order to adapt the data to the convolutional neural network, an additional channel dimension is added to the original two-dimensional data to be correctly processed by the network, where the number of channels is 1. In addition, signal amplification, filtering and normalization and other technical means can be used to improve the quality of the signal.

[0068] Step three: model design, training and testing

[0069] In view of the superiority of 1D-CNN in processing sequence data such as waveforms, time series, signals and audio, the model in the present application is constructed based on 1D-CNN. By sliding the convolution filter on the one-dimensional time series data, 1D-CNN can effectively capture the time pattern and extract detailed information from the SAW time domain curve. The attention mechanism can highlight key features in large-scale time series data and give higher weights to important signal parts, thereby improving the accuracy of feature extraction. For the data of SAW sensor, noise and interference signals may mask key features, and the attention mechanism can help the model focus on more representative parts and reduce the influence of interference.

[0070] Therefore, the structure of the one-dimensional CNN neural network classification model with attention mechanism adopted by the present application adopts an alternating sequence of stacking 1D-CNN convolution layers, attention mechanism modules, MaxPooling layers, Dropout layers and fully connected layers. The model parameters can be learning rate, convolution kernel size, filter number, pooling window size and Dropout ratio, but are not limited thereto. The training and testing of the model can be performed by local computing resources, or large-scale parallel training and optimization can be performed through a cloud computing platform. The implementation of the 1D-CNN neural network model can be based on TensorFlow, Keras, PyTorch, Matlab, R and other scientific computing and machine learning libraries.

[0071] As shown in Figure 5 In the present embodiment, the model includes:

[0072] The first layer is a one-dimensional convolutional layer one, used for preliminary extraction of time-frequency features of the radio frequency signal intensity data, the size of the convolution kernel is 1, the padding is 1, and the stride is 1. This layer can enhance the nonlinear ability of feature extraction and capture more complex patterns by applying a convolution operation with a convolution kernel size of 1, padding of 1, and stride of 1 on one-dimensional data combined with an activation function for nonlinear transformation.

[0073] The second layer is an attention mechanism module, used to improve the model's attention to key features. This module calculates the weighted average of time-frequency features, enabling the model to better focus on important features and improve overall performance and anti-interference ability.

[0074] The third layer is a one-dimensional convolutional layer two, also used for extracting time-frequency features.

[0075] The fourth layer is a max-pooling layer, used to reduce the dimension of features while retaining the most important information.

[0076] The fifth layer is a dropout layer one, used to randomly "turn off" a portion of neurons during the training process to help prevent model overfitting.

[0077] The sixth layer is a one-dimensional convolutional layer three, also used for extracting time-frequency features.

[0078] The seventh layer is a dropout layer two, also used to prevent model overfitting.

[0079] The eighth layer is a flattening layer, used to convert multi-dimensional time-frequency features into a one-dimensional feature vector to provide input for the fully connected layer.

[0080] The ninth layer is a fully connected layer one, using a ReLU activation function for nonlinear transformation. This layer has 128 neurons, which can further process the extracted features into higher-level representations.

[0081] The tenth layer is a dropout layer three, also used to prevent model overfitting.

[0082] The eleventh layer is a fully connected layer two using a Softmax activation function for three classification of noise, SAW response and interference of the surface acoustic wave echo signal. The Softmax function can convert the model output into a probability distribution, so that each class corresponds to a probability value, and the final classification result is the class with the highest probability value, thereby completing the classification task of the surface acoustic wave signal. The number of neurons in the output layer matches the number of target classification categories, i.e. the three categories of noise, response and interference, ensuring that the model can make predictions for all possible signal categories. This model can effectively extract the time-frequency features of the signal and enhance the model's ability to recognize different frequency and intensity noise and interference. To cope with changes in complex environmental interference or noise patterns, real-time monitoring and adaptive adjustment strategies can also be designed so that the model can automatically adjust parameters according to real-time changes in the environment, thereby adapting to new signal characteristics.

[0083] Model training and testing:

[0084] The present embodiment uses a gradient-based optimization algorithm to update the weights of the model. This algorithm finds the path of minimizing the loss by calculating the gradient of the loss function with respect to the model parameters. To further optimize the training process, the present invention then uses an adaptive learning rate adjustment strategy, which can dynamically adjust the learning rate according to the model's performance during training, thereby speeding up the convergence process and preventing premature convergence to a local optimal solution during training.

[0085] Model evaluation:

[0086] First, the performance of the model on the test set is preliminarily evaluated by two-dimensional indicators of loss and accuracy, as shown in Figure 6 Loss is a numerical value that measures the degree of inconsistency between the model's predicted results and the true labels. Ideally, a good model will have a decreasing loss value during training, as shown in Figure 6 , because the model is constantly optimizing its parameters through learning training data to reduce prediction errors. Therefore, the downward trend of the loss value usually reflects the effectiveness of the model's learning. Accuracy reflects the proportion of samples that the model predicts correctly out of the total number of samples. During training, an increase in accuracy means that the model's ability to identify data has improved, as shown in Figure 6 , the accuracy gradually increases with the increase of the number of iterations until it becomes relatively stable, proving that its ability to identify data has been continuously improved during training and has finally reached a relatively stable state.

[0087] Parameter optimization of the model:

[0088] According to the results of the previous model evaluation and the performance of the model on the validation set, further adjust the hyperparameters of the model, such as learning rate, regularization coefficient, number of layers, number of neurons, etc., to find the optimal model setting. In this way, it returns to the model design and repeats the cycle iteration until the model performance no longer improves significantly.

[0089] Step four: Collect new acoustic surface wave echo signals, and after the same processing as step two, input the one-dimensional CNN neural network model after training, and the model outputs the category of the new acoustic surface wave echo signal.

[0090] Save the trained one-dimensional CNN neural network classification model as an h5 file and deploy it to the host computer so that it can analyze and identify different types of signals received by the sensing system in real time.

[0091] Embodiment two

[0092] This embodiment provides a surface acoustic wave passive wireless sensing system based on artificial intelligence, as shown in Figure 3 , which includes a SAW sensing unit, a reader, a radio frequency cable and a host computer.

[0093] The SAW sensing unit is used to receive the query signal and transmit the acoustic surface wave signal with environmental information.

[0094] The reader is used to receive the signal received by the SAW sensing unit and convert it into processable data.

[0095] The radio frequency cable is used to transmit signals between the hardware of the reader and the antenna of the reader; the antenna of the reader is used to enhance the efficiency of signal transmission and reception.

[0096] The host computer includes a data acquisition module, a data processing module, a data display module, a model updating module and an intelligent power detection module.

[0097] The data acquisition module is used to cooperate with the surface acoustic wave sensing unit to realize the acquisition of the acoustic surface wave signal with environmental information.

[0098] The data processing module is used to preprocess the radio frequency signal intensity data of various acoustic surface wave echo signals so that all data lengths are consistent; then standardize the data and perform one-hot encoding on the standardized data; finally, add an additional channel dimension to the two-dimensional data, so that the data becomes three-dimensional. Specifically, the dimensions of these three-dimensional data are sample number, feature number and channel number, where the channel number is 1.

[0099] The intelligent power detection module deploys a one-dimensional CNN neural network classification model of the trained attention mechanism, analyzes the signal received by the SAW sensing unit, and combines the time domain features of different types of noise, normal response and interference signals learned to perform identification.

[0100] The data display module visualizes the model performance by using the confusion matrix, ROC curve and t-SNE method. The confusion matrix is a matrix that specifically shows the performance of the classification model, and its heat map is as shown in Figure 7 , which shows the comparison between the prediction results of the model for the test data and the actual labels. Each row of the confusion matrix represents the actual class, and each column represents the predicted class of the model, so that the values on the diagonal line represent the number of correct classifications, and the values on the non-diagonal line represent the number of incorrect classifications. From Figure 7 , it can be seen that the color on the diagonal line is darker and the color on the non-diagonal line is lighter and concentrated between the 0 and 1 categories, and it can be seen that the misjudgment rate of the model of the embodiment is not high and is mainly concentrated between the noise and the true response. The ROC curve shows the recognition ability of the model by plotting the true positive rate (TPR) and the false positive rate (FPR) at different thresholds. The area under the ROC curve (AUC) can quantitatively reflect the performance of the model, and the higher the AUC value, the better the performance of the model. From Figure 8 , it can be seen that the ROC curve of the model of the present application shows that the AUC of the three categories is 0.98 or above, which proves that the performance of the model of the present application is good. The t-SNE is a high-efficiency nonlinear dimensionality reduction technique, which is mainly used for visualization of high-dimensional data. It maps high-dimensional data points to two-dimensional or three-dimensional space to observe the distribution of data points on the plane (or space) graph. From Figure 9 , it can be seen that the t-SNE visualization of the three different layers in the model at the 99th iteration of the model training shows that there is a gradually obvious gap between the clusters of different colors representing different types as the layer deepens, which indicates that the model of the present application can well distinguish different categories.

[0101] The model updating module is configured to periodically input the processed acoustic surface wave echo signals in a new environment into the trained one-dimensional CNN neural network classification model, retrain the one-dimensional CNN neural network classification model, update the parameters of the trained one-dimensional CNN neural network classification model, and deploy the updated model to the host computer to adapt to the changes in the environment and new interference in long-term operation.

[0102] In this embodiment, the model is actually deployed to the system host computer part, and tests of noise, response and interference are respectively performed. First, the algorithm model is used for signal classification of a single frequency point, and the classification result is as shown in Figure 10 From the (a) noise test classification result in Figure 10 It can be seen that the interference signal mixed in the noise can also be distinguished when the single frequency point noise time domain test is performed. From the (b) response test classification result in Figure 10 It can be seen that the noise signal mixed in the response can also be distinguished when the single frequency point response time domain test is performed. From the (c) response test classification result in Figure 11 It can be seen that the interference signal is well identified. Therefore, the algorithm well distinguishes the signals of noise, response and interference.

[0103] After distinguishing different kinds of signals, then the analysis of the entire actual test frequency domain is performed. The frequency scanning is performed on the required frequency domain, the frequency step is 10 kHz, and the single frequency point is sampled 16 times, that is, 16 times of algorithm classification are performed on each frequency point, and the distinction result of the single frequency point is the mode of the 16 times of algorithm classification results. The frequency domain scanning is performed in the pure noise environment, the interference environment and the SAW response environment with the applied interference, and the measured results are as shown in Figure 11 It can be seen from the (a) in Figure 11 It can be seen that each frequency sampling point in the pure noise environment is correctly identified as a noise signal. Then, the 415.75 MHz interference signal is applied in the pure noise environment, and the result is as shown in the (b) in Figure 11 It can be seen that the interference signal 415.75 MHz is correctly identified as an interference signal, and the frequencies 415.6 MHz before and 415.9 MHz after the interference signal are correctly identified as noise signals. Finally, in the environment with the applied 415.75 MHz radio frequency interference signal, the SAW sensing unit with the measured resonance frequency of 415.6 MHz is put in as a device to be tested to perform the SAW response test in the interference environment. The result is as shown in Figure 11As shown in (c) in the figure, near the frequency 415.75MHz of the interference signal, the frequency sampling points are all correctly identified as interference signals. The signals of the frequency points which are not interfered on, before and after the resonance frequency 415.6MHz of the SAW sensor, are still correctly identified as SAW response signals. Moreover, the fitted frequency shown by the dotted line is consistent with the measured SAW sensor resonance frequency, which is 415.6MHz. In summary, through actual testing, the algorithm can not only measure effective real signals in an interfered environment, but also preserve the original noise or interference information, and has good implementation effect in actual application.

[0104] Those skilled in the art can understand that the above description is only preferred examples of the application and is not used to limit the application. Although the application has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples or make equivalent replacements for part of the technical features. Any modification, equivalent replacement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An anti-interference method for an artificial intelligence-based surface acoustic wave passive wireless sensing system, characterized in that, The method comprises the following steps: Step 1: obtaining radio frequency signal intensity data of various surface acoustic wave echo signals including noise, normal response and interference at different frequencies, and giving different labels to the surface acoustic wave echo signals; Step 2: first, pre-processing the radio frequency signal intensity data of various surface acoustic wave echo signals to make all data lengths consistent; then, standardizing the data and performing one-hot encoding on the standardized data; finally, adding an extra channel dimension to the two-dimensional data to make the data three-dimensional, i.e. the dimensions of the three-dimensional data are sample number, feature number and channel number, wherein the channel number is 1; Step 3: constructing a one-dimensional CNN neural network classification model with attention mechanism, which comprises one one-dimensional convolution layer and an attention mechanism module; the one-dimensional convolution layer is used to extract time-frequency features of the radio frequency signal intensity data, the size of the convolution kernel is 1, the padding is 1, and the stride is 1, which is used to enhance the nonlinear ability of feature extraction; the attention mechanism module performs weighted average on the time-frequency features extracted by the one-dimensional convolution layer; The data obtained in step 2 is used to train the one-dimensional CNN neural network classification model; Step 4: collecting new surface acoustic wave echo signals, processing them in the same way as in step 2, and inputting them into the trained one-dimensional CNN neural network model, and the model outputs the category of the new surface acoustic wave echo signals.

2. The anti-interference method of the artificial intelligence-based surface acoustic wave passive wireless sensing system according to claim 1, characterized in that, The label of the surface acoustic wave echo signal at different frequencies including only noise is 1, the label of the surface acoustic wave echo signal containing noise and normal response is 0, and the label of the surface acoustic wave echo signal containing only noise and interference is 2.

3. The anti-interference method of the artificial intelligence-based surface acoustic wave passive wireless sensing system according to claim 1, characterized in that, In step 2, the radio frequency signal intensity data of various surface acoustic wave echo signals is pre-processed to make all data lengths consistent, which specifically comprises: For data missing values, the mode is preferred to fill, and the median is used to fill in the case where the mode does not exist; For data whose length exceeds the preset threshold, the data is truncated.

4. The anti-interference method of the artificial intelligence-based surface acoustic wave passive wireless sensing system according to claim 1, wherein In addition to the one-dimensional convolution layer and the attention mechanism module, the one-dimensional CNN neural network classification model further comprises a second one-dimensional convolution layer, a maximum pooling layer, a first dropout layer, a third one-dimensional convolution layer, a second dropout layer, a flattening layer, a first full connection layer, a third dropout layer and a second full connection layer; The maximum pooling layer is used to reduce the feature dimension while retaining the most important information; The first dropout layer, the second dropout layer and the third dropout layer are used to randomly turn off a part of neurons during the training process to prevent model overfitting; The flattening layer is used to convert multi-dimensional time-frequency features into one-dimensional feature vectors; The first full connection layer and the second full connection layer are used for high-level representation and classification of the surface acoustic wave echo signals.

5. An artificial intelligence-based surface acoustic wave passive wireless sensing system, characterized by, The system comprises a surface acoustic wave sensing unit, a reader, a radio frequency cable and an upper computer; The surface acoustic wave sensing unit is used to receive inquiry signals and transmit surface acoustic wave signals with environmental information; The reader is used to receive the signals received by the surface acoustic wave sensing unit and convert them into processable data; The radio frequency cable is used to realize signal transmission between the hardware of the reader and the reader antenna; The host computer deploys the trained one-dimensional CNN neural network classification model of the attention mechanism, and is used for realizing the anti-interference method of the surface acoustic wave passive wireless sensing system based on artificial intelligence in any one of claims 1-4.

6. The artificial intelligence based surface acoustic wave passive wireless sensing system of claim 5, wherein, The host computer comprises a data acquisition module, a data processing module, a data display module, a model updating module and an intelligent power detection module. The data acquisition module is used for cooperating with the surface acoustic wave sensing unit to realize the acquisition of the surface acoustic wave signals with environmental information. The data processing module is used for pre-processing the radio frequency signal intensity data of various surface acoustic wave echo signals to make all data lengths consistent, then performing standardization processing on the data, and performing one-hot encoding on the standardized data, finally adding an additional channel dimension to the two-dimensional data to make the data into three-dimensional form, that is, the dimensions of the three-dimensional data are respectively the number of samples, the number of features and the number of channels, wherein the number of channels is 1. The intelligent power detection module deploys the trained one-dimensional CNN neural network classification model of the attention mechanism, analyzes the signals received by the SAW sensing unit, combines the time domain features of different types of noise, normal response and interference signals learned to identify; The data display module uses the confusion matrix, ROC curve and t-SNE method to perform model performance visualization; The model updating module is used for inputting the processed surface acoustic wave echo signals under the new environment into the trained one-dimensional CNN neural network classification model, retraining the one-dimensional CNN neural network classification model, and updating the parameters of the trained one-dimensional CNN neural network classification model to adapt to the environmental changes and new interference in long-term operation.

Images