A method for ice classification and adhesion force evaluation of mems

Abstract

A method for MEMS ice type classification and adhesion force assessment belongs to the technical field of assessment methods, and particularly relates to a method for MEMS ice type classification and adhesion force assessment based on active thermal perturbation response and temporal convolutional network. This invention proposes a MEMS icing detection and adhesion force assessment method that integrates active thermal perturbation physical excitation and deep temporal convolutional network (TCN). This method artificially creates a non-destructive thermal transient process on the sensor surface, utilizing the kinetic differences of different ice types (clear ice, frost ice) and adhesion states during phase change heat transfer to achieve high-precision ice type classification and adhesion strength regression. The specific implementation steps are as follows: Step S1: Icing trigger monitoring based on sliding window statistics; the system is initially in a low-power monitoring mode, and the MEMS sensor acquires the resonant frequency f(t) data stream in real time.

Description

Technical Field

[0001] This invention belongs to the technical field of evaluation methods, and particularly relates to a method for classifying ice types and evaluating adhesion of MEMS (Micro-Electro-Mechanical Systems, in this invention referring to a miniature sensor integrating a resonator and a heating electrode) based on active thermal perturbation response and temporal convolutional networks. Background Technology

[0002] Terminology Explanation:

[0003] Active thermal perturbation refers to a short-duration, low-energy thermal pulse actively applied by the system to excite the thermal transient response of the measured substance, rather than to completely de-ice it.

[0004] TCN (Temporal Convolutional Network): A deep neural network structure that combines causal convolution and dilated convolution.

[0005] Glaze: A dense, transparent, and highly adhesive type of ice, usually formed by the slow freezing of supercooled water droplets after they hit a surface.

[0006] Rime: A loose, opaque ice form with weak adhesion, usually formed by the rapid freezing of supercooled water droplets mixed with air.

[0007] MEMS (Micro-Electro-Mechanical Systems) icing sensors have broad application prospects in aerospace anti-icing, wind turbine blade monitoring, and power transmission cable maintenance due to their advantages such as small size, light weight, low power consumption, and ease of integration and arraying. Traditional MEMS icing detection technology mainly relies on the mass loading effect or changes in dielectric properties of the material. The most basic detection principle is based on the Sauerbrey equation, using piezoelectric quartz crystal microbalances (QCM) or surface acoustic wave (SAW) devices as sensing elements. When supercooled water droplets freeze into ice on the sensor surface, the added mass of the ice layer causes a linear decrease in the resonant frequency of the vibration system. By monitoring the frequency shift, the mass of the ice can be inferred. With the development of technology, researchers have proposed a variety of novel MEMS sensor structures to improve detection sensitivity and adapt to complex environments.

[0008] In the field of microwave detection, Xie Jianbing et al. (patent number CN114001635B) disclosed a MEMS microwave sensor for icing detection. This technical solution mainly consists of a bottom metal microstrip transmission line, a dielectric substrate, and a top layer of complementary interdigitated open resonant rings. Its working principle utilizes the significant differences in the dielectric constants of ice, water, and air (ice has a dielectric constant of approximately 3.2, water approximately 80, and air approximately 1). When ice forms on the test area of ​​the sensor surface, the equivalent dielectric constant of the medium changes, thereby altering the distributed capacitance of the microwave circuit, causing the sensor's resonant frequency to drift or the return loss (S11 parameter) to change. This solution effectively enhances the local electric field intensity of the sensitive area by etching an open resonant ring with an interdigitated structure on the metal reference surface, thus improving the detection sensitivity to the covering medium to a certain extent.

[0009] In the field of mechanical strain detection, Chen Deyong et al. (patent number CN201110106333.9) disclosed a MEMS strain-type icing sensor and detection method. This scheme employs a structural design based on a peripherally fixed square flat diaphragm, integrating a detection resistor, a reference resistor, and an ice-melting resistor. Its core principle is based on the piezoresistive effect, utilizing four foil strain resistors forming a Wheatstone bridge at the edge of the flat diaphragm. When ice accumulates on the surface of the flat diaphragm, the contractile stress or gravity generated by the ice causes a slight deformation of the diaphragm, which in turn causes a change in the resistance value on the bridge arms. By providing constant current drive to the bridge through an interface circuit and amplifying and temperature-compensating the output voltage signal, this technology can convert ice thickness information into a voltage signal output, achieving continuous monitoring of ice thickness.

[0010] Xie Jianbing et al. (patent number CN114001635B) proposed a MEMS microwave sensor based on a complementary interdigitated open resonant ring. By optimizing the interdigitated structure to enhance the local electric field strength, they improved the detection sensitivity of icing media. However, this technology mainly relies on the change in microwave resonance characteristics caused by the dielectric constant of the ice layer, which is a steady-state detection of a single physical field. When facing complex meteorological conditions, it is difficult to accurately identify ice types with different microstructures, such as clear ice and frost ice, through simple dielectric differences. Moreover, as a purely electrical detection method, it cannot sense the mechanical adhesion strength between the ice layer and the substrate interface, resulting in a lack of key mechanical control basis for the de-icing system.

[0011] Chen Deyong et al. (patent number CN201110106333.9) proposed a MEMS strain-type icing sensor, which determines the ice thickness by detecting changes in surface mechanical stress caused by ice accumulation through a Wheatstone bridge integrated on a flat diaphragm. However, this technology relies on static mechanical load effects, making it difficult to effectively distinguish between aerodynamic pressure loads and ice mass loads under high-speed airflow, which can easily lead to false alarms. At the same time, single stress detection cannot analyze the thermodynamic properties of the ice layer, thus failing to achieve classification and identification of different ice types such as clear ice and frost ice, as well as adhesion force assessment.

[0012] The aforementioned existing technologies generally suffer from limitations such as passive detection modes and single information dimensions. Whether it is microwave detection based on dielectric properties or strain detection based on mechanical load, they are essentially threshold judgments of steady-state physical quantities, lacking in-depth analysis of the transient thermodynamic characteristics of the ice-sensor interface. This "blind testing" mechanism prevents existing sensors from accurately identifying ice types with different microstructures, such as clear ice and frost ice, in complex environments, and further hinders the quantitative assessment of ice adhesion strength, which is crucial for flight safety. Moreover, they are highly susceptible to false alarms caused by aerodynamic noise interference. To address these shortcomings, the present invention aims to provide a MEMS ice type classification and adhesion force assessment method based on active thermal disturbance response and temporal convolutional networks. By combining actively excited unsteady-state thermal conduction processes with deep learning algorithms, it achieves a leap from passive qualitative detection to active multidimensional quantitative assessment. Summary of the Invention

[0013] This invention proposes a method for MEMS icing detection and adhesion force assessment that integrates active thermal perturbation physical excitation with a deep temporal convolutional network (TCN). This method artificially creates a non-destructive thermal transient process on the sensor surface, utilizing the kinetic differences in ice types (clear ice, frost ice) and adhesion states during phase change heat transfer to achieve high-precision ice type classification and adhesion strength regression. The specific implementation steps are as follows:

[0014] Step S1: Icing trigger monitoring based on sliding window statistics;

[0015] The system initially operates in a low-power monitoring mode, with the MEMS sensor acquiring the resonant frequency f(t) data stream in real time. To avoid false triggering caused by transient environmental noise, this step does not employ a simple single-point threshold determination. Instead, a time-sliding window of length W is constructed, and the statistical characteristics of the frequency signal within this window are calculated. The mean frequency within the window at time t is defined as... and variance A suspected icing event is identified and a subsequent active thermal disturbance mode is triggered only when both the "signal mean value drops significantly" and the "signal is in a steady state" conditions are met. The triggering logic is shown in formula (1):

[0016]

[0017] in, This is the reference frequency under ice-free conditions. The minimum icing detection threshold is set. The noise tolerance threshold is given. Formula (1) ensures that the system only starts heating when there is a stable deposit, thus avoiding false starts under airflow disturbances or strong vibrations.

[0018] Step S2: Application of active thermal disturbance pulse and generation of heat flux;

[0019] Once the triggering mechanism is activated, the control unit applies a timed voltage pulse V(t) to the platinum thin-film heater integrated on the MEMS surface. This pulse waveform is designed as a square wave or trapezoidal wave, with a duration of [duration missing]. (Typically 50ms-200ms). According to Joule's law, the applied voltage pulse generates an instantaneous heat flux q(t) at the sensor-ice interface. Since the heater's resistance R(T) changes only slightly with temperature, the resulting real-time heat power... As shown in formula (2):

[0020]

[0021] in, The resistance at the reference temperature, Temperature coefficient of resistance It is an indicator function; The value is 1 during the pulse duration and 0 at other times. This thermal power, as an energy source, disrupts the thermal equilibrium at the ice-sensor interface, aiming to induce a microscopic phase change at the interface rather than completely de-icing.

[0022] Step S3: Transient response mechanism based on multilayer medium heat transfer model;

[0023] Temperature field distribution on the sensor surface under thermal pulse action It follows a one-dimensional unsteady-state heat conduction differential equation. Different ice types (clear ice, frost ice) have significantly different thermal conductivity. and density Its response to the thermal pulse, i.e., the temperature decay process, is completely different. This physical process can be described by formula (3):

[0024]

[0025] in Volumetric heat capacity. Resonant frequency offset of the MEMS sensor. Not only is it affected by mass load, but it is also greatly affected by temperature effects during thermal disturbances. This invention establishes a coupling equation between frequency response and temperature change as shown in equation (4):

[0026]

[0027] The first term is the Sauerbrey mass load term, which is considered constant during short-duration thermal pulses; the second term is the thermal effect term. Let be the frequency temperature coefficient of each layer of material. This is determined by observing the frequency temperature coefficient in formula (4). The curve of change over time is actually the thermal conductivity k and heat capacity characteristics of the ice layer in the inversion formula (3), which is the physical essence of the present invention in distinguishing ice types.

[0028] Step S4: Construction and preprocessing of multidimensional temporal feature tensors;

[0029] The system synchronously collects data before and after the thermal pulse triggering time window. Frequency drift sequence within Impedance real part sequence and temperature gradient sequence To eliminate individual sensor variations, the raw data was first standardized using Z-scores. Then, the standardized sequences across the three dimensions were channel-concatenated to construct the model input tensor. Let the number of sampling points be... The input tensor is represented by formula (5):

[0030]

[0031] This tensor It fully encompasses the change in mechanical stiffness under thermal disturbance (by... Characterization), interface phase transition loss (by Characterization) and thermal conductivity (by (Representation) information.

[0032] Step S5: Construct a dilated causal convolutional network (TCN) to extract thermodynamic fingerprints;

[0033] tensor The input sequence is fed into a temporal convolutional network. To capture long-range thermal relaxation features without increasing computational complexity, the core layer of the TCN designed in this invention employs dilated causal convolution. For the input sequence of the l-th layer... Its output at time t Defined as formula (6):

[0034]

[0035] in This represents the dilated convolution operator. For convolution kernel, The expansion factor is used in this invention. This design allows the receptive field of deep networks to expand exponentially with the number of layers. This allows the network to capture both the intense high-frequency fluctuations at the onset of a thermal pulse (corresponding to micro-melting at the ice interface, extracted by shallow convolution with low expansion rate) and the long refreezing recovery process after the pulse ends (corresponding to latent heat release, extracted by deep convolution with high expansion rate), thus forming a complete "thermodynamic fingerprint".

[0036] Step S6: Weighted keyframes based on self-attention mechanism;

[0037] Because the thermal transient response curve contains a large amount of redundant information (such as the long tail segment of the steady-state recovery period), a self-attention module is introduced after the TCN layer to enhance the features. This module calculates the input feature sequence. The internal correlations are automatically focused on key time steps where phase transitions occur (such as the latent heat plateau). Attention weight matrix The calculation is as shown in formula (7):

[0038]

[0039] Final weighted feature vector The product of the weight matrix and the value vector: Using formula (7), the algorithm can adaptively identify the most distinctive instantaneous features that differentiate clear ice from frost ice, thus improving the robustness of the classification.

[0040] Step S7: Multi-task loss function optimization and output decision;

[0041] The network's output layer is designed as a dual-head structure, performing ice type classification and adhesion regression tasks respectively. To simultaneously optimize these two related physical properties, a joint loss function is designed in this invention. As shown in formula (8):

[0042]

[0043] The first term is the cross-entropy loss for the classification task, used to optimize ice type determination (ice-free / frost-covered / clear ice / mixed); the second term is the mean squared error loss for the regression task, used to approximate the true ice shear adhesion strength. ; To balance the hyperparameters of the two task weights, the system ultimately outputs the ice category probability and the specific adhesion force value (unit: kPa), achieving a leap from qualitative detection to quantitative evaluation.

[0044] The beneficial effects of this invention.

[0045] This invention changes the traditional single mode of passively waiting for mass load in MEMS sensors. After the sensor detects the initial frequency drift, it actively applies non-destructive microsecond to millisecond thermal pulses. By utilizing the physical differences in microscopic thermal conductivity, specific heat capacity and interfacial phase transition rate of different ice types (clear ice and frost ice), it excites and collects multidimensional transient thermodynamic response curves containing frequency, impedance and temperature gradient, which serve as physical fingerprints for identifying ice types and evaluating adhesion.

[0046] This invention constructs a deep temporal convolutional network (TCN) architecture specifically for processing the aforementioned thermal transient responses. It employs an exponentially growing dilated causal convolutional structure, achieving an exponential expansion of the receptive field through multi-layer stacking. This allows it to simultaneously capture both the high-frequency fluctuation features of the thermal pulse (corresponding to interface micro-melting) and the long-term low-frequency relaxation features (corresponding to the refreezing process). Furthermore, a self-attention mechanism is embedded in the network, enabling it to adaptively weight key phase transition time windows (such as the latent heat release plateau) in the thermal response sequence, thereby accurately extracting phase transition features even in noisy environments.

[0047] This invention's multi-task learning output strategy and joint loss function design differ from existing technologies that can only output a single icing signal. By designing a dual-head output network structure and combining a joint objective function composed of weighted cross-entropy loss (for classification) and mean squared error loss (for regression), it achieves the simultaneous output of discrete ice type categories (ice-free / frost-free / clear ice / mixed) and continuous ice layer adhesion shear strength prediction values ​​during the same forward propagation process. This provides a complete decision-making basis for the de-icing system, combining qualitative classification and quantitative assessment.

[0048] Compared with the microwave detection based on dielectric constant changes by Xie Jianbing et al. and the detection technology based on static stress changes by Chen Deyong et al., this invention achieves a fundamental leap from single-dimensional passive qualitative alarm to multi-dimensional anisotropic quantitative assessment. Existing technologies essentially rely on steady-state threshold determination of physical quantities (such as mass or dielectric constant reaching a certain value), ignoring the rich information brought by the differences in the microstructure within the ice layer, resulting in the inability to distinguish between clear ice and frost ice and the inability to sense the interfacial adhesion strength. This invention, by introducing an active thermal perturbation strategy, utilizes the significant differences in thermodynamic properties such as thermal conductivity, specific heat capacity, and latent heat release rate of different ice types, combined with a temporal convolutional network for deep decoding of transient thermal response curves, successfully establishing a nonlinear mapping relationship between interfacial transient thermal damping and mechanical adhesion force. This technical solution enables the sensor not only to answer "is there ice?", but also to accurately determine "what kind of ice" and "how tightly it adheres," thus solving the core pain point of existing technologies—namely, the lack of adhesion force data leading to crude control strategies and low energy efficiency in de-icing systems.

[0049] This invention significantly improves the robustness and response speed of sensors in complex aerodynamic environments at the algorithm level. Addressing the problem that existing strain gauge sensors are highly susceptible to false alarms caused by aerodynamic loads from high-speed airflow and random environmental vibrations, this invention employs a TCN network structure that combines dilated causal convolution with a self-attention mechanism. This allows it to automatically focus on microscopic phase transition thermodynamic features (such as latent heat plateaus) that are insensitive to environmental mechanical noise, effectively shielding the influence of external vibration noise and maintaining extremely high detection confidence even under harsh conditions of strong convection and vibration. Furthermore, compared to traditional energy-intensive schemes that rely on completely heating and melting ice layers to infer ice thickness, this invention only requires microsecond to millisecond-level micro-energy thermal pulses to complete detection. While maintaining high-precision classification, it significantly reduces system power consumption and extends the service life of wireless sensor nodes. Attached Figure Description

[0050] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The scope of protection of the present invention is not limited to the following description.

[0051] Figure 1 This is a flowchart of the present invention.

[0052] Figure 2 This is a block diagram illustrating the principle of the present invention.

[0053] Figure 3 This is a structural diagram of the model of the present invention.

[0054] Figure 4 This is a schematic diagram of the control process for a specific industrial implementation of the present invention.

[0055] Figure 5This is a thermal field distribution diagram of an embodiment of the present invention. Detailed Implementation

[0056] Experimental verification and quantitative analysis of the technical effects of this invention

[0057] Multiphase Comparison Experiment and Accuracy Verification Based on Icing Wind Tunnel

[0058] To verify the detection performance of this invention under meteorological conditions, the growth processes of Glaze, Rime, and mixed ice were artificially reproduced using a high-low temperature icing wind tunnel experimental platform by precisely controlling the liquid water content (LWC), median volume diameter (MVD) of water droplets, and ambient airflow temperature. The active thermal disturbance sensor proposed in this invention was then compared with a traditional passive pure frequency detection sensor under the same operating conditions.

[0059] The experiment included three typical operating conditions: open ice (ambient temperature -5℃, high liquid water content), frost ice (ambient temperature -15℃, low liquid water content), and mixed ice. Each condition underwent 100 independent icing detection cycles. The control group used a traditional passive pure frequency detection test piece, while the experimental group used the active thermal disturbance MEMS test piece of this invention.

[0060] Statistical results show that traditional passive methods are completely ineffective in classifying ice types due to the inherent limitations of their physical mechanisms. Specifically, traditional methods rely solely on the scalar accumulation of mass load perceived by the Sauerbrey equation. When 1 gram of dense clear ice and 1 gram of loose frost ice adhere to the sensor surface, the static frequency offsets they cause are extremely similar. Experimental data confirms that when faced with mixed blind testing of three ice types, the classification accuracy of traditional passive methods hovers only around 33% (approximately equal to 1 / 3 of a mathematically random guess), and due to the lack of interfacial thermal resistance and strain data, it is completely impossible to construct an adhesion force prediction model (output value is N / A).

[0061] In contrast, this invention achieves precise classification and quantitative assessment. Instead of relying on static mass absolute values, this invention captures the "thermodynamic fingerprints" of clear ice (high thermal conductivity, large latent heat release) and frost ice (low thermal conductivity, internal porosity) during phase change heat transfer by stimulating unsteady-state thermal responses. Experimental test data (see table below) show that the classification accuracy of this invention for clear ice and frost ice jumps to 96.8% and 95.5%, respectively. Simultaneously, through regression prediction using a dual-head TCN network, the output ice shear adhesion force values ​​highly match the actual physical values ​​simultaneously measured using a centrifugal tensile testing machine, with a root mean square error (RMSE) as low as 12.4 kPa.

[0062] Table 1: Performance Comparison of Passive Pure Frequency Detection and the Active Thermal Disturbance Method of the Present Invention under Multiphase Conditions

[0063] Glaze 33.1% 96.8% Unable to output (N / A) ± 12.4 kPa Rime (Frost Ice) 34.5% 95.5% Unable to output (N / A) ± 10.8 kPa Mixed ice 32.8% 91.2% Unable to output (N / A) ± 15.6 kPa

[0064] Experimental data statistics show that traditional passive methods, which can only sense the scalar accumulation of mass load, are completely ineffective in classifying ice types (the accuracy is approximately equal to random probability) and cannot output any adhesion force information.

[0065] In contrast, this invention, by capturing microscopic thermodynamic fingerprints and using a dual-headed TCN network for decoding, achieved classification accuracies of 96.8% and 95.5% for clear ice and frost ice, respectively. Even under complex mixed ice conditions, the accuracy remained above 91.2%. In the adhesion force prediction task, for the actual shear adhesion strength range of 0 to 500 kPa, the predicted values ​​output by the model of this invention highly matched the measured values ​​obtained by the centrifugal tensile testing instrument, with a root mean square error (RMSE) of only 12.4 kPa and a coefficient of determination (...). The accuracy reached 0.94, which fully demonstrates the excellent precision and reliability of this method in quantitatively assessing the mechanical properties of ice.

[0066] Robustness and generalization ability test under strongly coupled noise environment

[0067] Considering that aircraft wings or wind turbine blades are subjected to alternating loads of high-frequency mechanical vibration and strong aerodynamic noise for extended periods during actual service, this embodiment incorporates a rigorous anti-interference robustness test. Random broadband sweep vibrations with an amplitude of 5g and a frequency range of 10Hz to 2000Hz, along with turbulent aerodynamic disturbances at Mach number 0.3, were superimposed in the aforementioned wind tunnel environment.

[0068] Table 2: Comparison of robustness test results under combined aerodynamic and vibration noise environments (sample size N=200 per group)

[0069] Traditional piezoelectric / strain type passive detection 2.5% 36.2% Unable to perform effective statistics (due to false alarm interference). Frequency deviation ±150Hz / Stress ±45kPa This invention (Active Thermal Disturbance + TCN) 0.5% 1.5% 92.5% SNR tolerance improved by 16dB

[0070] Test results show that traditional piezoelectric or strain gauge icing sensors are affected by mechanical load coupling when subjected to airflow impact or resonant point frequency sweep. Specific data shows that the reference frequency drift of traditional passive sensors reaches ±150Hz (or equivalent stress baseline fluctuation reaches ±45kPa), exceeding the dead zone range set by the static threshold, causing the system's false alarm rate (FAR) to rise to 36.2%.

[0071] In contrast, this invention extracts the unsteady thermodynamic features of the microscopic heat transfer and phase change processes in ice layers (corresponding to low-frequency thermal relaxation time series), achieving frequency band isolation from high-frequency mechanical vibrations and aerodynamic loads through a physical mechanism. Combined with feature extraction using a temporal convolutional network (TCN), this invention controls the false alarm rate to 1.5% under conditions of superimposed composite noise. In ice type classification tasks, the classification accuracy of this invention is 92.5%; the system signal-to-noise ratio (SNR) tolerance is improved by 16 dB compared to traditional threshold determination methods, verifying the method's anti-interference capability under alternating load environments.

[0072] Influence Mechanism and Optimization Boundary Analysis of Core Thermal Excitation Parameters

[0073] Regarding the core excitation source of "active thermal disturbance," the duration of the thermal pulse ( ) and energy density ( The parameter is a key boundary parameter that determines the system's detection resolution and energy efficiency. This invention conducts a parametric sweep of the system through multiphysics coupled simulation (COMSOL) combined with controlled variable experiments.

[0074] The optimal operating parameter range of this invention was established through optimization scanning: the pulse duration was set to 80ms to 150ms, and the energy density was set to 2.0 to 5.0 W / cm². Within this range, a stable phase change layer of about 2 to 5 μm can be formed at the interface, the ice type classification accuracy is maintained above 95%, and the total energy consumption of a single thermal pulse is controlled within 20mJ.

[0075] Table 3: Active thermal disturbance core parameter (tp, qp) scan test data and system performance evaluation

[0076] 20 0.5 <0.1 62.4% 0.25 Insufficient stimulation, thermal characteristics not activated 50 1.0 0.5 78.5% 1.25 Critical state, low classification confidence 100 3.0 2.5 95.8% 7.5 Optimal parameter range, balancing high accuracy and low power consumption 150 5.0 4.2 96.5% 18.75 The optimal parameter domain yields the most pronounced characteristic responses. 300 10.0 >50.0 85.2% 75.0 Excessive stimulation leads to macroscopic slip disturbances in the ice layer.

[0077] Test analysis shows that the excitation parameters must match the phase change thermodynamics of the ice layer: when the pulse duration tp ≤ 20 ms or the energy density qp ≤ 0.5 W / cm², the thickness of the micro-melting layer formed at the sensor-ice interface is less than 0.1 μm. This depth fails to reach the physical threshold for effectively exciting the transient thermal damping difference between clear ice and frost ice, leading to the failure of model feature extraction and a drop in ice type classification accuracy to below 65%. Conversely, if tp ≥ 300 ms or qp ≥ 10 W / cm², the thickness of the micro-melting layer at the interface will exceed 50 μm, triggering the risk of macroscopic slippage of the ice layer under the action of airflow, destroying the non-destructive detection mechanism, and significantly increasing the energy consumption of a single detection.

[0078] The hardware architecture of this invention uses a high-frequency piezoelectric resonator as the core sensing unit. Specifically, the MEMS sensor substrate is an AT-cut quartz crystal microbalance (QCM) with a fundamental frequency of 5MHz. This cut has an extremely low frequency temperature drift coefficient over a wide temperature range, thus providing a stable detection reference. On the non-sensing surface (i.e., the back side) of the QCM, a 20nm thick titanium (Ti) bonding layer and a 150nm thick platinum (Pt) thin film are sequentially sputtered using a magnetron sputtering physical vapor deposition (PVD) process. This Pt thin film is photolithographically patterned into a 5mm × 5mm serpentine trace, which simultaneously serves as a micro-heater and a temperature-sensing resistance thermometer (RTD, initially designed with a Pt1000 standard resistance). This integrated thin-film deposition process not only ensures extremely low heat capacity, enabling the system to have millisecond-level thermal response capabilities, but also eliminates the interface thermal resistance caused by discrete devices.

[0079] To assess the uniformity and controllability of the thermal field distribution, transient heat conduction simulations of the integrated microheater were performed using COMSOL Multiphysics finite element software. Simulation results (thermal field distribution diagram, as shown) are presented below. Figure 5 As shown in the figure, under typical pulse width (e.g., 100 ms) and typical power density (e.g., 5 W / cm²), 2 During the square-wave thermal pulse, the temperature distribution within the heating area (5mm×5mm) is highly uniform. Employing a patterned serpentine trace design and a thin adhesive layer, the maximum transient temperature difference between the center and edge of the heating area does not exceed 0.5℃, effectively preventing localized overheating or "cold spots." This uniform, controllable, and hysteresis-free ideal thermal field ensures the smooth advancement of the micro-melting phase transition boundary at the ice-sensor interface, a core physical prerequisite for subsequent algorithms to extract pure, high signal-to-noise ratio "thermodynamic fingerprints."

[0080] In terms of signal acquisition and control subsystem, combined with Figure 2 The system block diagram shown can employ a hierarchical heterogeneous computing architecture. The underlying real-time control and signal conditioning unit uses STMicroelectronics' high-performance microcontroller STM32H743XI. This chip is connected via the SPI bus to a high-precision impedance converter network analyzer chip (Analog Devices' AD5933), responsible for applying the sweep excitation signal to the QCM and extracting the resonant frequency in real time at a sampling rate of 1kHz. With the real part of the complex impedance When the sliding window algorithm running inside the STM32 determines that the icing trigger condition is met (step S1), its internal DAC module will output a control signal to drive the external high-power MOSFET switch to apply a square wave voltage with an amplitude of 5V and a pulse width of 100ms to the Pt thin film heater (step S2), thereby accurately injecting a quantitative heat flux at the ice-sensor interface.

[0081] Deploying and executing at the algorithm edge, combined with Figure 3 The model structure diagram shown can be implemented using the NVIDIA Jetson Orin Nano edge computing core board as the host computer processing platform. The software environment is based on the Ubuntu 20.04 operating system, and the underlying algorithm is trained offline using Python 3.8 and the PyTorch 2.0 deep learning framework. To meet the real-time requirements of the embedded platform, the trained TCN dual-head network model undergoes graph optimization and precision quantization (FP16 half-precision) using Tensor RT, and is encapsulated as an independent inference engine. In the data acquisition phases of steps S3 and S4, the STM32 transmits the transient response data (3×2000 tensors) for a total of 2 seconds before and after the thermal pulse trigger to the Jetson platform at high speed via DMA. Subsequently, the preprocessing module within the Jetson platform performs Z-Score normalization on the tensors and feeds them into the TensorRT engine.

[0082] After forward propagation, the dilated causal convolutional layer in the TCN network backbone rapidly extracts the "thermodynamic fingerprint" containing micro-melting and re-icing characteristics, and the self-attention module assigns high weights to data frames during the latent heat release period (steps S5 and S6). Finally, in step S7, the engine's classification head outputs a Softmax probability vector of length 4 (e.g., a probability of classifying it as "visible ice" of 98.2%), and the regression head simultaneously outputs a linear scalar (e.g., a predicted adhesion shear strength of 215.4 kPa). This integrated decision result is finally sent to the aircraft's de-icing control computer (IPS) via a CAN bus or Ethernet interface, guiding it to activate the electrothermal anti-icing system with optimal power and timing, thus completing a full closed-loop intelligent detection process.

[0083] Example

[0084] The application scenario is the de-icing system on the leading edge of the wing of the Wing Loong-2H high-altitude long-endurance unmanned aerial vehicle (UAV). Three MEMS icing sensors are arrayed in the stagnation area of ​​the wing leading edge. The underlying sensing element of each sensor is an AT-cut quartz crystal microbalance (QCM) with a fundamental frequency of 5.000MHz. The non-sensing surface is integrated with an area of ​​[area missing] via magnetron sputtering. A Pt thin-film serpentine micro-heater with a nominal resistance of 1000Ω is used. The system's bottom-level control unit uses an STM32H743XI microcontroller (with an AD5933 impedance analysis chip), and the edge computing unit uses an onboard NVIDIA Jetson Orin Nano core board, which is connected to the UAV's electrothermal anti-icing system (IPS) via a CAN bus.

[0085] The initialization settings of the known parameters and variables involved in this embodiment are as follows:

[0086] : Sliding window length, set to 2 seconds (including 2000 sampling points).

[0087] : The reference resonant frequency in the ice-free state, 5,000,000 Hz.

[0088] Minimum icing detection threshold is set to decrease by 15Hz.

[0089] Noise tolerance threshold, set to 2Hz (to filter transient high-frequency jitter caused by airflow disturbance).

[0090] Active thermal disturbance pulse width, set to 120ms.

[0091] The amplitude of the active thermal disturbance pulse voltage is set to 5V.

[0092] Based on the above hardware scenario and parameter definitions, the specific industrial implementation control process of the method of this invention is described as follows: Figure 4 As shown, it includes the following steps:

[0093] Step 1: Real-time monitoring and trigger determination of low power consumption.

[0094] The drone cruised through clouds at an altitude of 3000 meters and an ambient temperature of -8°C. The STM32 microcontroller acquired the resonant frequency of the QCM in real time at a sampling rate of 1kHz. At this moment, supercooled water droplets strike the leading edge of the wing and freeze. The STM32 calculates the current 2-second sliding window. Frequency statistics within the window. When the mean frequency within the window is found to drop to 4,999,980 Hz (i.e., the offset exceeds...), the frequency statistics are considered to be... ), and the frequency variance is lower than At that time, the system determines that there is a stable ice load and triggers the active detection process.

[0095] Step 2: Apply a non-destructive microsecond-level thermal pulse.

[0096] Upon triggering, the STM32's internal DAC module immediately outputs an enable signal, closing the external power MOSFET switch and applying an amplitude of [value missing] to the Pt thin-film microheater. Duration A square wave pulse. This action injects a transient heat flux (power density approximately) into the sensor-ice interface. Without causing macroscopic ice layer detachment, a micro-melting phase transition layer with a thickness of about 2.5 μm was excited at the interface.

[0097] Step 3: Acquisition of multidimensional thermodynamic response.

[0098] Within a 2-second period from 0.5 seconds before the thermal pulse trigger to 1.5 seconds after it ends, the AD5933 chip synchronously acquires the resonant frequency shift sequence and the real part sequence of the complex impedance at a high frequency (1kHz), and determines the temperature gradient sequence using the temperature coefficient of resistance (TCR) of the Pt thin film itself. The STM32 packages a three-dimensional data packet containing 2000 discrete time points through a DMA channel.

[0099] Step 4: Edge side tensor construction and preprocessing.

[0100] After receiving data packets, the airborne Jetson node first performs Z-Score normalization to eliminate baseline drift, and then performs channel stacking to build a dimension of [missing information]. The temporal feature tensor.

[0101] Step 5: TCN network feature extraction and decoding.

[0102] The aforementioned tensors were input into a pre-deployed deep temporal convolutional network (TCN) model (accelerated by FP16 quantization using TensorRT). The shallow dilated causal convolutions at the beginning of the network quickly extracted the frequency transient recovery features (high-frequency fluctuations) caused by the micro-melting of the ice layer during the first 120ms heating period; the subsequent self-attention mechanism automatically focused on the "latent heat release relaxation period" from 100ms to 800ms after the pulse ended. Since the current ice is dense, highly thermally conductive open ice, this relaxation segment exhibits a characteristic smooth, long-tailed recovery curve.

[0103] Step 6: Dual-head task output and industrial closed-loop control.

[0104] After a single forward propagation (taking only 12ms), the TCN classification head outputs the judgment result: the current ice type is "Glaze", with a confidence level of 98.2%; the TCN regression head simultaneously outputs the predicted ice layer shear adhesion force: 285kPa.

[0105] The Jetson node immediately sent the status command "[Exposed ice, 285kPa]" to the UAV flight control center via the CAN bus. Based on this quantified physical indicator, the flight control center determined that the conventional anti-icing power was insufficient to deal with the highly adhesive exposed ice. Therefore, it instantly activated the "high-power short-time de-icing mode" of the wing leading edge electrothermal de-icing system (IPS), precisely peeling off the ice before it accumulated over a large area, completing an adaptive closed-loop anti-icing and de-icing control.

[0106] Comparative analysis of the effects of the examples:

[0107] To verify the superiority of the above embodiments in industrial applications, the present invention compared the performance of a UAV wing test piece equipped with the method with a test piece equipped with a conventional passive detection method (alarmed only by frequency threshold) in an icing wind tunnel.

[0108] 1) Regarding ice type classification and adhesion force sensing: As shown in Table 1, conventional methods, when faced with blind testing of clear ice and frost ice, only sense scalar mass, resulting in a classification accuracy of only 33.1% (in a failed state), and cannot output adhesion force. In contrast, this invention utilizes micro-thermal perturbation to excite phase transition fingerprints, achieving a stable classification accuracy of over 96.8%, and exhibiting extremely small adhesion force prediction error (RMSE = 12.4 kPa), providing accurate decision-making basis for the IPS system.

[0109] 2) Regarding noise robustness under extreme conditions: As shown in Table 2, when a wind tunnel experiences mechanical vibration with an amplitude of 5g and turbulent disturbance with Ma=0.3, the false alarm rate of the conventional method soars to 36.2% due to the severe drift of the stress baseline. However, this invention relies on the adaptive focusing of the TCN network on low-frequency phase transition thermal characteristics, physically isolating high-frequency mechanical noise at the algorithm level, strictly controlling the false alarm rate to 1.5%, thus ensuring the reliability of UAV detection in harsh convective environments.

[0110] 3) In terms of system energy consumption: Traditional detection methods that completely melt the ice layer consume energy in the hundreds of joules (J) per detection; while the present invention is based on the principle of "microscopic micro-melting", and the total energy consumption of a single active detection is as low as about 7.5 mJ. The impact on the overall energy system of the UAV is negligible, and the engineering contradiction between high-precision detection and low-power-constrained environment is completely solved.

[0111] In summary, the evaluation method based on active thermal disturbance response and temporal convolutional network proposed in this invention can achieve a leap from purely qualitative alarm to "ice type classification + adhesion force quantification" under complex aerodynamic and mechanical noise interference through low-power active physical excitation and deep temporal decoding. It is a technical solution with extremely high industrial implementation value in the field of aviation anti-icing and de-icing.

Claims

1. A method for classifying MEMS ice types and evaluating adhesion, characterized in that... The steps include the following: Step S1: Icing trigger monitoring based on sliding window statistics; Step S2: Application of active thermal disturbance pulse and generation of heat flux; Step S3: Transient response mechanism based on multilayer medium heat transfer model; Step S4: Construction and preprocessing of multidimensional temporal feature tensors; Step S5: Construct a dilated causal convolutional network to extract thermodynamic fingerprints; Step S6: Weighted keyframes based on self-attention mechanism; Step S7: Multi-task loss function optimization and output decision.

2. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step S1: The system is first in a low-power monitoring mode, and the MEMS sensor collects the resonant frequency f(t) data stream in real time. Construct a time sliding window of length W and calculate the statistical characteristics of the frequency signal within the window; Define the mean frequency within the window at time t. and variance The system is identified as potentially icing and triggers a subsequent active thermal disturbance mode only when both conditions are met: "significant decrease in average frequency" and "signal in steady state". The triggering logic is shown in formula (1). in, This is the reference frequency under ice-free conditions. The minimum icing detection threshold is set. The noise tolerance threshold is used; Formula (1) ensures that the system only starts heating when there is a stable attachment, avoiding false starts under airflow disturbance or strong vibration.

3. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step S2: Once the triggering mechanism is activated, the control unit applies a voltage pulse V(t) with a specific timing to the platinum thin-film heater integrated on the MEMS surface; this pulse waveform is designed as a square wave or trapezoidal wave, with a duration of [missing information]. According to Joule's law, the applied voltage pulse generates an instantaneous heat flux q(t) at the sensor-ice interface; due to the slight change in the heater's resistance R(T) with temperature, the generated real-time heat power... As shown in formula (2): in, The resistance at the reference temperature, Temperature coefficient of resistance The thermal power is the indicative function; this thermal power, as an energy source term, disrupts the thermal equilibrium state of the ice-sensor interface, aiming to induce a microscopic phase change at the interface rather than completely de-icing.

4. The MEMS ice type classification and adhesion assessment method according to claim 3, characterized in that... The duration in step S2 is The duration is 50ms-200ms; The value is 1 during the pulse duration and 0 at other times.

5. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step 3: Temperature field distribution on the sensor surface under the action of thermal pulse. It follows a one-dimensional unsteady-state heat conduction differential equation; Different types of ice (clear ice, frost ice) have different thermal conductivity. and density The response to the thermal pulse, i.e., the temperature decay process, is completely different; this physical process is described by formula (3): in Volumetric heat capacity; resonant frequency shift of MEMS sensor Not only is it affected by mass load, but it is also greatly affected by temperature effect during thermal disturbance; the coupling equation between frequency response and temperature change is established as shown in formula (4): The first term is the Sauerbrey mass load term, which is considered constant during short-duration thermal pulses; the second term is the thermal effect term. The frequency temperature coefficient of each layer of material; By observing formula (4) The curve of change over time is actually the thermal conductivity k and heat capacity characteristics of the ice layer in the inversion formula (3), which is the physical essence of distinguishing ice types.

6. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step 4: The system synchronously acquires the time window before and after the thermal pulse trigger. Frequency drift sequence within Impedance real part sequence and temperature gradient sequence To eliminate individual differences among sensors, the raw data is first standardized using Z-Score. Subsequently, the standardized sequences of the three dimensions are stacked in channels to construct the model input tensor. Let the number of sampling points be... The input tensor is represented by formula (5): This tensor Completely includes thermal disturbances caused by The change in mechanical stiffness is characterized by Characterized interfacial phase transition loss and its origin Information on the thermal conductivity rate.

7. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step 5: Transfer tensors The input is fed into a temporal convolutional network; the core layer of the TCN employs dilated causal convolution; for the input sequence of the l-th layer... Its output at time t Defined as formula (6): in This represents the dilated convolution operator. For convolution kernel, The inflation factor; an inflation factor with exponential growth. This causes the receptive field of deep networks to expand exponentially with the number of layers.

8. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step 6: Introduce a self-attention module after the TCN layer; the self-attention module calculates the input feature sequence. The internal correlation of the attention weight matrix automatically focuses on the key time step where the phase transition occurs. The calculation is as shown in formula (7): Final weighted feature vector The product of the weight matrix and the value vector: ; Using formula (7), the algorithm adaptively identifies the most distinctive instantaneous features that distinguish clear ice from frost ice.

9. The MEMS ice type classification and adhesion assessment method according to claim 1, characterized in that... Step 7: The network's output layer is designed as a dual-head structure, performing ice type classification and adhesion regression tasks respectively; a joint loss function is designed. As shown in formula (8): The first term is the cross-entropy loss for the classification task, used to optimize ice type determination; the second term is the mean squared error loss for the regression task, used to approximate the true ice shear adhesion strength. ; To balance the hyperparameters of the two task weights, the system ultimately outputs the ice category probability and the specific adhesion force value.

10. The method for MEMS ice type classification and adhesion assessment according to claim 1, characterized in that... The MEMS sensor substrate uses an AT-cut quartz crystal microbalance with a fundamental frequency of 5MHz. On the non-sensing surface of the QCM, a titanium bonding layer with a thickness of 20nm and a platinum thin film with a thickness of 150nm are sequentially sputtered using a magnetron sputtering physical vapor deposition process. The platinum thin film is photolithographically patterned into a serpentine trace with an area of ​​5mm×5mm, which also serves as a micro heater and a temperature measuring resistance thermometer.

Images