A method and apparatus for self-cleaning of an immersion probe window

Abstract

The present application provides a kind of immersion probe window self-cleaning method and device, belong to immersion probe technical field, the present application constructs acoustic pressure distribution prediction model using physical information neural network to optimize transducer installation position, adopts square wave signal generator to realize the time-sharing control of cleaning and detection, utilizes dynamic time warping algorithm to match the real-time transducer impedance change time sequence with cleaning success time sequence template library to adaptively determine the cleaning completion state, determines the detection equipment start time dynamically through vibration attenuation real-time monitoring module, and the cleaning intensity is closed-loop regulated by artificial intelligence light window pollution evaluation model and driving power adaptive adjustment function, solve the technical problems that cleaning completion state cannot be adaptively determined in the self-cleaning process of immersion probe window, leading to fixed cleaning time and unable to adapt to different pollution conditions.

Description

Technical Field

[0001] This invention belongs to the field of immersion probe technology, and more specifically, relates to a method and apparatus for self-cleaning the window of an immersion probe. Background Technology

[0002] Immersion probes are widely used in industrial process analysis, online water quality monitoring, and chemical reaction monitoring. The probe's viewing window is in direct contact with the medium for extended periods, making its surface susceptible to contaminant adhesion such as particulate matter, crystal precipitates, and biofilms, severely impacting the quality of the acquired signal. To maintain normal probe operation, current technologies typically employ ultrasonic vibration cleaning. This method applies high-frequency vibration energy to the viewing window, utilizing acoustic radiation and microjets to peel away surface deposits, achieving periodic cleaning. In engineering practice, existing ultrasonic self-cleaning devices generally employ a fixed-duration control strategy. This means a pre-set, uniform cleaning duration is maintained, and cleaning is performed for the fixed duration regardless of changes in the window's contamination level, medium viscosity, or temperature conditions, forcibly switching the detection state accordingly.

[0003] However, the adhesion strength of pollutants, the rheological properties of the medium, and the coupling state of the transducer vary significantly under different operating conditions. Fixed-duration control strategies cannot guarantee thorough cleaning under heavy pollution conditions and cause unnecessary monitoring interruptions under light pollution conditions, severely compressing the effective working time of real-time monitoring equipment. Furthermore, existing technologies lack real-time status sensing methods for the cleaning process, failing to determine the current cleaning stage based on dynamic changes in the transducer load state during cleaning, resulting in a systematic deviation between the cleaning termination timing and the actual cleaning completion status.

[0004] In existing technologies, ultrasonic self-cleaning devices employ a fixed-duration cleaning control strategy, lacking the ability to perceive and adaptively determine the real-time status of the cleaning process. This prevents the dynamic determination of when to terminate the cleaning based on varying degrees of contamination, media viscosity, and temperature changes, leading to an irreconcilable contradiction between thorough cleaning and effective detection duty cycle. In other words, existing technologies suffer from the technical problem of being unable to adaptively determine the completion status of the cleaning process during the self-cleaning of the immersion probe window, resulting in a fixed cleaning time that cannot adapt to different contamination conditions. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for self-cleaning an immersion probe window, which can solve the technical problem in the prior art that the cleaning completion status of the immersion probe window cannot be adaptively determined during the self-cleaning process, resulting in a fixed cleaning time that cannot adapt to different pollution conditions.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides a self-cleaning method for an immersion probe window, comprising the following steps:

[0007] An acoustic impedance gradient matching intermediate layer is filled into the contact interface between the ultrasonic transducer and the viewing window glass. A pre-tightening force is applied by a fixed metal pressure ring to ensure uniform contact at the interface. The acoustic impedance range and thickness value of the acoustic impedance gradient matching intermediate layer are solved based on the optimal coupling layer thickness solution algorithm of acoustic impedance gradient descent, and the optimal acoustic impedance range and optimal thickness value are output.

[0008] The optimal acoustic impedance range, optimal thickness value, transducer driving frequency, transducer driving power, and window glass geometric parameters are input into the sound pressure distribution prediction model. The output is the sound pressure distribution field and vibration displacement distribution on the window glass surface. Based on the sound pressure distribution field and vibration displacement distribution on the window glass surface, the transducer installation position parameters are determined to guide the installation of the ultrasonic transducer.

[0009] The square wave signal generator generates a square wave signal. The rising edge triggers the ultrasonic signal generator to drive the ultrasonic transducer to apply high-frequency vibration to the viewing window glass to perform self-cleaning. The falling edge triggers the signal switching module to switch to the real-time detection equipment working state.

[0010] After the ultrasonic transducer completes a single cleaning, the transducer impedance change time sequence is collected. The dynamic time warping cleaning process identification algorithm is used to match the transducer impedance change time sequence with the cleaning success time sequence template library. The cleaning completion judgment signal, the current cleaning stage and the predicted remaining cleaning time are output. The cleaning process is dynamically terminated or continued according to the cleaning completion judgment signal.

[0011] After cleaning, the vibration attenuation real-time monitoring module collects the residual vibration amplitude of the window glass through the accelerometer and compares the residual vibration amplitude with the attenuation threshold. When the residual vibration amplitude is lower than the attenuation threshold, the trigger delay time is output. The signal switching module starts the real-time detection device according to the trigger delay time.

[0012] The real-time detection device acquires window images, and the lightweight window contamination assessment model evaluates the degree of contamination in the window images and outputs a contamination score. The drive power adaptive adjustment function receives the contamination score, transducer impedance offset, and vibration decay time, and outputs a comprehensive adjustment index. The transducer drive power is adjusted according to the comprehensive adjustment index. When the contamination score is higher than the contamination trigger threshold, the self-cleaning step is re-triggered. When the contamination score is lower than the contamination trigger threshold, the real-time detection device remains in operation.

[0013] The acoustic impedance of the intermediate layer in the acoustic impedance gradient matching is between the acoustic impedance of the ultrasonic transducer material and the acoustic impedance of the window glass material, forming a gradient structure in which the acoustic impedance gradually transitions from the ultrasonic transducer side to the window glass side, which is used to reduce the reflection loss of acoustic energy at the interface.

[0014] Among them, the optimal coupling layer thickness solution algorithm of acoustic impedance gradient descent takes the acoustic energy transmission coefficient as the objective function, takes the acoustic impedance and thickness of the acoustic impedance gradient matching intermediate layer as the optimization variables, uses the transfer matrix method to establish the acoustic transmission equation system of the three media of ultrasonic transducer, acoustic impedance gradient matching intermediate layer and window glass, and uses the gradient descent method to iteratively update the optimization variables to maximize the vibration displacement amplitude of the window glass surface.

[0015] The iteration termination condition of the optimal coupling layer thickness solution algorithm for acoustic impedance gradient descent is that the change in vibration displacement amplitude between two adjacent iterations is less than a set threshold, and finally outputs the optimal acoustic impedance range and the optimal thickness value.

[0016] The sound pressure distribution prediction model is based on a physical information neural network. The input layer receives the optimal acoustic impedance range and optimal thickness value, transducer driving frequency, transducer driving power and window glass geometric parameters. The output layer outputs the sound pressure value and vibration displacement value at each position on the window glass surface. The loss function embeds the acoustic wave equation and boundary condition constraints, and the network output satisfies the physical constraints through automatic differential iteration.

[0017] The training dataset for the sound pressure distribution prediction model is obtained by performing finite element simulations on different combinations of transducer driving frequencies, transducer driving power, and window glass geometric parameters. Accelerometers are then deployed on the physical device to collect measured data, and the simulation results are fused with the measured data.

[0018] The training of the sound pressure distribution prediction model involves dividing the training dataset into a training set and a validation set in an 8:2 ratio. The total loss function is the weighted sum of the mean squared error loss and the physical residual loss. The Adam optimizer is used for iterative training, and training is terminated when the validation set loss no longer decreases.

[0019] Among them, the dynamic time warping cleaning process identification algorithm pre-establishes a cleaning success time sequence template library covering different types of contamination. The dynamic time warping distance metric method is used to calculate the dynamic time warping distance between the transducer impedance change time sequence and each template in the cleaning success time sequence template library. The template with the smallest dynamic time warping distance is selected as the matching template for the current cleaning process.

[0020] The dynamic time warping cleaning process identification algorithm determines the current cleaning stage and outputs the predicted remaining cleaning time based on the alignment position of the transducer impedance change timing in the matching template. When the normalized dynamic time warping distance between the transducer impedance change timing and the final state of the matching template is less than the final state neighborhood judgment threshold, a cleaning completion judgment signal is output.

[0021] Among them, the vibration attenuation real-time monitoring module continuously collects the residual vibration amplitude of the window glass after the ultrasonic transducer stops excitation, compares the residual vibration amplitude with the attenuation threshold and outputs the trigger delay time, and the signal switching module dynamically determines the start time of the real-time detection device based on the trigger delay time.

[0022] Among them, the adaptive adjustment function of driving power calculates the comprehensive adjustment index. The comprehensive adjustment index is obtained by weighted summation of three factors: the ratio of pollution score to pollution score reference value, the ratio of transducer impedance offset to transducer impedance reference value, and the ratio of vibration decay time to vibration decay time reference value. The sum of each weight coefficient is 1.

[0023] The adaptive adjustment function of driving power adjusts the transducer driving power in segments according to the range of the comprehensive adjustment index: when the comprehensive adjustment index is less than 0.3, the current driving power is maintained; when the comprehensive adjustment index is between 0.3 and 0.7, the driving power is increased by 15%; when the comprehensive adjustment index is between 0.7 and 1.0, the driving power is increased by 30%; when the comprehensive adjustment index is not less than 1.0, the driving power is increased to the rated upper limit and the extended cleaning process is triggered.

[0024] A second aspect of the present invention provides an immersion probe window self-cleaning device, comprising a probe housing, a sealing gasket, a viewing window, an ultrasonic transducer, a fixing metal ring, an ultrasonic signal generator, a square wave signal generator, a signal switching module, and a real-time detection device; the viewing window is installed at the open end of the probe housing via the sealing gasket, and the other end of the viewing window is connected to the ultrasonic transducer; the fixing metal ring fixes the ultrasonic transducer and the viewing window inside the probe housing.

[0025] The fixed metal pressure ring ensures that the sealing gasket is tightly fitted to the probe housing, the viewing window is tightly fitted to the sealing gasket, and the viewing window is tightly fitted to the ultrasonic transducer. The ultrasonic signal generator generates a high-frequency electrical signal, which is amplified and then transmitted to the ultrasonic transducer.

[0026] Among them, the lightweight window contamination assessment model is obtained by compressing the large image segmentation model using the knowledge distillation method. Combined with INT8 quantization and TensorRT deployment, it is used to segment and score the contamination level of window images in real time and output a normalized contamination score.

[0027] The device also includes a vibration attenuation real-time monitoring module, which is an acceleration sensor signal acquisition and processing unit integrated into the device circuit. It is used to continuously acquire the residual vibration amplitude of the window glass after the ultrasonic transducer stops excitation, and output the trigger delay time after comparing the residual vibration amplitude with the attenuation threshold.

[0028] This invention maximizes the transmission efficiency of vibration energy to the viewing window glass by constructing an algorithm for solving the thickness of the intermediate layer and the optimal coupling layer for acoustic impedance gradient matching; optimizes the transducer installation position through a sound pressure distribution prediction model to ensure uniform distribution of vibration energy on the viewing window glass surface; matches the transducer impedance change sequence with a cleaning success sequence template library through a dynamic time warping cleaning process identification algorithm to identify the current cleaning stage in real time and adaptively determine the cleaning completion status; dynamically determines the start-up time of the detection equipment through a vibration attenuation real-time monitoring module; and achieves coordinated control of the cleaning and detection processes by using a lightweight viewing window contamination assessment model and a closed-loop control function for adaptive adjustment of driving power, thus solving the technical problem of the inability to adaptively determine the cleaning completion status.

[0029] This invention utilizes a dynamic time warping distance measurement method to match the real-time acquired transducer impedance change time sequence with a cleaning success time sequence template library covering different types of pollution, under the condition that the time sequence pattern has elastic expansion and contraction. This makes the identification of the cleaning stage unaffected by the difference in cleaning rate, thereby ensuring the thoroughness of cleaning under heavy pollution conditions and terminating cleaning in a timely manner under light pollution conditions, minimizing unnecessary cleaning time, and improving the effective duty cycle of the real-time detection equipment.

[0030] In summary, the present invention solves the technical problem mentioned in the background art that the cleaning completion status of the immersion probe window cannot be adaptively determined during the self-cleaning process, resulting in a fixed cleaning time that cannot adapt to different pollution conditions. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention.

[0032] Figure 2 This is a schematic diagram of the device provided by the present invention.

[0033] Figure 3 A comparison chart showing the predicted results of vibration displacement distribution on the surface of the viewing window glass.

[0034] The reference numerals in the attached diagram are explained as follows: 1. Probe, housing; 2. Sealing gasket; 3. Viewing window glass; 4. Ultrasonic transducer; 5. Fixing metal pressure ring; 6. Ultrasonic signal generator; 7. Square wave signal generator; 8. Signal switching module; 9. Real-time detection equipment. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0036] like Figure 1 As shown, the first aspect of the present invention provides a self-cleaning method for an immersion probe window, comprising the following steps:

[0037] S01. Fill the contact interface between the ultrasonic transducer and the window glass with the acoustic impedance gradient matching intermediate layer, and apply a pre-tightening force by fixing the metal pressure ring to ensure uniform contact of the interface; solve the acoustic impedance range and thickness value of the acoustic impedance gradient matching intermediate layer based on the optimal coupling layer thickness solution algorithm of acoustic impedance gradient descent, and output the optimal acoustic impedance range and optimal thickness value.

[0038] S02. Input the optimal acoustic impedance range, optimal thickness value, transducer driving frequency, transducer driving power, and window glass geometric parameters into the sound pressure distribution prediction model, output the sound pressure distribution field and vibration displacement distribution on the window glass surface, determine the transducer installation position parameters based on the sound pressure distribution field and vibration displacement distribution on the window glass surface, and guide the installation of the ultrasonic transducer.

[0039] S03: The square wave signal generator generates a square wave signal. The rising edge triggers the ultrasonic signal generator to drive the ultrasonic transducer to apply high-frequency vibration to the viewing window glass to perform self-cleaning. The falling edge triggers the signal switching module to switch to the real-time detection equipment working state.

[0040] S04. After the ultrasonic transducer completes a single cleaning, the transducer impedance change time sequence is collected. The dynamic time warping cleaning process identification algorithm is used to match the transducer impedance change time sequence with the cleaning success time sequence template library. The cleaning completion judgment signal, the current cleaning stage and the predicted remaining cleaning time are output. The cleaning process is dynamically terminated or continued according to the cleaning completion judgment signal.

[0041] S05. After cleaning, the vibration attenuation real-time monitoring module collects the residual vibration amplitude of the window glass through the acceleration sensor, compares the residual vibration amplitude with the attenuation threshold, and outputs the trigger delay time when the residual vibration amplitude is lower than the attenuation threshold. The signal switching module starts the real-time detection device according to the trigger delay time.

[0042] S06: The real-time detection device acquires window images, and the lightweight window contamination assessment model assesses the degree of contamination in the window images and outputs a contamination score. The drive power adaptive adjustment function receives the contamination score, transducer impedance offset, and vibration decay time, and outputs a comprehensive adjustment index. The transducer drive power is adjusted according to the comprehensive adjustment index. When the contamination score is higher than the contamination trigger threshold, S03 is re-triggered. When the contamination score is lower than the contamination trigger threshold, the real-time detection device remains in operation.

[0043] The acoustic impedance gradient matching intermediate layer is an intermediate medium layer filled between the ultrasonic transducer and the window glass contact interface. Its acoustic impedance value is between the acoustic impedance of the ultrasonic transducer material and the acoustic impedance of the window glass material, forming a gradient structure in which the acoustic impedance gradually transitions from the ultrasonic transducer side to the window glass side. This is used to reduce the reflection loss of acoustic energy at the interface and improve the transmission efficiency of acoustic energy to the window glass.

[0044] The specific steps of the algorithm for solving the optimal coupling layer thickness by acoustic impedance gradient descent are as follows: taking the acoustic energy transmission coefficient as the objective function, the acoustic impedance and thickness of the acoustic impedance gradient matching intermediate layer as optimization variables, and using the transfer matrix method to establish a set of acoustic transmission equations for the three media layers of the ultrasonic transducer, acoustic impedance gradient matching intermediate layer, and window glass. The acoustic transmission equations are expressed as follows: ;in The normalized sound energy transmission coefficient. For the first The ratio of the acoustic impedance of the layered medium to the reference acoustic impedance. For the first The product of the wavenumber and thickness of the medium layer is divided by the dimensionless reference wavelength; the acoustic impedance and thickness optimization variables are iteratively updated using the gradient descent method to maximize the vibration displacement amplitude of the window glass surface; the iteration terminates when the change in vibration displacement amplitude between two adjacent iterations is less than a set threshold, and finally the optimal acoustic impedance range and optimal thickness value are output.

[0045] The proposed algorithm for finding the optimal coupling layer thickness using acoustic impedance gradient descent offers the following technical advantages: Traditional solid-solid interface acoustic energy transfer relies on a single contact pressure, and abrupt changes in acoustic impedance at the interface lead to significant acoustic energy reflection, resulting in low actual transmission efficiency. By introducing this algorithm, an acoustic impedance gradient transition structure is constructed between the ultrasonic transducer and the viewing window glass. The transfer matrix method is used to accurately model the acoustic transmission behavior of the multi-layered media. Iterative optimization using the gradient descent method ensures optimal matching between the acoustic impedance gradient, the intermediate layer parameters, and the acoustic impedance of the materials on both sides. This maximizes the transmission of vibration energy to the viewing window glass surface, resulting in a more uniform distribution of vibration energy and improved removal of particles and crystal deposits. Simultaneously, it avoids the problem of insufficient matching caused by relying on empirical material selection, thus improving the device's cleaning adaptability under different media conditions.

[0046] The specific structure of the sound pressure distribution prediction model is as follows: based on a physical information neural network architecture, the input layer receives the optimal acoustic impedance range and optimal thickness value, transducer driving frequency, transducer driving power and window glass geometric parameters, the hidden layer adopts a fully connected structure, and the output layer outputs the sound pressure value and vibration displacement value at each position on the window glass surface; the loss function embeds the acoustic wave equation and boundary condition constraints, and the network output satisfies the physical constraints through automatic differential iteration.

[0047] The steps for establishing the training dataset of the sound pressure distribution prediction model specifically include: performing finite element simulations on different combinations of transducer driving frequencies, transducer driving power, and window glass geometric parameters to obtain simulation results of the sound pressure distribution field and vibration displacement distribution on the window glass surface; arranging accelerometers on the physical device to collect measured sound pressure and vibration displacement data; and fusing the simulation results with the measured data to form the training dataset.

[0048] The specific steps for training the sound pressure distribution prediction model include: dividing the training dataset into a training set and a validation set in an 8:2 ratio; using the weighted sum of mean squared error loss and physical residual loss as the total loss function, and iteratively training with the Adam optimizer; terminating training when the validation set loss no longer decreases, and saving the optimal parameters.

[0049] The sound pressure distribution prediction model brings the following technical benefits to the overall solution: The sound pressure distribution on the window glass surface is jointly affected by the transducer driving frequency, transducer driving power, and window glass geometric parameters. Simply relying on finite element simulation results in long calculation cycles and makes real-time optimization difficult. This model, by embedding acoustic wave equations and boundary condition constraints into the loss function, allows the network to satisfy physical laws while fitting data. This enables rapid and accurate prediction of the sound pressure distribution field and vibration displacement distribution on the window glass surface under limited measured data conditions. It also assists in optimizing transducer installation position parameters, resulting in a more uniform distribution of vibration energy on the window glass surface, eliminating cleaning dead zones, and improving overall cleaning coverage.

[0050] The specific calculation steps for the adaptive adjustment function of the driving power are as follows: Collect pollution scores, transducer impedance offset, and vibration decay time; calculate the comprehensive adjustment index; the formula for the comprehensive adjustment index is as follows: ;in It is a dimensionless comprehensive adjustment index. The ratio of the pollution score to the pollution score reference value. This is the ratio of the transducer impedance offset to the transducer impedance reference value. This is the ratio of the vibration decay time to the reference value for the vibration decay time. , , The weighting coefficients and ;when When the transducer drive power is maintained at the current set value, when When the transducer drive power is increased by 15% based on the current set value; when When the transducer drive power is increased by 30% based on the current set value; when At this time, the transducer drive power is increased to the rated upper limit and an extended cleaning process is triggered.

[0051] The specific steps of the dynamic time warping cleaning process identification algorithm are as follows: A successful cleaning time sequence template library covering different types of contamination is pre-established. Each template in the library represents a time sequence showing the change in transducer impedance over time during a successful cleaning process. For each real-time cleaning process, the transducer impedance change time sequence is collected. The dynamic time warping distance is calculated using a dynamic time warping distance metric method, comparing the transducer impedance change time sequence with each template in the successful cleaning time sequence template library. The template with the smallest dynamic time warping distance is selected as the matching template for the current cleaning process. The formula for the dynamic time warping distance is as follows: ;in To normalize the dynamic time-warped distance, For the timing of transducer impedance changes, To match the template timing, and These represent the normalized impedance values ​​in the transducer impedance change timing sequence and the matching template timing sequence, respectively. To align the path, The normalized path length reference value is used; the current cleaning stage is determined based on the alignment position of the transducer impedance change timing in the matching template, and the predicted remaining cleaning time is output; when the normalized dynamic time warping distance between the transducer impedance change timing and the final state of the matching template is less than the final state neighborhood determination threshold, a cleaning completion determination signal is output.

[0052] The dynamic time-warping cleaning process identification algorithm brings the following technical benefits to the overall solution: Fixed-duration cleaning control cannot adapt to the differences in cleaning rates caused by varying degrees of contamination, media viscosity, and temperature, resulting in insufficient cleaning or excessively long cleaning times leading to prolonged detection interruptions. By introducing this algorithm, and dynamically warping the transducer impedance change timing sequence with a cleaning success timing template library, the algorithm can accurately identify the current cleaning stage even with flexible timing patterns. This enables adaptive determination of the cleaning completion status, ensuring thorough cleaning while minimizing unnecessary cleaning time, thereby improving the effective duty cycle of the real-time detection equipment.

[0053] The vibration attenuation real-time monitoring module is an acceleration sensor signal acquisition and processing unit integrated into the device circuit. It is used to continuously acquire the residual vibration amplitude of the window glass after the ultrasonic transducer stops excitation, compare the residual vibration amplitude with the attenuation threshold, and output the trigger delay time. The signal switching module dynamically determines the start time of the real-time detection device based on the trigger delay time.

[0054] The pollution trigger threshold is a pollution score boundary value preset according to the minimum requirements of the real-time detection equipment for the clarity of the window glass. When the pollution score output by the lightweight window pollution assessment model exceeds the pollution trigger threshold, it is determined that the degree of pollution of the window glass has affected the detection quality of the real-time detection equipment, and S03 is triggered.

[0055] The attenuation threshold is a preset upper limit for the residual vibration amplitude of the window glass. When the residual vibration amplitude is lower than the attenuation threshold, it is determined that the residual vibration has reduced the signal interference of the real-time detection device to an allowable range, and the signal switching module is allowed to start the real-time detection device.

[0056] The lightweight window contamination assessment model is a lightweight image segmentation model obtained by compressing a large image segmentation model using a knowledge distillation method. It combines INT8 quantization and TensorRT deployment to perform real-time contamination segmentation and scoring of window images, and outputs a normalized contamination score.

[0057] The final state neighborhood determination threshold is a preset upper limit value of the normalized dynamic time warping distance. When the normalized dynamic time warping distance between the transducer impedance change timing sequence and the final state segment of the matching template is less than the final state neighborhood determination threshold, it is determined that the cleaning process has reached the final state corresponding to the template in the timing template library.

[0058] The transducer impedance offset is the difference between the real-time acquired transducer impedance value and the transducer impedance reference value, reflecting the change in the load state of the ultrasonic transducer under the current operating conditions.

[0059] The vibration decay time is the time elapsed from when the ultrasonic transducer stops being excited until the residual vibration amplitude of the viewing window glass decreases below the decay threshold, and is output in real time by the vibration decay real-time monitoring module.

[0060] like Figure 2 As shown, a second aspect of this invention provides a real-time online immersion probe window self-cleaning device, comprising a probe housing, a sealing gasket, a window glass, an ultrasonic transducer, a fixing metal ring, an ultrasonic signal generator, a square wave signal generator, a signal switching module, and other components. The window glass is located at the open end of the probe housing, with a sealing gasket between the housing and the window glass. The ultrasonic transducer is installed at the rear end of the window glass, and its structure is annular with an outer plastic or rubber protective sleeve. A fixing metal ring is installed behind the ultrasonic transducer. During operation, the ultrasonic signal generator provides a signal, and the ultrasonic transducer vibrates at high frequency, thereby cleaning small particles and other impurities adhering to the window, thus achieving the purpose of cleaning the window and improving clarity. The square wave generator and the signal switching module are used for real-time switching between testing equipment and ultrasonic cleaning, thereby achieving the effect of real-time online detection.

[0061] Furthermore, the probe housing needs to meet IP68 or higher waterproof requirements.

[0062] Furthermore, the materials used for viewing windows include tempered glass, borosilicate glass, ceramic glass, and sapphire glass.

[0063] Furthermore, the materials for the sealing gaskets include: silicone gaskets, perfluoroether, polytetrafluoroethylene, etc.

[0064] Furthermore, the ultrasonic transducer has a ring-shaped structure with an outer plastic protective layer such as PTFE or rubber. It can generate mechanical vibrations from 1MHz to 100MHz. When this vibration is applied to the viewing window, it can clean the surface of the viewing window, removing small particles, crystals, and other adhering substances, thereby making the viewing window clearer.

[0065] Furthermore, the fixing metal ring secures the ultrasonic transducer and viewing window glass within the probe housing, ensuring their sealing and waterproofing. Another function of the fixing metal ring is to ensure tight contact between the ultrasonic transducer and the viewing window glass, guaranteeing the transmission of mechanical vibrations.

[0066] Furthermore, the ultrasonic signal generator produces high-frequency electronic oscillations: the high-frequency electrical signal generated by the main circuit board is amplified and transmitted to the ultrasonic transducer.

[0067] Furthermore, the square wave signal generator produces a square wave signal, where the rising signal serves as the trigger signal for the ultrasonic generator for automatic cleaning; and the falling signal serves as the trigger signal for the real-time detection equipment. This real-time detection equipment includes, but is not limited to, cameras, infrared, near-infrared, ultraviolet, Raman, and turbidity measurement devices.

[0068] The specific implementation of step S01 is as follows: An acoustic impedance gradient matching intermediate layer is filled between the contact interface between the ultrasonic transducer and the viewing window glass. The acoustic impedance value of the intermediate layer material is between that of the ultrasonic transducer material (typically piezoelectric ceramic, with an acoustic impedance of approximately...). ) and window glass material (acoustic impedance approximately Between the layers, an impedance gradient transition structure is formed. A fixed metal pressure ring is pre-tightened via threads to apply a uniform contact force, with a reference pre-tightening force range of 20–50 N, ensuring consistency of the interface acoustic coupling conditions. The optimal coupling layer thickness solution algorithm for acoustic impedance gradient descent uses the acoustic energy transmission coefficient as the objective function and the acoustic impedance value and thickness of the intermediate layer for acoustic impedance gradient matching as optimization variables. It uses the transfer matrix method to establish a set of acoustic transmission equations for the three media layers (ultrasonic transducer layer, intermediate layer, and window glass layer), describing the continuity conditions of sound pressure and particle velocity at each layer interface. The acoustic impedance and thickness are iteratively updated using the gradient descent method, with a learning rate reference value of [value missing]. The change in vibration displacement amplitude between two adjacent iterations is less than The iteration terminates at time m, outputting the optimal acoustic impedance range and optimal thickness value. The typical optimal thickness value is about one-quarter of the wavelength corresponding to the transducer's operating frequency.

[0069] The specific implementation of step S02 is as follows: The sound pressure distribution prediction model is based on a physical information neural network architecture. The hidden layer adopts a 5-layer fully connected structure, with 128 neurons in each layer. The activation function is the hyperbolic tangent function. When establishing the training dataset, finite element simulations are performed on combinations of transducer driving frequency (20–60 kHz), driving power (10–100 W), and window glass geometric parameters (diameter 20–80 mm, thickness 2–10 mm), generating no less than 2000 sets of simulation samples. Simultaneously, no less than 9 accelerometers are deployed on the physical device to collect measured data, which is then fused with the simulation results to form the training dataset. During training, the training set and validation set are divided in an 8:2 ratio. The total loss function is the weighted sum of the mean square error loss and the physical residual loss (sound wave equation residual and boundary condition residual). The physical residual weight reference value is 0.1. The Adam optimizer is used, and the initial learning rate is... Training terminates when the loss on the validation set no longer decreases after 50 consecutive rounds. The model outputs the sound pressure and vibration displacement values ​​at various locations on the viewing window glass surface, thereby determining the transducer installation position parameters with the largest and most uniform vibration displacement amplitude, guiding the actual installation of the ultrasonic transducer.

[0070] The specific implementation of step S03 is as follows: A square wave signal generator generates a square wave signal with an adjustable duty cycle. The reference value for the square wave frequency is 0.1–1 Hz, and the reference value for the ratio of cleaning time to detection time is 1:4. When the rising edge of the square wave arrives, the signal switching module switches the system to the ultrasonic cleaning state. The ultrasonic signal generator generates a high-frequency sinusoidal signal that matches the resonant frequency of the transducer. After being amplified by a power amplifier, the signal drives the ultrasonic transducer, causing the viewing window glass to vibrate at a high frequency. The reference value for the vibration frequency is 40 kHz, and the reference value for the initial driving power is 50 W. The particulate matter and crystal precipitates adhering to the surface of the viewing window glass are removed using acoustic radiation force and micro-jet effect. When the falling edge of the square wave arrives, the signal switching module automatically switches the system to the real-time detection equipment working state, realizing time-sharing control of the cleaning process and the detection process, and avoiding interference of ultrasonic vibration with the detection signal.

[0071] The specific implementation of step S04 is as follows: During the ultrasonic transducer cleaning process, the change sequence of the real part and mode of the transducer impedance is continuously acquired at a sampling frequency of 1 kHz to form the transducer impedance change time sequence. The pre-established successful cleaning time sequence template library covers different contamination types such as particulate contamination, crystal precipitation contamination, and oil film contamination. Each type of contamination contains at least 10 impedance time sequence templates for successful cleaning processes. The dynamic time warping algorithm constructs an optimal nonlinear alignment path, calculates the normalized dynamic time warping distance between the transducer impedance change time sequence and each template, and selects the template with the smallest distance as the matching template for the current cleaning process. Based on the alignment position of the transducer impedance change time sequence in the matching template, the current cleaning stage (initial stripping stage, main cleaning stage, or final stage) and the predicted remaining cleaning time are output. When the normalized dynamic time warping distance between the real-time time sequence and the final state segment of the matching template is less than the final state neighborhood judgment threshold (reference value is 0.05), a cleaning completion judgment signal is output, and the cleaning process is dynamically terminated; otherwise, cleaning continues and matching continues until the termination condition is met.

[0072] The specific implementation of step S05 is as follows: The vibration attenuation real-time monitoring module continuously collects the residual vibration acceleration signal of the window glass at a sampling frequency of 10 kHz after the ultrasonic transducer stops excitation, using an accelerometer arranged at the edge of the window glass. After bandpass filtering (the filtering frequency band is centered on the transducer resonant frequency, and the bandwidth reference value is ±10% of the transducer resonant frequency), the root mean square value of the residual vibration amplitude is calculated. The root mean square value is compared with the attenuation threshold (reference value is 1% of the initial vibration amplitude) in real time. When the residual vibration amplitude is lower than the attenuation threshold, the time difference between the current time and the time when the transducer stops excitation is recorded as the vibration attenuation time, and a trigger delay time (reference value is 1.2 times the vibration attenuation time) is output to the signal switching module. The signal switching module starts the real-time detection device after the trigger delay time is reached to ensure that the detection process is not disturbed by residual vibration.

[0073] The specific implementation of step S06 is as follows: After the real-time detection device acquires the window image, the lightweight window contamination assessment model performs real-time semantic segmentation on the image, dividing the effective area of ​​the window glass into clean and contaminated areas. The contamination score is defined as the ratio of the contaminated area area to the total effective area area, and the output is a normalized contamination score ranging from 0 to 1, with a contamination trigger threshold reference value of 0.15. The drive power adaptive adjustment function simultaneously receives the contamination score, transducer impedance offset (the difference between the real-time impedance value and the reference impedance value, where the reference value is the transducer resonant impedance), and vibration decay time, normalizes them according to their respective reference values, and then uses a weighted coefficient (reference value) to adjust the output. , , The weighted summation yields the comprehensive adjustment index, and the drive power adjustment command is output according to the segmented intervals of the comprehensive adjustment index. When the pollution score is higher than the pollution trigger threshold, the system re-triggers step S03 to perform cleaning; when the pollution score is lower than the pollution trigger threshold, the system maintains the real-time detection equipment's working status, continuously collects window images and cyclically evaluates the degree of pollution, forming an adaptive closed-loop control for cleaning and detection.

[0074] It should be noted that the first key technical idea of ​​this invention is to solve for the intermediate layer parameters based on the transfer matrix method and gradient descent algorithm for acoustic impedance gradient matching. Traditional solid-solid interface ultrasonic coupling relies on contact pressure to maintain acoustic energy transfer. The abrupt change in acoustic impedance between the materials on both sides of the interface causes a large amount of acoustic energy to be lost as reflected waves, resulting in limited effective vibration energy reaching the viewing window glass. This invention constructs a gradient transition structure at the interface and accurately models the acoustic transmission behavior of the multilayer medium using the transfer matrix method. It then uses the gradient descent method to optimize the intermediate layer parameters, achieving optimal acoustic impedance matching between the intermediate layer parameters and the materials on both sides. This maximizes the acoustic energy transmission coefficient and improves the actual transmission efficiency of vibration energy.

[0075] The second key technological approach is the prediction of sound pressure distribution and optimization of transducer installation location driven by a physical information neural network. Relying solely on finite element simulation results in long calculation cycles and cannot support real-time parameter optimization, while purely data-driven models lack generalization ability under small sample conditions. The physical information neural network embeds the acoustic wave equation and boundary condition constraints into the loss function, enabling the network to satisfy physical laws while fitting measured data. This allows for rapid and accurate prediction of the sound pressure distribution field and vibration displacement distribution on the window glass surface under limited sample conditions, assisting in the quantitative optimization of transducer installation location parameters and eliminating cleaning dead zones.

[0076] The third key technical approach is the adaptive determination of the cleaning process driven by the dynamic time warping algorithm. Fixed-duration control cannot adapt to the differences in cleaning rates under different working conditions. The dynamic time warping algorithm overcomes the interference of cleaning rate differences on similarity measurement by allowing the time sequence to be non-linearly scaled and matched on the time axis, thus realizing the adaptive determination of the working condition of the cleaning completion status.

[0077] The synergistic effect of the three technical approaches is as follows: the acoustic impedance gradient matching intermediate layer ensures that the vibration energy of sufficient intensity is transmitted to the viewing window glass, which is the physical basis for the cleaning effectiveness; the sound pressure distribution prediction model ensures that the vibration energy is evenly distributed on the surface of the viewing window glass, eliminating local cleaning blind spots; and the dynamic time warping algorithm, on the basis of ensuring cleaning quality by the first two approaches, further realizes the dynamic adaptation of the cleaning time, so that the entire system can minimize the detection interruption time while ensuring the thoroughness of cleaning. Together, the three constitute a complete technical chain from energy transmission and distribution optimization to process control.

[0078] It should be noted that this invention also solves the following technical problem: In the prior art, acoustic energy transfer between the ultrasonic transducer and the viewing window glass is achieved through single-material contact or simple filling with a coupling agent. This lack of quantitative optimization methods for interface acoustic impedance matching parameters results in acoustic energy transmission efficiency being limited by the inherent differences in material acoustic impedance, making it impossible to flexibly adjust coupling parameters for different combinations of transducer and glass materials. This invention, through an optimal coupling layer thickness solution algorithm based on acoustic impedance gradient descent, establishes an accurate multi-layer acoustic transmission model using the transfer matrix method and optimizes it using the gradient descent method. This achieves quantitative optimization of coupling layer parameters, solving the technical problem of insufficient matching caused by reliance on empirical material selection for acoustic energy transmission efficiency.

[0079] Furthermore, in existing technologies, after the ultrasonic cleaning device stops excitation, the viewing window glass retains residual vibration for a certain period. If the real-time detection device is activated before the residual vibration has sufficiently decayed, the residual vibration will interfere with the optical or acoustic detection signals, affecting the accuracy of the detection results. Conversely, using a fixed delay time to wait for vibration decay would cause unnecessary waiting under conditions where vibration decay is rapid, further compressing the effective detection time. This invention continuously collects the residual vibration amplitude using a real-time vibration decay monitoring module and compares it with the decay threshold in real time, dynamically outputting the trigger delay time. This ensures that the activation time of the detection device precisely corresponds to the moment when the residual vibration decays to an acceptable range, solving the technical problem that a fixed delay time cannot adapt to differences in residual vibration decay rates under different operating conditions.

[0080] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned core technical problems is that during the process of vibrating and cleaning the viewing window glass, the impedance characteristics of the ultrasonic transducer dynamically evolve with the change of load state. The peeling state of contaminants, the medium coupling conditions, and the cleanliness of the glass surface during the cleaning process all leave identifiable characteristic trajectories on the transducer impedance time sequence. The dynamic time warping algorithm, by constructing a nonlinear alignment path, allows the time sequences to be compared to be elastically stretched and matched on the time axis, thereby overcoming the problem of time sequence morphology changes caused by differences in cleaning rates under different working conditions. This makes the similarity measurement between the impedance change time sequence and the cleaning success time sequence template library adaptable to different working conditions. When the normalized dynamic time warping distance between the real-time impedance time sequence and the template final state segment is less than the final state neighborhood judgment threshold, the system determines that the cleaning process has reached a successful final state and outputs a completion judgment signal, thereby achieving dynamic termination of the cleaning duration and replacing the fixed duration control strategy. Logically, the above mechanism constitutes a complete closed loop from process perception to state judgment to control output, which conforms to the basic principles of adaptive control systems. Therefore, it can fundamentally solve the technical problem of the inability to adaptively determine the cleaning completion state.

[0081] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0082] The specific implementation of step S01 is as follows: An acoustic impedance gradient matching intermediate layer is filled into the contact interface between the ultrasonic transducer and the viewing window glass. A pre-tightening force is applied by a fixed metal pressure ring to ensure uniform contact at the interface. The algorithm for solving the optimal coupling layer thickness by reducing the acoustic impedance gradient uses the acoustic energy transmission coefficient as the objective function, and the acoustic impedance and thickness of the acoustic impedance gradient matching intermediate layer as optimization variables. The transfer matrix method is used to establish a set of acoustic transmission equations for the three media layers: the ultrasonic transducer, the acoustic impedance gradient matching intermediate layer, and the viewing window glass. The normalized acoustic energy transmission coefficient formula is expressed as follows:

[0083] ;

[0084] In the formula, is the normalized sound energy transmission coefficient, with a dimension of 1; is the actual sound energy transmission coefficient, with a dimension of 1; The reference sound energy transmission coefficient has a dimensionless value of 1 and a value of 1. For the first Acoustic impedance of the layered medium, in units of ; For the first Acoustic impedance of the layered medium, in units of ; For reference acoustic impedance, the acoustic impedance value of the ultrasonic transducer material is taken, in units of . ; For the first The thickness of the layer medium, in units of ; It is the imaginary unit, with a dimension of 1; The total number of dielectric layers, dimensionless (1). For the first Wave number in a layered medium, in units of Calculated by the following formula:

[0085] ;

[0086] For the first Reference wavelength in layered medium, in units of Calculated by the following formula:

[0087] ;

[0088] In the formula, The transducer drive frequency, in units of ; For the first The speed of sound in a layered medium, in units of This is determined by the properties of the medium material. Therefore, it can be known that... , is a dimensionless quantity with dimension 1, consistent with the requirement for the independent variable of a trigonometric function. Iterative updates are performed using the gradient descent method. and To maximize the vibration displacement amplitude of the viewing window glass surface, the iteration terminates when the change in vibration displacement amplitude between two consecutive iterations is less than a set threshold, which is empirically considered to be... The final output is the optimal acoustic impedance range and the optimal thickness value.

[0089] The specific implementation of step S02 is as follows: The optimal acoustic impedance range, optimal thickness value, transducer driving frequency, transducer driving power, and window glass geometric parameters are input into the sound pressure distribution prediction model. This model is based on a physical information neural network architecture, with the acoustic wave equation and boundary condition constraints embedded in the loss function. Through automatic differential iteration, the network output satisfies the physical constraints, outputting the sound pressure distribution field and vibration displacement distribution on the window glass surface. The training dataset is constructed by performing finite element simulations on different combinations of transducer driving frequencies, transducer driving powers, and window glass geometric parameters, and by collecting measured data from accelerometers on a physical device. The training set and validation set are divided in an 8:2 ratio. The total loss function is the weighted sum of mean square error loss and physical residual loss. The Adam optimizer is used for iterative training. Training is terminated and the optimal parameters are saved when the validation set loss no longer decreases. The transducer installation position parameters are determined based on the sound pressure distribution field and vibration displacement distribution output by the model to guide the installation of the ultrasonic transducer, ensuring uniform distribution of vibration energy on the window glass surface and eliminating cleaning dead zones.

[0090] The specific implementation of step S03 is as follows: the square wave signal generator generates a square wave signal, the rising edge triggers the ultrasonic signal generator to drive the ultrasonic transducer to apply high-frequency vibration to the viewing window glass to perform self-cleaning; the falling edge triggers the signal switching module to switch to the real-time detection equipment working state to realize time-sharing control of the cleaning process and the detection process.

[0091] The specific implementation of step S04 is as follows: After the ultrasonic transducer completes a single cleaning cycle, the transducer impedance change time sequence is collected. A dynamic time warping cleaning process identification algorithm is used to match the transducer impedance change time sequence with a successful cleaning time sequence template library. The successful cleaning time sequence template library is pre-established, covering different types of contamination. Each template in the library represents the time sequence of transducer impedance changes over time during a successful cleaning process. The normalized dynamic time warping distance formula is expressed as follows:

[0092] ;

[0093] In the formula, The normalized dynamic time-warped distance has a dimension of 1. This is the real-time acquisition timing of transducer impedance changes; To match the template timing; For time sequence The Middle The transducer impedance value at a given moment, in units of ; For time sequence The Middle The template impedance value at each moment, in units of ; This is the normalized reference value for the transducer impedance, in units of... The nominal impedance of the transducer under rated operating conditions is obtained through experimental calibration. The specific steps are as follows: Step 1, place the probe in a clean medium and drive the transducer to operate at the rated frequency; Step 2, measure the transducer port impedance using an impedance analyzer and record the stabilized impedance value as the nominal impedance. ; For time sequence With time series Alignment paths between; The normalized path length reference value is taken as the maximum time series length, with a dimension of 1. The template with the smallest normalized dynamic time warp distance is selected as the matching template for the current cleaning process, based on the time series. Determine the current cleaning stage based on the alignment position in the matching template and output the predicted remaining cleaning time. When the value is less than the final state neighborhood determination threshold, a cleaning completion determination signal is output. The cleaning process is dynamically terminated or continued based on the cleaning completion determination signal. The empirical value of the final state neighborhood determination threshold is 0.05.

[0094] The specific implementation of step S05 is as follows: After cleaning, the vibration attenuation real-time monitoring module continuously collects the residual vibration amplitude of the window glass through an accelerometer. It compares the residual vibration amplitude with an attenuation threshold. When the residual vibration amplitude is lower than the attenuation threshold, it outputs a trigger delay time. The signal switching module then activates the real-time detection device based on the trigger delay time. The attenuation threshold is a pre-set upper limit for the residual vibration amplitude of the window glass. When the residual vibration amplitude is lower than this threshold, it is determined that the residual vibration's interference with the real-time detection device's signal has been reduced to an acceptable range. Vibration attenuation time. The time elapsed from when the ultrasonic transducer stops excitation until the residual vibration amplitude decreases below the attenuation threshold is output in real time by the vibration attenuation real-time monitoring module, in units of 1. .

[0095] The specific implementation of step S06 is as follows: A real-time detection device acquires window images. A lightweight window contamination assessment model evaluates the contamination level of the window images and outputs a normalized contamination score. This model is obtained by compressing a large image segmentation model using a knowledge distillation method, combined with INT8 quantization and TensorRT deployment, supporting real-time inference. The drive power adaptive adjustment function receives the contamination score, transducer impedance offset, and vibration decay time, and calculates the comprehensive adjustment index. The formula for the comprehensive adjustment index is as follows:

[0096] ;

[0097] In the formula, It is a dimensionless comprehensive regulation index with a dimension of 1; The normalized pollution score output by the lightweight window pollution assessment model has a dimension of 1. This is a reference value for pollution scoring, with a dimension of 1, and the default value is 1. Transducer impedance offset, in units of Calculated by the following formula:

[0098] ;

[0099] In the formula, The transducer impedance measurement value is acquired in real time, and the unit is... ; The nominal impedance of the transducer under clean operating conditions, in units of... The impedance was obtained through experimental calibration. The specific steps are as follows: Step 1, place the probe in a clean medium and drive the transducer to operate at the rated frequency; Step 2, use an impedance analyzer to measure the transducer port impedance and record the stabilized impedance value as the measured value. ; This is a reference value for vibration decay time, in units of Take the typical decay time under normal operating conditions; , , For the weighting coefficients, satisfying The empirical values ​​are 0.5, 0.3, and 0.2 respectively. When When the transducer drive power is maintained at the current set value, when When the transducer drive power is increased by 15% based on the current set value; when When the transducer drive power is increased by 30% based on the current set value; when At this time, the transducer drive power is increased to the rated upper limit and the extended cleaning process is triggered. When the contamination score is higher than the contamination trigger threshold, step S03 is retried. When the contamination score is lower than the contamination trigger threshold, the real-time detection equipment remains in operation. The contamination trigger threshold is a contamination score boundary value preset according to the minimum requirements of the real-time detection equipment for the clarity of the viewing window glass.

[0100] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0101] To verify the effectiveness of the invention, technicians set up a test environment and tested the mixed medium containing crystalline precipitates and particulate suspensions in a chemical reactor by installing the device of the invention on an industrial online Raman spectroscopy probe. The probe window glass had a diameter of 40 mm and a thickness of 5 mm, and the ultrasonic transducer resonant frequency was 40 kHz.

[0102] like Figure 2 As shown, the automatic cleaning device in this embodiment includes components such as a probe housing 1, a sealing gasket 2, a viewing window 3, an ultrasonic transducer 4, a fixed metal pressure ring 5, an ultrasonic signal generator 6, a square wave signal generator 7, a signal switching module 8, and a real-time detection device 9.

[0103] The viewing window 3 is installed at the open end of the probe housing 1 via the sealing gasket 2. The other end of the viewing window 3 is connected to the ultrasonic transducer 4 and secured with a fixing metal ring 5. The fixing metal ring 5 ensures a tight fit between the sealing gasket 2 and the probe housing 1; the tight fit between the viewing window 3 and the sealing gasket 2 ensures the airtightness of the probe. The fixing metal ring 5 also ensures a tight fit between the viewing window 3 and the ultrasonic transducer 4, ensuring that the ultrasonic transducer efficiently transmits high-frequency vibrations to the surface of the viewing window.

[0104] During self-cleaning, the ultrasonic signal generator provides signal 6, and the ultrasonic transducer 4 vibrates at high frequency, causing the viewing window glass to also vibrate at high frequency. This causes small particles and other impurities adhering to the viewing window to detach, thereby achieving the purpose of cleaning the viewing window and improving clarity. When the real-time detection device 9 is working, the signal switching module 8 performs time-division control of cleaning and testing based on the signal from the square wave signal generator 7.

[0105] In step S01, the acoustic impedance gradient matching intermediate layer is made of epoxy resin-based composite material, and its acoustic impedance is approximately Between piezoelectric ceramic transducers (acoustic impedance approximately...) ) and window quartz glass (acoustic impedance approximately The optimal coupling layer thickness solution algorithm for acoustic impedance gradient descent establishes a three-layer acoustic transmission equation system using the transfer matrix method, and then iterates through gradient descent (learning rate...). After convergence (with 2000 iterations), the optimal thickness value is 18 μm, at which point the normalized acoustic transmission coefficient reaches 0.94. A preload of 35 N is applied to the fixed metal pressure ring to ensure uniform contact at the interface.

[0106] In step S02, the sound pressure distribution prediction model is based on a physical information neural network. The training dataset contains 1800 finite element simulation samples and 120 measured samples. The simulation covers a combination of transducer driving frequencies of 20–60 kHz, driving power of 10–100 W, and window glass geometric parameters. The training and validation sets are divided in an 8:2 ratio, with the physical residual loss weight set to 0.1. The Adam optimizer is used for training, and the validation set loss converges and training stops at the 320th round. The model predicts the vibration displacement distribution on the window glass surface, and based on this, the transducer installation eccentricity is determined to be 3 mm to ensure that the vibration displacement uniformly covers the effective area of ​​the window. Figure 3 As shown, the predicted results of vibration displacement distribution on the window glass surface show that the uniformity is significantly better than that of the center-mounted installation scheme.

[0107] In step S03, the square wave signal generator is set to a square wave frequency of 0.2 Hz and a duty cycle of 20%, meaning that the cleaning time is 1 s and the detection time is 4 s in each cycle. The initial transducer drive power is set to 50 W. The ultrasonic signal generator outputs a 40 kHz sine wave signal to drive the transducer to apply high-frequency vibration to the viewing window glass.

[0108] In step S04, the successful cleaning timing template library pre-recorded 15 transducer impedance change timing sequences for each of the three types of successful cleaning processes: crystal precipitation contamination, particulate suspension contamination, and mixed contamination. The impedance sampling frequency was 1 kHz. The dynamic time warping algorithm calculated the normalized dynamic time warping distance between the real-time impedance timing sequence and each template in the template library. When a mixed contamination template was matched, the alignment position indicated that the main cleaning stage was in progress, and the predicted remaining cleaning time was 2.3 s. When the normalized dynamic time warping distance between the real-time timing sequence and the final state segment of the matched template dropped to 0.038 (below the final state neighborhood judgment threshold of 0.05), the system output a cleaning completion judgment signal, dynamically terminating the current cleaning process. The actual cleaning time was 3.1 s, saving unnecessary cleaning time compared to fixed duration control (preset 5 s). The dynamic time warping matching results for each contamination type are shown in Table 1.

[0109] Table 1. Dynamic Time Warping Matching Results of Cleaning Processes for Each Type of Contamination

[0110]

[0111] In step S05, the vibration attenuation real-time monitoring module acquires the accelerometer signal at a sampling frequency of 10 kHz, calculates the root mean square value of the residual vibration after bandpass filtering, and the initial value is... m / The attenuation threshold is set to 1% of the initial value. m / The time for residual vibration to drop below the attenuation threshold is 0.35 s. The vibration attenuation real-time monitoring module outputs a trigger delay time of 0.42 s (1.2 times the attenuation time). The signal switching module starts the real-time detection device after the trigger delay time is reached.

[0112] In step S06, after the real-time detection device acquires the window image, the lightweight window contamination assessment model (knowledge distillation from a large segmentation model, INT8 quantization, TensorRT deployment, inference time approximately 8 ms) outputs a normalized contamination score. The weight coefficients of the adaptive adjustment function for driving power are set as follows: , , The contamination trigger threshold was set to 0.15. After the first cleaning, the contamination score was 0.07, lower than the contamination trigger threshold, and the system maintained real-time monitoring of the equipment. After a period of time, the contamination score rose to 0.21, higher than the contamination trigger threshold. At the same time, the transducer impedance offset was 0.18 of the reference value, the vibration decay time was 0.35 of the reference value, and the comprehensive adjustment index was calculated to be 0.43, falling within the 0.3 to 0.7 range. The drive power was automatically increased by 15% to 57.5 W, and step S03 was retried to perform cleaning. The adaptive adjustment process of drive power under various operating conditions is shown in Table 2.

[0113] Table 2 Record of Drive Power Adaptive Adjustment Process

[0114]

[0115] Therefore, the predicted vibration displacement distribution on the window glass surface verifies the effect of optimizing the transducer installation position. When the transducer is installed at an eccentric position of 3 mm, the uniformity of vibration displacement in the effective area of ​​the window is significantly better than that of a centrally installed transducer, thus eliminating vibration nodes in the central area of ​​the window.

[0116] The technological advancements of this invention compared to traditional fixed-duration cleaning control methods are reflected in the following aspects. First, in traditional schemes, acoustic energy relies on contact pressure for transmission, and abrupt changes in interface acoustic impedance lead to the reflection of a large amount of acoustic energy. This invention, however, constructs an acoustic impedance gradient transition structure and quantitatively optimizes the intermediate layer parameters using the transfer matrix method, transforming the acoustic energy transmission mechanism from passive contact coupling to active impedance matching, thus improving the physical efficiency of vibration energy transmission to the viewing window glass. Second, in traditional schemes, the transducer installation position is determined empirically, resulting in insufficient cleaning energy at vibration nodes on the viewing window surface, forming cleaning dead zones. This invention, however, predicts the sound pressure distribution field using a physical information neural network and quantitatively optimizes the transducer installation position accordingly, ensuring uniform distribution of vibration energy within the effective area of ​​the viewing window, fundamentally eliminating the mechanism for the formation of cleaning dead zones. Third, traditional fixed-duration control cannot perceive the cleaning process status. However, this invention utilizes the cleaning status information carried by the timing of transducer impedance changes and achieves adaptive determination of the cleaning completion status under different pollution conditions through the elastic scaling matching mechanism of dynamic time warping algorithm. This ensures that the cleaning termination time corresponds precisely to the actual cleaning completion status, fundamentally solving the inherent limitations of fixed-duration control strategies.

[0117] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0118] Table 3. Variable Explanation Table

[0119]

[0120] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-cleaning method for an immersion probe window, characterized in that, Includes the following steps: An acoustic impedance gradient matching intermediate layer is filled into the contact interface between the ultrasonic transducer and the viewing window glass. A pre-tightening force is applied by a fixed metal pressure ring to ensure uniform contact at the interface. The acoustic impedance range and thickness value of the acoustic impedance gradient matching intermediate layer are solved based on the optimal coupling layer thickness solution algorithm of acoustic impedance gradient descent, and the optimal acoustic impedance range and optimal thickness value are output. The optimal acoustic impedance range, optimal thickness value, transducer driving frequency, transducer driving power, and window glass geometric parameters are input into the sound pressure distribution prediction model. The output is the sound pressure distribution field and vibration displacement distribution on the window glass surface. Based on the sound pressure distribution field and vibration displacement distribution on the window glass surface, the transducer installation position parameters are determined to guide the installation of the ultrasonic transducer. The square wave signal generator generates a square wave signal. The rising edge triggers the ultrasonic signal generator to drive the ultrasonic transducer to apply high-frequency vibration to the viewing window glass to perform self-cleaning. The falling edge triggers the signal switching module to switch to the real-time detection equipment working state. After the ultrasonic transducer completes a single cleaning, the transducer impedance change time sequence is collected. The dynamic time warping cleaning process identification algorithm is used to match the transducer impedance change time sequence with the cleaning success time sequence template library. The cleaning completion judgment signal, the current cleaning stage and the predicted remaining cleaning time are output. The cleaning process is dynamically terminated or continued according to the cleaning completion judgment signal. After cleaning, the vibration attenuation real-time monitoring module collects the residual vibration amplitude of the window glass through the accelerometer and compares the residual vibration amplitude with the attenuation threshold. When the residual vibration amplitude is lower than the attenuation threshold, the trigger delay time is output. The signal switching module starts the real-time detection device according to the trigger delay time. The real-time detection device acquires window images, and the lightweight window contamination assessment model evaluates the degree of contamination in the window images and outputs a contamination score. The drive power adaptive adjustment function receives the contamination score, transducer impedance offset, and vibration decay time, and outputs a comprehensive adjustment index. The transducer drive power is adjusted according to the comprehensive adjustment index. When the contamination score is higher than the contamination trigger threshold, the self-cleaning step is re-triggered. When the contamination score is lower than the contamination trigger threshold, the real-time detection device remains in operation.

2. The self-cleaning method for the immersion probe window according to claim 1, characterized in that, The acoustic impedance value of the acoustic impedance gradient matching intermediate layer is between the acoustic impedance of the ultrasonic transducer material and the acoustic impedance of the window glass material, forming a gradient structure in which the acoustic impedance gradually transitions from the ultrasonic transducer side to the window glass side, which is used to reduce the reflection loss of acoustic energy at the interface.

3. The self-cleaning method for the immersion probe window according to claim 2, characterized in that, The optimal coupling layer thickness solution algorithm of acoustic impedance gradient descent takes the acoustic energy transmission coefficient as the objective function, and uses the acoustic impedance and thickness of the acoustic impedance gradient matching intermediate layer as optimization variables. It uses the transfer matrix method to establish a set of acoustic transmission equations for the three media of ultrasonic transducer, acoustic impedance gradient matching intermediate layer and window glass, and uses the gradient descent method to iteratively update the optimization variables to maximize the vibration displacement amplitude of the window glass surface.

4. The self-cleaning method for the immersion probe window according to claim 3, characterized in that, The iteration termination condition of the optimal coupling layer thickness solution algorithm by acoustic impedance gradient descent is that the change in vibration displacement amplitude between two adjacent iterations is less than a set threshold, and finally outputs the optimal acoustic impedance range and the optimal thickness value.

5. The self-cleaning method for the immersion probe window according to claim 4, characterized in that, The sound pressure distribution prediction model is based on a physical information neural network. The input layer receives the optimal acoustic impedance range and optimal thickness value, transducer driving frequency, transducer driving power and window glass geometric parameters. The output layer outputs the sound pressure value and vibration displacement value at each position on the window glass surface. The loss function embeds the acoustic wave equation and boundary condition constraints, and the network output satisfies the physical constraints through automatic differential iteration.

6. The self-cleaning method for the immersion probe window according to claim 5, characterized in that, The training dataset for the sound pressure distribution prediction model is obtained by performing finite element simulations on different combinations of transducer driving frequencies, transducer driving power, and window glass geometric parameters. Accelerometers are then deployed on the physical device to collect measured data, and the simulation results are fused with the measured data.

7. The self-cleaning method for the immersion probe window according to claim 6, characterized in that, The training dataset for the sound pressure distribution prediction model is divided into training and validation sets in an 8:2 ratio. The total loss function is the weighted sum of mean squared error loss and physical residual loss. The Adam optimizer is used for iterative training, and training is terminated when the validation set loss no longer decreases.

8. The self-cleaning method for the immersion probe window according to claim 7, characterized in that, The dynamic time warping cleaning process identification algorithm pre-establishes a cleaning success time sequence template library covering different types of contamination. It uses the dynamic time warping distance metric method to calculate the dynamic time warping distance between the transducer impedance change time sequence and each template in the cleaning success time sequence template library, and selects the template with the smallest dynamic time warping distance as the matching template for the current cleaning process.

9. The self-cleaning method for the immersion probe window according to claim 8, characterized in that, The dynamic time warping cleaning process identification algorithm determines the current cleaning stage and outputs the predicted remaining cleaning time based on the alignment position of the transducer impedance change timing in the matching template. When the normalized dynamic time warping distance between the transducer impedance change timing and the final state of the matching template is less than the final state neighborhood judgment threshold, a cleaning completion judgment signal is output.

10. A real-time online immersion probe window self-cleaning device for implementing the method of any one of claims 1 to 9, characterized in that, It includes a probe housing, a sealing gasket, a viewing window, an ultrasonic transducer, a fixing metal ring, an ultrasonic signal generator, a square wave signal generator, a signal switching module, and a real-time detection device; the viewing window is installed at the open end of the probe housing via the sealing gasket, and the other end of the viewing window is connected to the ultrasonic transducer. The fixing metal ring fixes the ultrasonic transducer and the viewing window inside the probe housing.

Images