Liquid identification method and system based on triboelectrification-convolutional neural hybrid network

Abstract

The invention relates to a liquid identification method and system based on a triboelectrification-convolutional neural hybrid network, and belongs to the technical field of liquid identification, and the method comprises the following steps: continuously mixing a pulse gas flow with a to-be-detected liquid to form a gas-liquid two-phase flow, inputting the gas-liquid two-phase flow into a triboelectrification nanogenerator, and collecting a voltage signal at the same time; according to the method, liquid to be detected is fully sheared through pulse airflow, gas-liquid two-phase flow enters the friction nano-generator to be in high-frequency contact-separation with the friction nano-generator, and the friction nano-generator is subjected to high-frequency contact-separation with the gas-liquid two-phase flow, so that the high-frequency contact-separation between the gas-liquid two-phase flow and the friction nano-generator is realized. The friction nanometer generator generates stable and repeatable voltage signals with high spatial and temporal distribution characteristics, the voltage signals are collected and preprocessed to obtain time sequence data, then the time sequence data are input into a convolutional neural network and long and short term memory network mixed model, and therefore the category or concentration grade of liquid to be detected is accurately obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of liquid identification, and in particular relates to a liquid identification method and system based on a triboelectric-convolutional neural hybrid network. Background Art

[0002] Identifying liquid composition and concentration has broad and important application prospects in fields such as environmental monitoring, medical diagnosis, food safety, and industrial control. Common detection methods include spectroscopy, chromatography, and electrochemical sensors. However, these methods generally suffer from large equipment size, long detection cycles, strong dependence on the environment, and high costs, making them difficult to meet the needs of real-time on-site identification. Furthermore, with the growing demand for new technologies such as "non-invasive detection" and "rapid response," traditional liquid detection solutions are no longer able to meet the requirements of modern intelligent applications.

[0003] In recent years, the triboelectric nanogenerator (TENG), an emerging energy harvesting and sensing technology based on the principles of triboelectric charging and electrostatic induction, has garnered widespread attention. Without the need for an external power source, TENGs can convert mechanical energy into electrical signals for energy harvesting or sensing applications. They offer significant advantages in terms of simple structure, low cost, and high integration. In the field of liquid identification, TENGs, through the contact-separation behavior between liquids and solid materials, can generate voltage signals that reflect the liquid's composition and concentration.

[0004] Currently, TENG-based liquid identification generally uses methods such as liquid column injection, droplet impact, or static contact. Liquid column injection involves injecting a section of the liquid to be tested into the TENG, droplet impact involves a droplet of the liquid to be tested hitting the TENG through free fall, and static contact involves inserting the TENG into the liquid to be tested and then removing it. During these processes, a voltage signal is generated on the TENG, which is then collected and analyzed using a deep learning model to accurately determine the composition and concentration of the liquid.

[0005] However, the above-mentioned method of using TENG for liquid identification has a long cycle time for the liquid droplet or liquid column to impact, withdraw, or statically contact with the TENG, resulting in a slow speed for obtaining periodic voltage signals, which leads to the following defects in liquid identification: (10) Low signal dimension: Since the output signal generated by TENG is mainly affected by the conductivity of the liquid, the slow contact-separation speed between the liquid and TENG increases the difficulty of sample acquisition. Usually, only a single or a few slowly changing pulse voltage signals can be obtained. This low-speed pulse voltage signal cannot excite or capture the contact interface between the liquid and TENG, and the fine time domain features such as rich high-frequency response, complex relaxation process or multimodal oscillation that may be generated during the fast and dynamic contact and separation process. It lacks multidimensional information that can distinguish complex liquids. (11) Large environmental interference: When liquid hits TENG or is in static contact with it, it is easily affected by gravity, air disturbance, etc., resulting in poor reproducibility of contact and separation between liquid and TENG; (12) Difficult to support deep learning training: The voltage signal generated by the low-speed contact between the liquid and the TENG has a simple time domain structure and lacks features. The powerful ability of the deep learning model in processing time series data depends on the input signal having a sufficiently long time span and rich dynamic change characteristics. The short and simple pulse signal generated lacks a time evolution process with sufficient complexity and information content, making it difficult for the model to learn a high-dimensional spatiotemporal feature representation with high discrimination and fine distinction between multiple liquids.

[0006] In summary, the existing TENG-based liquid recognition method will reduce the accuracy of liquid composition and concentration recognition due to the above defects. Summary of the Invention

[0007] In view of this, the present invention provides a liquid identification method and system based on a triboelectric-convolutional neural hybrid network to address the shortcomings of the prior art. The present invention enhances the contact-separation rate and contact area of ​​the liquid-solid interface, so that the triboelectric nanogenerator generates a stable, repeatable voltage signal with a high spatiotemporal distribution characteristic, thereby accurately obtaining the type or concentration level of the liquid to be tested.

[0008] The technical solution of the present invention is: a liquid identification method based on a triboelectric-convolutional neural hybrid network, comprising the following steps: Continuously mixing the pulsed gas flow with the liquid to be tested to form a gas-liquid two-phase flow; The gas-liquid two-phase flow is input into the triboelectric nanogenerator, and the voltage signal is collected simultaneously; Preprocessing the collected voltage signal to obtain time series data; The time series data is input into a hybrid model of convolutional neural network and long short-term memory network to extract the local features of the time series data and capture the dynamic changes of local features in the time series, thereby obtaining the category or concentration level of the liquid to be tested.

[0009] Preferably, preprocessing the collected voltage signal to obtain time series data includes: Convert the collected voltage signal into digital time series data, and then perform low-pass filtering on the digital time series data; Perform peak detection on the digitized time series data after low-pass filtering, and then slice and extract adjacent time windows centered on the peak value as an independent sample signal; The extracted signal samples are normalized to obtain time series data.

[0010] Preferably, the normalization process to obtain the time series data includes: performing a standard transformation on the amplitude of each signal sample so that the data of the signal sample has statistical characteristics of zero mean and unit variance.

[0011] Preferably, the convolutional neural network and long short-term memory network hybrid model includes: a CNN layer and an LSTM layer connected in sequence; The CNN layer consists of two layers of one-dimensional convolutional networks connected in series to extract local features of the input time series data and output feature maps; The LSTM layer consists of two stacked LSTM networks, each containing 128 hidden units. It captures the dynamic changes of local features in the time series based on the input feature map and outputs the hidden state at the final moment. The output hidden state at the final moment is mapped to the N-dimensional category space through the fully connected layer, where N is the type of liquid to be identified or the number of concentration levels.

[0012] Preferably, each layer of the one-dimensional convolutional network is sequentially connected with batch normalization and ReLU activation, and a maximum pooling is configured at the end to downsample the feature map. Each layer of the LSTM network is respectively configured with a Dropout layer, and the dropout rate of the Dropout layer is 0.5.

[0013] Preferably, after the hidden state at the final moment is mapped to the N-dimensional category space through the fully connected layer, the probability of each category is generated through the Softmax layer.

[0014] A liquid recognition system based on a triboelectric-convolutional neural network hybrid network, comprising: The gas output module includes: an air compressor, a solenoid valve and a signal generator. The solenoid valve is connected to the output port of the air compressor through a pipeline, and the signal generator is connected to the solenoid valve signal to output a pulsed airflow; The liquid output module includes: a container for the liquid to be tested and a pump body, wherein the input port of the pump body is connected to the container for the liquid to be tested through a pipeline to transport the liquid to be tested; The triboelectric signal detection module includes a triboelectric nanogenerator, a ring electrode, and a gas-liquid mixing tube. One end of the gas-liquid mixing tube is connected to the output port of the solenoid valve and the pump body through a three-way connector and a pipeline, and the other end is connected to the triboelectric nanogenerator. The liquid to be tested mixes with the pulsed airflow to form a gas-liquid two-phase flow. The droplets in the gas-liquid two-phase flow collide and rub against the inner wall of the triboelectric nanogenerator. The ring electrode is set on the triboelectric nanogenerator to induce a voltage signal. The deep learning recognition module includes: an oscilloscope and a processor. The oscilloscope is electrically connected to the ring electrode through a voltage probe to collect voltage signals. The processor is connected to the oscilloscope signal. The processor preprocesses the collected voltage signals to obtain time series data, and then uses a hybrid model of a convolutional neural network and a long short-term memory network to extract local features of the time series data and capture the dynamic changes of local features in the time series to obtain the category or concentration level of the liquid to be tested.

[0015] Preferably, the gas-liquid mixing tube is a Laval nozzle-like structure.

[0016] Preferably, the inner diameter of the gas-liquid mixing tube at one end close to the friction nanogenerator is smaller than the inner diameter of the other end.

[0017] Compared with the prior art, the present invention provides a liquid identification method and system based on a triboelectric-convolutional neural hybrid network. The pulsed airflow is continuously mixed with the liquid to be tested, and the pulsed airflow fully shears the liquid to be tested to form a high-speed pulsed gas-liquid two-phase flow. The gas-liquid two-phase flow enters the friction nanogenerator and undergoes a high-frequency contact-separation process with it, thereby enhancing the contact-separation rate and contact area of ​​the liquid-solid interface, so that the friction nanogenerator generates a stable, repeatable voltage signal with a high temporal and spatial distribution characteristic. The voltage signal is collected and preprocessed to obtain time series data, and the time series data is input into a hybrid model of a convolutional neural network and a long short-term memory network to extract the local features of the time series data and capture the dynamic changes of the local features in the time series, thereby accurately obtaining the category or concentration level of the liquid to be tested, thereby improving the accuracy of liquid composition and concentration identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of a liquid identification system of the present invention; Figure 2 This invention Figure 1 A is an enlarged schematic diagram; Figure 3 This is a flow chart of the working principle of triboelectric charging of the present invention; Figure 4 It is a training flow chart of the hybrid model of convolutional neural network and long short-term memory network of the present invention; Figure 5 It is a confusion matrix diagram of the classification results predicted by the hybrid model of the present invention; Figure 6 It is a confusion matrix diagram of the concentration results predicted by the hybrid model of the present invention; Figure 7 This is a performance comparison diagram of the gas-liquid mixing tube of the present invention before and after optimization; Figure 8 This is the impedance matching diagram of the friction nanogenerator of the present invention. DETAILED DESCRIPTION

[0019] The present invention provides a liquid identification method and system based on triboelectric-convolutional neural hybrid network. Figures 1 to 8 The present invention is described with reference to a structural schematic diagram of FIG.

[0020] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "axial", "radial", "circumferential" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the technical solutions of the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0021] The present invention provides a liquid identification method based on a triboelectric-convolutional neural hybrid network, comprising the following steps: Continuously mixing the pulsed gas flow with the liquid to be tested to form a gas-liquid two-phase flow; The gas-liquid two-phase flow is input into the triboelectric nanogenerator, and the voltage signal is collected simultaneously; Preprocessing the collected voltage signal to obtain time series data; The time series data is input into a hybrid model of convolutional neural network and long short-term memory network to extract the local features of the time series data and capture the dynamic changes of local features in the time series, thereby obtaining the category or concentration level of the liquid to be tested.

[0022] In this embodiment, a liquid identification method based on a triboelectric-convolutional neural hybrid network is disclosed. A pulsed airflow is continuously mixed with a liquid to be tested, and the pulsed airflow fully shears the liquid to be tested to form a high-speed pulsed gas-liquid two-phase flow. The gas-liquid two-phase flow enters the friction nanogenerator and undergoes a high-frequency contact-separation process with it, thereby enhancing the contact-separation rate and contact area of ​​the liquid-solid interface, so that the friction nanogenerator generates a stable, repeatable voltage signal with a high temporal and spatial distribution characteristic. The voltage signal is collected and preprocessed to obtain time series data. The time series data is input into a hybrid model of a convolutional neural network and a long short-term memory network to extract local features of the time series data and capture the dynamic changes of the local features in the time series, thereby accurately obtaining the category or concentration level of the liquid to be tested, thereby improving the accuracy of liquid composition and concentration identification.

[0023] In the above embodiment, signal acquisition does not require external power supply. After the gas-liquid two-phase flow enters the triboelectric nanogenerator, a high-frequency contact-separation friction process occurs with it to generate a voltage signal. This part of the electrical energy, i.e., the voltage signal, is collected. Figure 8 , Figure 8 This is the impedance diagram for the triboelectric nanogenerator of this embodiment. The triboelectric nanogenerator achieves maximum output power at 500 megohms, reaching a peak output power of 0.56mW. This converts the kinetic energy of the gas-liquid two-phase flow into electrical energy. This electrical energy drives signal acquisition. Under typical experimental conditions, the average latency for completing a single recognition is less than 1 second.

[0024] As a further optimization solution, in this embodiment, preprocessing the collected voltage signal to obtain time series data includes: Convert the collected voltage signal into digital time series data, and then perform low-pass filtering on the digital time series data; Perform peak detection on the digitized time series data after low-pass filtering, and then slice and extract adjacent time windows centered on the peak value as an independent sample signal; The extracted signal samples are normalized to obtain time series data.

[0025] In this embodiment, the collected original voltage signal is first converted into digital time series data. In order to retain the intrinsic characteristics of the signal to the greatest extent and remove noise and unnecessary interference, the digital time series data is low-pass filtered. This step can effectively suppress high-frequency noise caused by equipment vibration, electromagnetic interference, etc., and retain the key features within the main signal frequency band. Then, a peak detection operation based on the threshold difference method is performed to identify the location of the main pulse of the voltage signal in each cycle. After detecting a stable peak, the adjacent time window is sliced ​​and extracted as an independent sample signal with the peak as the center. This step can ensure that the time length of all input data is consistent, the phase is aligned, and the structure is unified.

[0026] In addition, this embodiment can also extract additional interpretable time domain parameters such as "rise time", "fall time", "signal width", "maximum / minimum amplitude" and so on during the preprocessing stage of the collected voltage signal without interfering with the original waveform structure, for auxiliary use in model training or feature visualization, for example: introducing an automated feature engineering auxiliary module to extract the fir tree time domain parameters.

[0027] As a further optimization solution, in this embodiment, normalizing the extracted signal samples to obtain time series data includes: performing a standard transformation on the amplitude of each signal sample so that the data of the signal sample has statistical characteristics of zero mean and unit variance.

[0028] In this embodiment, all extracted signal samples will be uniformly processed in the normalization module. The normalization process uses the Z-score standardization method to perform a standard transformation on the amplitude of each signal sample, so that the data of the signal sample has the statistical characteristics of zero mean and unit variance, which helps to eliminate systematic offsets caused by slight differences in liquid concentration or environmental disturbances.

[0029] In the above embodiment, a standardized, structured, and automated preprocessing process is used to provide high-quality and highly consistent input data for the subsequent deep learning model, ensuring that the hybrid model of the convolutional neural network and the long short-term memory network has strong generalization capabilities and fast response characteristics.

[0030] As a further optimization solution, the hybrid model of convolutional neural network and long short-term memory network in this embodiment includes: a CNN layer and an LSTM layer connected in sequence; The CNN layer consists of two layers of one-dimensional convolutional networks connected in series to extract local features of the input time series data and output feature maps; The LSTM layer consists of two stacked LSTM networks, each containing 128 hidden units. It captures the dynamic changes of local features in the time series based on the input feature map and outputs the hidden state at the final moment. The output hidden state at the final moment is mapped to the N-dimensional category space through the fully connected layer, where N is the type of liquid to be identified or the number of concentration levels.

[0031] Reference Figure 4 , Figure 4 This paper presents a training flowchart of a hybrid model of convolutional neural networks and long short-term memory networks. In order to fully exploit the complex features contained in the voltage signal of gas-liquid two-phase flow and achieve high-precision recognition of various liquid types and concentrations, a hybrid deep learning model is developed by integrating the convolutional neural network (CNN) and the long short-term memory network (LSTM) architecture. This hybrid model not only has the ability to extract local time domain structures, but also can effectively model the global dependencies of signals in the time dimension. It is particularly suitable for processing triboelectric time series data with coexisting periodicity, weak nonlinearity and multi-scale features.

[0032] In terms of the structure of the hybrid model of convolutional neural network and long short-term memory network in this embodiment, the CNN layer serves as a front feature extractor. The CNN layer is composed of two layers of one-dimensional convolutional networks connected in series. Through one-dimensional convolution operations, it slides on the time axis to identify key local patterns in the signal, such as the edge steepness of the waveform, the change in pulse amplitude and the rhythm of signal fluctuations.

[0033] The feature map is then fed into the LSTM layer, which consists of two stacked LSTM layers, each containing 128 hidden units. This layer effectively captures long-range dependencies in time series and models the complex behavior of liquids over multiple signal cycles due to changes in viscosity, polarity, and conductivity. The LSTM layer is configured to output only the hidden state at the final moment, representing the temporal features of the entire signal.

[0034] Finally, a fully connected layer is used to map the final hidden state output by the LSTM layer to an N-dimensional category space, where N is the number of liquid types or concentration levels to be identified.

[0035] The hybrid convolutional neural network and long short-term memory network model in the above embodiment uses a cross-entropy loss function as the optimization objective during training, and uses the Adam optimizer for parameter updates. A learning rate scheduler is introduced during training to dynamically reduce the learning rate when validation set accuracy stagnates, thus avoiding falling into a local optimum.

[0036] To adapt to data distribution differences caused by environmental changes or task shifts, a differentiated parameter freezing strategy can be added. Specifically, during the model fine-tuning phase, some parameters of the convolutional network are frozen to retain its general feature extraction capabilities, while only the LSTM and fully connected layers are retrained, achieving rapid adaptation and accurate recognition with a small amount of new data.

[0037] Through the structural design and training mechanism of the hybrid model of the above-mentioned convolutional neural network and long short-term memory network, the deep learning recognition model of the present invention not only fully fits the multi-scale and temporal characteristics of the friction-generated voltage signal, but also demonstrates excellent accuracy and robustness in actual liquid recognition tasks.

[0038] As a further optimization solution, in this embodiment, each layer of the one-dimensional convolutional network is sequentially connected with batch normalization and ReLU activation, and a maximum pooling is configured at the end to downsample the feature map. A Dropout layer is configured after each layer of the LSTM network, and the dropout rate of the Dropout layer is 0.5.

[0039] In this example, batch normalization and ReLU activation are sequentially applied to each layer of the one-dimensional convolutional network to enhance the model's nonlinear expression capabilities and accelerate network convergence. Max pooling is configured at the end of the convolutional layer to downsample the feature maps, compressing the data dimension while preserving the core feature information, thereby highlighting the most discriminative signal segments.

[0040] In this embodiment, a Dropout layer is introduced after the output of each LSTM network layer, with a dropout rate of 0.5, to randomly block some neuron connections, which can suppress the risk of overfitting and further improve the generalization ability.

[0041] As a further optimization solution, in this embodiment, the hidden state at the final moment is mapped to the N-dimensional category space through the fully connected layer, and then the probability of each category is generated through the Softmax layer.

[0042] In this embodiment, the prediction results of the final hidden state at the final moment are mapped to an N-dimensional category space through a fully connected layer using a Softmax layer, and then converted into probability distributions for each category. The hybrid model architecture of the convolutional neural network and long short-term memory network in the above embodiment is adaptable to multi-classification tasks (such as liquid type identification) or regression tasks (such as concentration level prediction), with high task flexibility and deployment adaptability.

[0043] The core idea of ​​this invention is to combine the liquid-solid contact electrification mechanism in the triboelectric effect with the deep learning recognition model to construct a liquid recognition method with self-powered characteristics, high spatiotemporal resolution signal acquisition capabilities and high-precision recognition capabilities. High-speed pulsed airflow is used to shear a continuous liquid column to generate a gas-liquid two-phase flow, so that the liquid hits the inner wall of the solid material of the triboelectric nanogenerator in a high-frequency and irregular form, thereby greatly enhancing the contact-separation rate and contact area of ​​the liquid-solid interface, and generating a triboelectric signal with a high spatiotemporal distribution characteristic in the triboelectric nanogenerator.

[0044] At the same time, a deep learning recognition model uses this high-dimensional, multi-parameter voltage signal as input and constructs a CNN+LSTM hybrid neural network to accurately distinguish different liquid types and concentrations. This hybrid model, a convolutional neural network and a long short-term memory network, uses convolutional layers to extract local characteristic patterns of the signal and then uses LSTM units to capture its dynamic changes over time, accurately characterizing the differences in the physical and chemical properties of the liquids.

[0045] The liquid identification method of the present invention not only retains the low cost and high sensitivity advantages of triboelectric sensing, but also fully releases the potential of deep learning in complex data mining, achieving a deep integration of "hardware structure optimization" and "software intelligent empowerment".

[0046] Liquid type or concentration identification experiment Example 1: Detection of different types of liquids This example demonstrates the effectiveness and accuracy of a triboelectric-convolutional neural network-based liquid recognition method for identifying multiple liquid types. Six typical liquids, including deionized water, ethanol, sodium chloride solution, glucose solution, acetic acid solution, and ammonia solution, were selected as test objects. By setting the gas flow rate to 40 L / min and the liquid flow rate to 32 mL / min, the liquid-to-gas ratio was 0.08%, and the pulsed airflow frequency was 1 Hz.

[0047] Under these conditions, the liquid under test and the pulsed gas mix to form a high-speed gas-liquid two-phase flow. This flow then flows through the triboelectric nanogenerator, stimulating the liquid-solid contact and generating a voltage signal. This signal is collected and preprocessed, including filtering, before being fed into a hybrid model of a convolutional neural network and a long-short-term memory network for liquid classification inference.

[0048] Reference Figure 5 , Figure 5 The following is a confusion matrix diagram of the classification results predicted by the hybrid model in this implementation. The confusion matrix results for the recognition of six liquids based on this hybrid model show excellent classification performance across liquid categories, with the model achieving an overall recognition accuracy of 97.25%, with recognition accuracy approaching 100% for most liquid categories. Only minor confusion between a few categories exists, demonstrating that the hybrid convolutional neural network and long short-term memory network model in the liquid recognition method of this invention has strong expressive power in pattern learning and classification of complex voltage signals.

[0049] Compared to traditional methods such as spectral analysis and electrochemical sensing, the present invention's liquid identification method, based on a hybrid triboelectric-convolutional neural network, offers advantages such as requiring no external power supply, lightweight construction, fast response, and high recognition accuracy. The experimental results not only validate the present method's stability and reliability in liquid classification, but also demonstrate its broad applicability and technological advancement for low-cost, on-site, real-time identification tasks.

[0050] Example 2: Detection of liquids of different concentrations To further verify the sensitivity and discrimination ability of the liquid recognition method of the present invention to changes in liquid concentration, this example uses sodium chloride (NaCl) solutions of different concentrations as test objects to construct a concentration gradient recognition model. In the experiment, six NaCl solutions with concentration gradients were selected, with a concentration range of 1×10 -5 mol / L to 1 mol / L, and keeping other experimental parameters consistent with Example 1, that is, the gas flow rate is 40 L / min, the liquid-gas ratio is 0.08%, and the frequency of the pulse gas flow is 1 Hz.

[0051] The time series data is obtained through the same data collection and preprocessing, and is input into the hybrid model of convolutional neural network and long short-term memory network to classify and predict the concentration level. Figure 6 , Figure 6 The confusion matrix diagram of the concentration results predicted by the mixed model for this implementation is as follows: Figure 6 The confusion matrix at different concentrations is shown. The results demonstrate that the hybrid model can accurately distinguish six different concentration levels, with an overall recognition accuracy of 99.13%. The model exhibits excellent resolution, particularly at low concentrations, demonstrating its high sensitivity to subtle changes in the voltage signal.

[0052] Verified by multiple experiments, the liquid recognition method based on the triboelectric charging-convolutional neural hybrid network of the present invention has demonstrated ultra-high sensitivity, precise detection capability and good adaptability in the detection of liquids of different concentrations and types. These experimental results fully demonstrate the broad application prospects and feasibility of the recognition of the present invention.

[0053] These experimental results further demonstrate that the present liquid identification method is not only capable of classifying liquids but also offers high-precision quantitative identification capabilities, making it suitable for applications requiring discernment of minute concentration differences, such as salt concentration monitoring and biological fluid concentration screening. Compared to traditional detection methods, this method offers faster response times, lower costs, and improved system integration, demonstrating significant practical value and potential for widespread adoption.

[0054] Reference Figure 1 , Figure 1 This is a schematic diagram of the liquid identification system implemented in this invention, a liquid identification system based on a triboelectric-convolutional neural hybrid network, comprising: a gas output module comprising: an air compressor 1, a solenoid valve 2 and a signal generator 3, the solenoid valve 2 is connected to the output port of the air compressor 1 through a pipeline, the signal generator 3 is connected to the solenoid valve 2 signal to output a pulsed airflow, a liquid output module comprising: a container for the liquid to be tested 4 and a pump body 5, the input port of the pump body 5 is connected to the container for the liquid to be tested 4 through a pipeline to transport the liquid to be tested, a triboelectric signal detection module comprising: a triboelectric nanogenerator 12, a ring electrode 9 and a gas-liquid mixing tube 8, one end of the gas-liquid mixing tube 8 is connected to the output ports of the solenoid valve 2 and the pump body 5 respectively through a three-way connector 7 and a pipeline, and the other end is connected to the The friction nanogenerator 12 is connected, the liquid to be tested is mixed with the pulsed airflow to form a gas-liquid two-phase flow, and the droplets in the gas-liquid two-phase flow collide and rub against the inner wall of the friction nanogenerator 12. The annular electrode 9 is set on the friction nanogenerator 12 to induce a voltage signal. The deep learning recognition module includes: an oscilloscope 10 and a processor 11. The oscilloscope 10 is electrically connected to the annular electrode 9 through a voltage probe to collect voltage signals. The processor 11 is signal-connected to the oscilloscope 10. The processor 11 preprocesses the collected voltage signals to obtain time series data, and then uses a hybrid model of a convolutional neural network and a long short-term memory network to extract local features of the time series data and capture the dynamic changes of local features in the time series to obtain the category or concentration level of the liquid to be tested.

[0055] In this embodiment, the liquid identification system based on the triboelectric-convolutional neural hybrid network uses the air compressor 1 of the gas output module to output gas, and then uses the signal generator 3 to control the solenoid valve 2 to control the air flow at a preset frequency and duty cycle. For example, the pulse frequency range is 1Hz~5Hz. When the solenoid valve 2 is opened for 200ms and closed for 800ms per cycle, the duty cycle is 20%, and a stable 1Hz pulse air flow injection frequency is achieved, thereby providing high-speed, high-pressure pulse air flow, and the frequency can be flexibly adjusted according to the liquid characteristics and experimental requirements; the pump body 5 of the liquid output module (a programmable peristaltic pump) is used. ) The liquid in the liquid container 4 to be tested is uniformly transported at a precise flow rate, so that the liquid and the pulse airflow begin to mix at the three-way joint 7, and the pulse airflow shears the liquid into a large number of irregular tiny droplets to form a gas-liquid two-phase flow. The gas-liquid two-phase flow enters the friction nanogenerator 12 after passing through the three-way joint 7. During the contact and friction between the droplets in the gas-liquid two-phase flow and the inner wall of the friction nanogenerator 12, the friction nanogenerator 12 easily obtains negative charge, while the droplets lose electrons and become positively charged. When the droplets pass through the annular electrode, the annular electrode 9 on the friction nanogenerator 12 is used to sense the voltage signal generated by friction.

[0056] In the above embodiment, the triboelectric nanogenerator 12 utilizes a fluorinated polymer pipe. A ring electrode 9 is coaxially mounted and fixed in the middle of the pipe, serving as a sensing electrode for the friction-generated voltage signal. The width of the ring electrode 9 is approximately 1 cm, which can be flexibly adjusted based on the liquid's characteristics and needs. The copper material of the ring electrode 9 allows for efficient voltage signal sensing. The fluorinated polymer pipe exhibits excellent electrical insulation and contact electrification properties.

[0057] At the same time, the fluorine-containing polymer pipe can also be replaced by a pipe made of polytetrafluoroethylene and other materials. Polytetrafluoroethylene and other materials have negative electrification properties. In addition, fluoroelastomer rubber, polyformaldehyde, cellulose, polyvinyl chloride, polytetrafluoroethylene, polyurethane rubber, acrylonitrile-butadiene-styrene copolymer (ABS plastic), polycarbonate, polystyrene, polyetherimide, polydimethylsiloxane, silicone rubber, polyimide, polyvinylidene fluoride, polyetheretherketone, polyethylene, polystyrene, nylon, acrylic, silicone, polyphenylene sulfide, polypropylene and other materials can be used to make pipes to achieve contact and separation friction between the gas-liquid two-phase flow and the inner wall of the pipe. The inner wall of the pipe is easy to obtain negative charge, while the droplets lose electrons and become positively charged. When the droplets pass through the annular electrode, the annular electrode 9 mounted on the outside of the pipe senses the voltage signal generated by friction.

[0058] In the above embodiment, a one-way valve is provided on the pipe between the pump body 5 and the three-way joint 7 to prevent backflow, and the one-way valve is as close to the three-way joint 7 as possible to prevent gas from flowing back into the pipe connected to the output port of the pump body 5 and affecting the stability of liquid transportation.

[0059] In the above embodiment, the gas output module and the liquid output module work together to finely control the shear intensity and contact frequency of the gas-liquid two-phase flow with an adjustable gas-liquid ratio and pulse frequency.

[0060] It should be noted that when the liquid identification system is activated, pump 5 pushes the liquid at a constant speed to tee 7. The high-speed pulsed airflow controlled by solenoid valve 2 begins to mix with the liquid to be tested at tee 7, forming a high-speed, intermittent gas-liquid two-phase flow. Upon entering the fluorinated polymer pipe, the irregular microdroplets in this gas-liquid two-phase flow continuously impact the inner wall of the pipe, generating a series of high-frequency contact and separation behaviors.

[0061] According to the electron cloud-potential well model, when water molecules approach a fluorinated polymer pipe, the electron cloud shifts and transfers to the pipe surface, giving the pipe a negative charge and the liquid a positive charge. When the liquid enters the region around the annular electrode 9 on the fluorinated polymer pipe, charge accumulates and the electric field perturbs at the liquid-solid interface, generating a periodic current between the external annular electrode 9 and ground, producing a typical pulse voltage signal on the oscilloscope.

[0062] Reference Figure 3 , each pulse process can be divided into four stages: State I: The pipeline has not yet entered the two-phase flow, the electrode is stationary and has no signal; State 2: The two-phase flow enters the electrode area, forming an electric double-layer shielding effect. Electrons flow from the ground to the electrode, generating a forward voltage. State 3: The two-phase flow completely shields the electrode, and the signal is stable; State 4: The two-phase flow leaves the electrode area, the shielding effect weakens, and the electrons flow back to the ground from the electrode, generating a reverse voltage.

[0063] This cycle repeats at a set gas pulse frequency, generating a highly repeatable and sensitive triboelectric signal sequence. Different liquids exhibit varying amplitudes, waveform characteristics, rise times, and fall times in this signal, depending on their physical and chemical properties, such as polarity, viscosity, surface tension, and ion concentration.

[0064] As a further optimization solution, in this embodiment, the gas-liquid mixing tube 8 is a Laval nozzle-like structure.

[0065] The gas-liquid mixing tube 8 in this embodiment is a Laval nozzle-like structure. The pulsed airflow, liquid, and the gas-liquid two-phase flow formed at the three-way joint 7 enter the gas-liquid mixing tube 8 and exert a pressurizing and accelerating effect on it, thereby enhancing the degree of gas-liquid mixing. This can further increase the efficiency of the pulsed airflow and liquid mixing and shearing, as well as the speed of the gas-liquid two-phase flow, thereby achieving more sufficient shearing of the liquid into tiny droplets, so that the pulsed gas and liquid are fully sheared and mixed according to a predetermined gas-liquid ratio, further improving the contact-separation rate and contact area of ​​the liquid-solid interface.

[0066] In addition, the gas-liquid mixing tube 8 in the above embodiment may also be a Venturi tube, which can also pressurize and accelerate the pulsed airflow, liquid and gas-liquid two-phase flow.

[0067] In the above embodiment, the fluorine-containing polymer pipe is connected to the nozzle outlet end of the gas-liquid mixing pipe 8 , and the inner diameter of the fluorine-containing polymer pipe matches that of the gas-liquid mixing pipe 8 .

[0068] Reference Figure 2 , Figure 2 This is an enlarged schematic diagram of position A of the liquid identification system of this embodiment. As a further optimization solution, the inner diameter of the gas-liquid mixing tube 8 at one end close to the triboelectric nanogenerator 12 in this embodiment is smaller than the inner diameter at the other end.

[0069] In this embodiment, a trumpet-shaped structure with an inner diameter at one end larger than that at the other end is used to replace the Laval nozzle structure. The inner diameter of the inlet end is 8 mm, and the inner diameter of the outlet end is 8 mm. Compared with the Laval nozzle structure, a section of the pipeline is reduced, which can further enhance the gas-liquid mixing efficiency and the injection speed. Compared with the use of a Laval nozzle structure, it is more effective in improving the contact-separation rate and contact area of ​​the liquid-solid interface. At the same time, it ensures that a stable and repeatable droplet-shaped gas-liquid two-phase flow is formed within the 200ms ventilation and 800ms air-off cycles controlled by the solenoid valve.

[0070] The existing Laval nozzle structure is necking-straightening-expansion, denoted as W. In this embodiment, the structure with one end having a larger inner diameter than the other end (optimized structure) only contains the necking-straightening section, denoted as W / O. The performance comparison of the gas-liquid two-phase flow passing through the triboelectric nanogenerator (TENG) is shown in Figure 2. Figure 7 shown.

[0071] The optimized W / O structure increased the output voltage amplitude by 137% compared to the W structure, and the charge transfer momentum caused by a single gas-liquid two-phase flow impact increased by 246%. The fundamental reason for this is that for the gas-liquid two-phase flow TENG, only the flow velocity gain brought by the necking section is needed to enhance charge separation, while the gas-liquid atomization and diffusion caused by the expansion section actually reduce the system's impact efficiency.

[0072] As a further optimization solution, this embodiment also includes: a voltage acquisition and processing module, which includes: an oscilloscope 10 and a processor 11. The oscilloscope 10 is electrically connected to the annular electrode 9 through a voltage probe, and the processor 11 is signal-connected to the oscilloscope 10.

[0073] In this embodiment, the voltage probe of the voltage acquisition and processing module oscilloscope 10 receives the voltage signal sensed by the ring electrode 9 and converts it into digital time series data. The digital time series data is transmitted to the processor 11 using the VISA protocol. The processor 11 uses a pre-processed and pre-trained convolutional neural network and long short-term memory network hybrid model to infer the category or concentration level of the liquid, outputs the inference result, and uses a display device for visual expression.

[0074] In addition to the convolutional neural network and long short-term memory network hybrid model in this application, this liquid recognition system also supports the use of other mainstream machine learning or deep learning models for liquid recognition tasks, including but not limited to: 1. Traditional machine learning models, such as support vector machines, random forests, and extreme gradient boosting trees, can achieve efficient classification in scenarios with small samples or sufficient feature engineering; 2. Other deep learning architectures, such as fully convolutional neural networks, residual networks, gated recurrent units, and temporal convolutional networks, can also be selected based on data characteristics and application requirements. Therefore, the recognition system of the present invention has a high degree of algorithm compatibility and flexible expansion capabilities, is not limited to a specific network structure, and can adjust and replace models according to specific application scenarios.

[0075] Specifically, the liquid identification system uses a high-impedance voltage probe on oscilloscope 10 to collect friction-generated voltage signals in real time. This high-speed, high-amplitude, periodic voltage waveform is then transmitted to processor 11 (the computer) via the oscilloscope. Because the high-speed shear and contact-separation behavior of gas-liquid two-phase flow on the inner wall of a fluorinated polymer pipe is highly repeatable, the collected signal exhibits a distinct periodic waveform structure, rich in physical meaning and liquid characteristic information.

[0076] The oscilloscope 10 in the above embodiment may be of a model including but not limited to Tektronix MSO44, Tektronix MSO44 coupled with a 100 MΩ high impedance voltage probe Tektronix P6015.

[0077] In the above embodiment, after the raw voltage signal is acquired, the voltage signal data undergoes a series of standardization and structured processing to adapt it to the input requirements of the subsequent convolutional neural network and long short-term memory network hybrid model. The entire data processing flow was highly integrated into the software architecture of processor 11 from the outset of system design, enabling fully automated and efficient pipeline processing from signal acquisition to model input.

[0078] In the above embodiment, the voltage probe of the oscilloscope 10 is a high-impedance voltage probe, and the processor 11 also has a human-computer interaction module. The human-computer interaction module uses PYQT5 to create a UI interface, which integrates data acquisition, model reasoning and visualization interface. Users can obtain recognition results within seconds. It is easy to operate and suitable for multi-scenario deployment.

[0079] The present invention proposes a liquid identification method and system based on a triboelectric-convolutional neural network hybrid network. This method innovatively integrates a triboelectric nanogenerator mechanism with a convolutional neural network and a long-short-term memory network hybrid model. By constructing a high-speed pulsed gas-liquid two-phase flow, a high-frequency contact-separation process is achieved within a fluorinated polymer pipeline, generating a stable and repeatable self-powered triboelectric signal. Through preprocessing steps such as automated signal filtering, peak slicing, and Z-score normalization, signal features are efficiently extracted and then input into a deep learning model for identification. This method effectively addresses key technical bottlenecks in existing liquid identification technologies, such as expensive detection equipment, slow response speed, and strong dependence on power supply. It not only significantly improves the real-time and accuracy of liquid identification (with a liquid type identification accuracy of 97.25% and a concentration prediction accuracy of 99.13%), but also achieves a low-cost, self-powered, and rapidly deployable system.

[0080] The above method provides a more efficient and reliable solution for liquid identification in various application scenarios such as environmental monitoring, biological diagnosis, food safety and industrial automation, and has important engineering practical value and promotion prospects.

[0081] Compared with the existing technology, the liquid identification method and system based on the triboelectric-convolutional neural hybrid network of the present invention has brought significant beneficial effects in multiple dimensions. First, at the signal acquisition level, the gas-liquid two-phase flow breaks the problems of low frequency and small contact area of ​​the traditional liquid column TENG device, greatly improving the amplitude and feature diversity of the triboelectric signal, and providing a richer data basis for subsequent recognition algorithms. Secondly, at the intelligent recognition level, the traditional manual feature extraction method is difficult to capture the nonlinear relationships and subtle differences hidden in the signal, and the CNN+LSTM hybrid model structure adopted by the present invention can automatically learn the key space-time patterns in the signal, thereby achieving an accuracy of 97.25% when identifying six representative liquids, and an accuracy of 99.13% when predicting salt solutions of different concentrations, which is significantly better than existing similar methods.

[0082] The invented identification system has a simple structure, universal devices, low manufacturing cost, good modularity and scalability, and can be easily promoted to multiple fields such as environmental protection, biological testing, food safety and emergency response, and has broad industrial application prospects.

[0083] In summary, the present invention has broken through the bottlenecks of existing methods in speed, accuracy, portability and intelligence through a systematic upgrade of the structural design and intelligent algorithm of triboelectric liquid identification technology, providing an innovative path and technical support for building a new generation of high-performance liquid identification systems.

[0084] The above disclosure is only a preferred specific embodiment of the present invention. However, the embodiments of the present invention are not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the scope of protection of the present invention.

Claims

1. A liquid identification method based on triboelectric-convolutional neural network, characterized in that: The following steps are involved: Continuously mixing the pulsed gas flow with the liquid to be tested to form a gas-liquid two-phase flow; The gas-liquid two-phase flow is input into the triboelectric nanogenerator, and the voltage signal is collected simultaneously; Preprocessing the collected voltage signal to obtain time series data; The time series data is input into a hybrid model of convolutional neural network and long short-term memory network to extract the local features of the time series data and capture the dynamic changes of local features in the time series, thereby obtaining the category or concentration level of the liquid to be tested.

2. The liquid identification method based on triboelectric charging-convolutional neural network according to claim 1 is characterized in that: The preprocessing of the collected voltage signal to obtain time series data includes: Convert the collected voltage signal into digital time series data, and then perform low-pass filtering on the digital time series data; Perform peak detection on the digitized time series data after low-pass filtering, and then slice and extract adjacent time windows centered on the peak value as an independent sample signal; The extracted signal samples are normalized to obtain time series data.

3. The liquid identification method based on triboelectric-convolutional neural network according to claim 2 is characterized in that: The normalization process to obtain time series data includes: performing a standard transformation on the amplitude of each signal sample so that the data of the signal sample has statistical characteristics of zero mean and unit variance.

4. The liquid identification method based on triboelectric-convolutional neural network according to claim 1, characterized in that: The convolutional neural network and long short-term memory network hybrid model includes: a CNN layer and an LSTM layer connected in sequence; The CNN layer includes two layers of one-dimensional convolutional networks connected in series to extract local features of the input time series data and output feature maps; The LSTM layer includes two stacked LSTM networks, each of which contains 128 hidden units. It captures the dynamic changes of local features in the time series based on the input feature map and outputs the hidden state at the final moment. The output hidden state at the final moment is mapped to the N-dimensional category space through a fully connected layer, where N is the type of liquid to be identified or the number of concentration levels.

5. The liquid identification method based on triboelectric charging-convolutional neural network according to claim 4 is characterized in that: Each layer of the one-dimensional convolutional network is sequentially connected to batch normalization and ReLU activation, and a maximum pooling is configured at the end to downsample the feature map. Each layer of the LSTM network is respectively configured with a Dropout layer, and the dropout rate of the Dropout layer is 0.

5.

6. The liquid identification method based on triboelectric-convolutional neural network according to claim 4 is characterized in that: The final hidden state is mapped to the N-dimensional category space through the fully connected layer, and the probability of each category is generated through the Softmax layer.

7. A liquid recognition system based on a triboelectric-convolutional neural network, characterized in that: include: A gas output module comprises: an air compressor (1), a solenoid valve (2) and a signal generator (3), wherein the solenoid valve (2) is connected to the output port of the air compressor (1) through a pipeline, and the signal generator (3) is connected to the solenoid valve (2) by signal to output a pulsed airflow; The liquid output module comprises: a container (4) for the liquid to be tested and a pump body (5), wherein the input port of the pump body (5) is connected to the container (4) for the liquid to be tested via a pipeline to transport the liquid to be tested; A triboelectric signal detection module comprises: a triboelectric nanogenerator (12), a ring electrode (9) and a gas-liquid mixing tube (8), one end of the gas-liquid mixing tube (8) is connected to the output ports of the electromagnetic valve (2) and the pump body (5) respectively through a three-way joint (7) and a pipeline, and the other end is connected to the triboelectric nanogenerator (12), the liquid to be tested is mixed with the pulsed airflow to form a gas-liquid two-phase flow, and the droplets in the gas-liquid two-phase flow collide and rub against the inner wall of the triboelectric nanogenerator (12), and the ring electrode (9) is arranged on the triboelectric nanogenerator (12) to induce a voltage signal; A deep learning recognition module comprises an oscilloscope (10) and a processor (11), wherein the oscilloscope (10) is electrically connected to the annular electrode (9) via a voltage probe to collect voltage signals, and the processor (11) is signal-connected to the oscilloscope (10). The processor (11) pre-processes the collected voltage signals to obtain time series data, and then uses a convolutional neural network and a long short-term memory network hybrid model to extract local features of the time series data and capture the dynamic changes of the local features in the time series to obtain the category or concentration level of the liquid to be tested.

8. The liquid identification system based on triboelectric-convolutional neural network according to claim 7, characterized in that: The gas-liquid mixing tube (8) is a Laval nozzle-like structure.

9. The liquid identification system based on triboelectric-convolutional neural network according to claim 7, characterized in that: The inner diameter of one end of the gas-liquid mixing tube (8) close to the friction nanogenerator (12) is smaller than the inner diameter of the other end.

Images