Piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm

Abstract

The invention belongs to the technical field of multidisciplinary crossing, and discloses a piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm, which comprises the following specific steps: S1, determining key design parameters of a piezoelectric acoustic sensor packaging structure, S2, determining multiple optimization indexes of the piezoelectric acoustic sensor packaging structure, and S3, determining a plurality of optimization parameters of the piezoelectric acoustic sensor packaging structure. S3, establishing a multi-physics field coupling model through COMSOL software; and S4, converting a multi-objective optimization problem of the packaging parameters into a multi-objective decision problem based on deep reinforcement learning, and carrying out interactive learning through a DRL intelligent agent and an optimization environment. The improved piezoelectric acoustic sensor design and optimization method is suitable for application scenes needing high-precision sound wave detection, the core lies in that state space parameter selection is meticulously designed, it is ensured that the sensor keeps the optimal performance under different conditions, and particularly, the optimal performance of the sensor is ensured. The thickness of the thin film serving as a key assembly is set in a reasonable range.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of multi-disciplinary cross technology, and specifically relates to a piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm. BACKGROUND

[0002] The piezoelectric acoustic sensor is a device capable of realizing sound-electric energy conversion, and the core plays a role through the piezoelectric material arranged thereon. Because the piezoelectric material has piezoelectric effect, when external force generated by external sound vibration is applied thereto, polarization charge is generated on the surface of the material, and then the sound signal is converted into an extractable electric signal, thereby completing the sound-electric conversion process. The sensor is widely used in external sound pickup scenes, and is suitable for underwater detection and other special environments, and can also be used in industrial equipment detection fields. For example, in an industrial scene, the piezoelectric acoustic sensor picks up the sound signals emitted by the vibration of the parts of the running equipment, converts the sound signals into electric signals, and then uses the useful information such as the vibration characteristics of the equipment carried by the electric signals to diagnose the mechanical operation state of the equipment, thereby providing data support for equipment maintenance and fault warning. The design of the packaging parameters of the piezoelectric acoustic sensor in the prior art mainly depends on empirical formulas or single requirements (such as lightweight) proposed by customers. This design mode often only focuses on the improvement of a single performance, but ignores the mutual restriction and trade-off relationship between different characteristics associated with the packaging parameters. For example, when the thickness or structure of the shell is optimized to meet the lightweight requirement, the acoustic sensitivity may abnormally fluctuate. If the piezoelectric film thickness or cavity size is adjusted to improve the acoustic sensitivity, the problem of excessive frequency deviation may be caused. The other performance defects caused by the single performance optimization make it difficult for the existing piezoelectric acoustic sensor to achieve a balance in multiple dimensions, and the overall comprehensive performance still has a significant optimization space. SUMMARY

[0003] The application aims to provide a piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm to solve the problems in the background.

[0004] In order to achieve the above-mentioned purpose, the application provides the following technical scheme: a piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm, and the specific steps of the optimization algorithm are as follows: S1: determining the key design parameters of the packaging structure of the piezoelectric acoustic sensor, wherein the key design parameters are parameters that have a direct or indirect influence on the performance of the sensor and have a priority in the degree of influence; S2: determining multiple optimization indexes of the packaging structure of the piezoelectric acoustic sensor, wherein the optimization indexes aim to maximize the comprehensive performance of the sensor, and there is a mutual dependence or restriction relationship between the indexes; S3: A multi-physical field coupling model is established by COMSOL software, and the model is used to simulate the structural characteristics and actual working conditions of the piezoelectric acoustic sensor; S4: The multi-objective optimization problem of the packaging parameters is converted into a multi-objective decision problem based on deep reinforcement learning. Through the interaction learning of the DRL agent and the optimization environment, the optimal solution set of the key design parameters is searched, and the collaborative balance of multiple optimization indicators is realized.

[0005] Preferably, the key design parameters in S1 include the thickness of the thin-film piezoelectric material, the size of the internal cavity of the sensor, the electrode layout method, the thickness and shell material of the shell, and the material and thickness of the acoustic matching layer; wherein the thin-film piezoelectric material is selected from at least one of PVDF film, PZT film, ZnO film, and AlN film, and the shell material is a metal material with electromagnetic shielding and structural protection functions.

[0006] Preferably, the multiple optimization indicators in S2 include acoustic pressure sensitivity, signal-to-noise ratio, temperature drift, frequency shift, and process disturbance; wherein the acoustic pressure sensitivity is defined as the ratio of the output electric signal of the sensor to the input acoustic pressure under unit acoustic pressure, with a unit of mV / Pa or mV / µbar, and the expression is S=output electric signal / input acoustic pressure; the multiple optimization indicators aim to simultaneously improve the acoustic pressure sensitivity and signal-to-noise ratio, and reduce the influence of temperature drift, frequency shift, and process disturbance on the performance of the sensor.

[0007] Preferably, the multi-physical field coupling model in S3 is constructed by adding acoustic, mechanical, electric field, and thermal modules in COMSOL; in the simulation process of the model, the acoustic impedance matching effect of the acoustic matching layer between the piezoelectric material and the measured object, the electromagnetic shielding effect of the shell, and the influence of the cavity on the vibration and stress distribution of the piezoelectric material are considered, and the model also needs to restore the vibration-stress-potential difference conversion process of the piezoelectric material under the action of acoustic signals.

[0008] Preferably, the deep reinforcement learning in S4 is realized by combining deep learning and reinforcement learning; wherein the reinforcement learning optimizes the behavior strategy through the trial-and-error interaction learning of the DRL agent and the optimization environment, aiming to maximize the long-term cumulative reward signal; the deep learning constructs a multi-layer neural network as a value function approximator or a policy approximator, the value function approximator is used to estimate the long-term cumulative reward value corresponding to different packaging parameter combinations, the policy approximator is used to directly generate the adjustment strategy of the packaging parameters, and the neural network can extract features from sensor performance data to realize end-to-end training and complex environment modeling.

[0009] Preferably, the reward function of the DRL in S4 is designed as a multi-objective reward function, the output value of which is directly related to multiple optimization indicators in S2: when the sound pressure sensitivity of the sensor is improved, the signal-to-noise ratio is improved, the temperature drift is reduced, the frequency offset is reduced, and the process disturbance is reduced, the reward function outputs a positive reward; when the performance of the above optimization indicators deteriorates, a negative reward is output; and the reward function realizes the trade-off between different optimization indicators through a weight adjustment mechanism, avoiding the imbalance of the overall performance caused by the optimization of a single indicator.

[0010] wherein the calculation formula of the reward function is: One design of the reward function is as follows: R = ω1(S p S min ) ω2∣Δf f target ∣ ω3A resonance , wherein ω1, ω2, and ω3 are weight coefficients for balancing the priority of the target, which are pre-adjusted through experiments, S p is the simulated sound pressure sensitivity, S min is the minimum sound pressure sensitivity required by the task, Δf is the simulated frequency offset value, f target is the target frequency offset value, and A resonance is the simulated parasitic resonance amplitude combined with the process disturbance simulated by the generative adversarial network to improve the robustness of the algorithm. The generation of the simulation data depends on the pre-established multi-physical field coupling model, and the output strategy is defined as: selecting the action a that maximizes as the optimal parameter adjustment after convergence.

[0011] Preferably, the construction of the optimization environment in S4 needs to take into account the influence of process disturbance: the parasitic resonance amplitude analysis method and the generative adversarial network are used to simulate the process disturbance scenarios, including film thickness deviation, cavity size error, and electrode misplacement. The simulation obtains the parasitic parameter changes and resonance characteristic degradation data of the sensor under different process disturbances, and inputs this data as the environmental state of the DRL agent, so that the optimized packaging parameters have the ability to resist process disturbance.

[0012] Preferably, the training process of the DRL agent in S4 includes: S41: initializing the feasible region of the packaging parameters, which is determined based on the physical constraints of the key design parameters in S1; S42: the DRL agent randomly selects an initial packaging parameter combination from the feasible region and inputs it into the COMSOL multi-physical field coupling model constructed in S3 to obtain the corresponding sensor performance data; S43: calculating a reward function value according to the performance data, and the DRL agent updating parameters of the neural network through a gradient descent algorithm based on the reward value; S44: repeating steps S42-S43 until the reward function value converges to a stable range, and the outputted encapsulation parameter combination constituting an optimal solution set; S45: selecting an arbitrary parameter combination from the optimal solution set, verifying performance stability under different working conditions through a COMSOL model, and determining the final optimization parameter after confirming that there is no local optimal solution.

[0013] Preferably, the parameter optimization for the shell and the acoustic matching layer in S4 specifically comprises: adjusting the shell thickness and the material type and thickness of the acoustic matching layer by the DRL agent, so that the shell minimizes the attenuation of sound wave propagation under the premise of meeting electromagnetic shielding and structural strength; meanwhile, the acoustic matching layer realizes acoustic impedance matching between the piezoelectric material and the measured object, and the sound wave transmittance is improved to above a preset threshold, and the adjustment process needs to be optimized in coordination with the piezoelectric material thickness and the cavity size parameter, so as to avoid overall performance conflicts caused by single component parameter optimization.

[0014] Preferably, the method further comprises S5: dynamic adjustment and iteration of the optimization result; when the application scenario of the piezoelectric acoustic sensor changes, the weights of the optimization indicators in step S2 and the parameters of the reward function in step S4 are readjusted, the DRL model is incrementally trained based on historical training, and the optimal solution of the encapsulation parameters adapting to the new scenario is quickly outputted without the need to rebuild the COMSOL multi-physics field coupling model and the DRL basic framework.

[0015] The beneficial effects of the present application are as follows: 1. The present application proposes an improved piezoelectric acoustic sensor design and optimization method, which is suitable for application scenarios requiring high-precision detection of sound waves. The core lies in carefully designing state space parameter selection to ensure that the sensor maintains optimal performance under different conditions. Specifically, by setting the film thickness as a key component within a reasonable range, it not only ensures sufficient flexibility to respond to changes in external sound waves, but also reduces internal stress accumulation to some extent and maintains long-term stable performance. By designing a reasonable cavity size, it provides sufficient space for sound waves to propagate effectively inside, while avoiding the impact of excessive volume on device compactness. By adopting diversified electrode layout schemes such as circular, square or ring-shaped, it adapts to different installation needs and optimization directions. By determining a reasonable shell thickness, it effectively protects the internal structure from external environmental influences and ensures the structural firmness and lightweight. Finally, through multi-dimensional parameter collaborative design, the sensor's high-precision sound wave detection capability is stably outputted.

[0016] 2、The application realizes the maximization of the performance of the piezoelectric acoustic sensor by comprehensively considering the core targets such as sound pressure sensitivity and frequency response, and strictly controlling the amplitude of parasitic resonance, which not only greatly improves the core performance indicators of the sensor, but also effectively reduces unnecessary energy loss, ensures stable performance and consistent output performance in various complex application environments, and further enhances the function expansion capability of the sensor, significantly improves its market competitiveness, unlike the traditional idea of pursuing single indicator optimization, multi-objective optimization realizes the overall optimization of the overall performance of the sensor by reasonably configuring the key design parameters, which not only builds the core ability of high-precision sound wave detection, but also prolongs the service life, and finally greatly improves the overall efficiency of the sensor, fully adapts to the diversified needs of high-precision detection scenes.

[0017] 3、The application realizes the dual improvement of the performance and optimization efficiency of the piezoelectric acoustic sensor through multi-dimensional technology cooperation: through the use of deep reinforcement learning for multi-objective optimization, the complex feature engineering and model design link in traditional machine learning is avoided, and it can flexibly adapt to various problems and application environments, relying on the interaction feedback with the environment, and learning to get balanced solutions on multiple targets, providing more comprehensive and reliable support for optimization decision; by constructing a multi-physical field coupling model to simulate the working condition of the sensor, the accuracy of the simulation parameters is significantly improved, laying a data foundation for the improvement of the optimization effect; by introducing the adversarial generative network to effectively simulate the process disturbance, the robustness and adaptability of the optimization algorithm in the actual process are enhanced, and the performance fluctuation caused by manufacturing differences is greatly reduced; at the same time, by using active inverse sound waves in the working process of the sensor to realize resonance cancellation, the traditional static packaging is upgraded to a dynamic adaptive system, further improving the detection accuracy and operation reliability of the sensor in complex acoustic environment, and fully meeting the needs of high-precision sound wave detection. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a schematic diagram of the pickup system based on the piezoelectric acoustic sensor of the application; Figure 2 It is a packaging structure diagram of the piezoelectric acoustic sensor of the application; Figure 3 It is a DQN principle schematic diagram of the application; Figure 4 It is a DQN algorithm training schematic diagram of the application; Figure 5 It is a whole use flow schematic diagram of the optimization algorithm of the application; Figure 6 It is a training process schematic diagram of the DRL agent of the application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0020] As shown in Figures 1 to 6 The piezoelectric acoustic sensor packaging parameter multi-objective collaborative optimization algorithm provided by the embodiments of the present application has the following specific steps: S1: determining the key design parameters of the piezoelectric acoustic sensor packaging structure, the key design parameters being the parameters that have direct or indirect influence on the sensor performance and have priority in the degree of influence; S2: determining the multiple optimization indexes of the piezoelectric acoustic sensor packaging structure, the optimization indexes aiming to maximize the overall performance of the sensor, and there being a mutual dependence or constraint relationship between the indexes; S3: establishing a multi-physics field coupling model through COMSOL software, the model being used to simulate the structural characteristics and actual working conditions of the piezoelectric acoustic sensor; S4: converting the multi-objective optimization problem of the packaging parameters into a multi-objective decision problem based on deep reinforcement learning, searching for the optimal solution set of the key design parameters through the interaction learning of the DRL agent and the optimization environment, and realizing the collaborative balance of the multiple optimization indexes.

[0021] In the S1, the key design parameters include the thickness of the thin-film piezoelectric material, the size of the internal cavity of the sensor, the electrode layout mode, the thickness and shell material of the shell, and the material and thickness of the acoustic matching layer; the thin-film piezoelectric material is selected from at least one of PVDF film, PZT film, ZnO film and AlN film, and the shell material is a metal material having electromagnetic shielding and structural protection functions.

[0022] In the S2, the multiple optimization indexes include acoustic pressure sensitivity, signal-to-noise ratio (SNR), temperature drift, frequency offset and process disturbance; the acoustic pressure sensitivity is defined as the ratio of the output electric signal of the sensor to the input acoustic pressure under unit acoustic pressure, with the unit of mV / Pa or mV / µbar (1µbar≈0.1Pa), and the expression is S=output electric signal (voltage / charge) / input acoustic pressure; the multiple optimization indexes aim to simultaneously improve the acoustic pressure sensitivity and the signal-to-noise ratio, and reduce the influence of temperature drift, frequency offset and process disturbance on the performance of the sensor.

[0023] The multi-physical field coupling model in S3 is constructed by adding an acoustic module, a mechanical module, an electric field module, and a thermal module in COMSOL; in the model simulation process, the sound impedance matching effect of the acoustic matching layer between the piezoelectric material and the measured object, the electromagnetic shielding effect of the shell, and the influence of the cavity on the vibration and stress distribution of the piezoelectric material need to be considered, and the model needs to restore the vibration-stress-potential difference conversion process of the piezoelectric material under the action of the acoustic signal.

[0024] In S4, the deep reinforcement learning is realized by combining deep learning and reinforcement learning; the reinforcement learning learns and optimizes the behavior strategy through trial and error interaction between the DRL agent and the optimization environment, with the goal of maximizing the long-term cumulative reward signal; the deep learning constructs a multi-layer neural network as a value function approximator or a policy approximator, the value function approximator is used to estimate the long-term cumulative reward value corresponding to different packaging parameter combinations, and the policy approximator is used to directly generate the adjustment strategy of the packaging parameters, and the neural network can extract features from sensor performance data to realize end-to-end training and complex environment modeling.

[0025] In S4, the reward function of the DRL is designed as a multi-objective reward function, and the output value of the reward function is directly related to the multiple optimization indicators in S2: when the sound pressure sensitivity of the sensor is improved, the signal-to-noise ratio is improved, the temperature drift is reduced, the frequency offset is reduced, and the process disturbance is reduced, the reward function outputs positive reward; when the performance of the above optimization indicators degrades, it outputs negative reward; and the reward function realizes the trade-off between different optimization indicators through a weight adjustment mechanism, avoiding the imbalance of the overall performance caused by the optimization of a single indicator.

[0026] The calculation formula of the reward function is: One design method of the reward function is as follows: R = ω1(S p S min ) ω2∣Δf f target ∣ ω3A resonance , wherein ω1, ω2, and ω3 are weight coefficients for balancing the priority of the target, which can be pre-adjusted through experiments, S p is the simulated sound pressure sensitivity, S min is the minimum sound pressure sensitivity required by the task, Δf is the simulated frequency offset value, f target is the target frequency offset value (preset value), and A resonance is the simulated parasitic resonance amplitude combined with the generative adversarial network (GAN) to simulate process disturbance and improve algorithm robustness. The generation of simulation data depends on the pre-established multi-physical field coupling model, and the output strategy is defined as: after convergence, the maximum The action 'a' is adjusted as the optimal parameter.

[0027] In particular, the construction of the optimized environment in S4 needs to incorporate the influence of process disturbances: by using the parasitic resonance amplitude analysis method and generative adversarial network (GAN) to simulate process disturbance scenarios, such as thin film thickness deviation, cavity size error, and electrode misalignment, the parasitic parameters (inductance / capacitance) changes and resonance characteristic degradation data of the sensor under different process disturbances are simulated, and this data is used as the environmental state input of the DRL agent, so that the optimized packaging parameters have the ability to resist process disturbances.

[0028] The training process of the DRL agent in step S4 includes: S41: Initialize the feasible region of the packaging parameters, which is determined based on the physical constraints of the key design parameters in S1 (such as the lower limit of material thickness and the upper limit of cavity volume). S42: The DRL agent randomly selects an initial combination of encapsulation parameters from the feasible domain and inputs it into the COMSOL multiphysics coupling model constructed in S3 to obtain the corresponding sensor performance data (i.e., the value of the optimization index). S43: Calculate the reward function value based on the performance data, and update the parameters (value function or policy) of the neural network based on the reward value using the gradient descent algorithm. S44: Repeat steps S42-S43 until the reward function value converges to a stable range. At this point, the combination of the output encapsulated parameters constitutes the optimal solution set. S45: Select any combination of parameters from the set of optimal solutions, verify its performance stability under different working conditions (such as temperature change, sound pressure intensity change) through the COMSOL model, and determine it as the final optimized parameter after confirming that there is no local optimum.

[0029] Specifically, the parameter optimization for the shell and acoustic matching layer in S4 includes: adjusting the shell thickness and the material type and thickness of the acoustic matching layer through the DRL agent, so that the shell minimizes the attenuation of sound wave propagation while meeting the requirements of electromagnetic shielding and structural strength; at the same time, the acoustic matching layer achieves acoustic impedance matching between the piezoelectric material and the object under test, increasing the sound wave transmittance to above a preset threshold. Moreover, the adjustment process needs to be coordinated with the piezoelectric material thickness and cavity size parameters to avoid overall performance conflicts caused by the optimization of parameters of a single component.

[0030] The method also includes S5: dynamic adjustment and iteration of optimization results; when the application scenario of the piezo-acoustic sensor (such as the frequency range of the measured sound signal and the ambient temperature range) changes, the weights of the optimization index in step S2 and the parameters of the reward function in step S4 are readjusted, and incremental training is performed based on the historically trained DRL model to quickly output the optimal solution of the encapsulation parameters adapted to the new scenario without having to rebuild the COMSOL multiphysics coupling model and the DRL basic framework.

[0031] Example of a multi-objective collaborative optimization algorithm for piezoacoustic sensor packaging parameters

[0032] This embodiment uses a piezo-acoustic sensor for industrial equipment fault diagnosis as the application carrier. This sensor needs to meet the high-precision pickup requirement of vibration acoustic signals from components in the frequency band of ≥20-2000Hz, while also possessing the ability to resist environmental temperature drift of ≥20~80℃ and adapt to manufacturing process fluctuations. Based on the aforementioned optimization algorithm, the packaging parameters are collaboratively optimized. The specific implementation process is as follows: I. Determination of Key Design Parameters and Optimization Indicators (I) Selection and Constraints of Key Design Parameters Based on the requirements of industrial scenarios and physical process limitations, the key design parameters and feasible domains are determined as follows: A PZT-5H thin film with strong piezoelectric effect and excellent stability is selected as the piezoelectric material, with a thickness of 5-20 μm (too thin and it is easily damaged; too thick and the sensitivity is insufficient); the internal cavity size of the sensor is limited to 4-8 mm × 4-8 mm × 1-3 mm, adapting to an overall installation size of ≤10 × 10 × 5 mm; the electrodes adopt a ring layout (to ensure uniform force distribution and reduce signal distortion), with an electrode width of 0.2-0.5 mm and a spacing of 0.1-0.3 mm; the shell is made of 6061 aluminum alloy, which combines electromagnetic shielding and lightweight characteristics, with a thickness of 0.5-2 mm (to balance structural strength and sound wave attenuation); the acoustic matching layer uses epoxy resin with excellent acoustic impedance compatibility, with a thickness of 10-50 μm (to achieve acoustic impedance matching between the piezoelectric material and the metal device).

[0033] (ii) Quantitative setting of multiple optimization indicators

[0034] With "improving sound-to-electric conversion efficiency by ≥ + reducing environmental and process interference" as the core, the following quantitative targets are set: sound pressure sensitivity ≥ 50mV / Pa (covering weak vibration signals, calculated based on S ≥ = = ≥ output voltage ≥ / ≥ input sound pressure); signal-to-noise ratio (SNR) ≥ 60dB (resisting industrial electromagnetic noise); temperature drift ≤ 0.1mV / (Pa). ±0.05mm, the sensor performance attenuation needs to be ≤10%.

[0035] II. COMSOL multi-physical field coupling model construction

[0036] Based on COMSOL 6.0 software, four modules of acoustics, mechanics, electric field, and heat are integrated to build a coupling model to restore the actual working condition of the sensor: the frequency domain analysis is adopted in the acoustic module, and a sound pressure load of 0.1-1 Pa, a frequency range of 20-2000 Hz is set; the linear elastic physical field is selected in the mechanical module, and the mechanical parameters of PZT film (elastic modulus of 70 GPa), aluminum alloy shell (69 GPa), and epoxy resin acoustic matching layer (3.5 GPa) are defined; the voltage monitoring points are set on the upper and lower electrodes of the PZT film in the electric field module, and the conversion of “vibration - stress - potential difference” is simulated; the temperature load of -20~80℃ is set in the thermal module, and the change of piezoelectric coefficient (d33) with temperature is monitored. The model also simulates key effects: the acoustic matching layer makes the acoustic transmittance greater than 85% by adjusting the thickness, the cavity avoids resonance with the equipment vibration frequency, and the shell realizes electromagnetic noise attenuation of more than 20 dB.

[0037] III. Deep reinforcement learning (DRL) multi-objective optimization implementation

[0038] (I) DRL core component design

[0039] The deep deterministic policy gradient (DDPG) algorithm is adopted to adapt to continuous parameter optimization, and a double-layer neural network is constructed: the policy network (3 layers of full connection) takes “parameter combination package + COMSOL performance data” as input and outputs parameter adjustment amount; the value function network (3 layers of full connection) inputs the same as above and outputs the long-term cumulative reward estimate value. The reward function design adapts to the industrial scene, and the formula is: , where , , , (prioritize sensitivity and frequency stability), is the simulated sensitivity, , is the frequency offset, , To generate a Generative Adversarial Network (GAN) to simulate the parasitic resonance amplitude after process disturbance, To generate a temperature drift.

[0040] (ii) DRL is greater than or equal to the training and validation of the agent

[0041] Initialization: Based on the parameter feasible region to exclude physically contradictory combinations, determine the initial search range; Initial simulation: Randomly select a parameter combination greater than or equal to PZT greater than or equal to thin film greater than or equal to 10 μm, cavity greater than or equal to 5 × 5 × 2 mm, shell greater than or equal to 1 mm, acoustic matching layer greater than or equal to 30 μm, input the model to get the initial performance: sensitivity greater than or equal to 45 mV / Pa, SNR 55 dB, temperature drift greater than or equal to 0.15 mV / (Pa ℃), frequency offset greater than or equal to 80 Hz, resonance amplitude under process disturbance greater than or equal to 0.2 V, and the negative reward is greater than or equal to -24.035; Iterative update: Update the network parameters through the gradient descent algorithm, repeat the simulation and update steps greater than or equal to 5000 times, and when the reward function is stable in greater than or equal to [-1.0, greater than or equal to 0.5] greater than or equal to 100 times, the optimal parameter combination (PZT greater than or equal to thin film greater than or equal to 12 μm, cavity greater than or equal to 6 × 6 × 1.5 mm, shell greater than or equal to 0.8 mm, acoustic matching layer greater than or equal to 25 μm) is output; Performance verification: Substitute the optimal parameters into the model to verify the performance under different working conditions: sensitivity greater than or equal to 58 mV / Pa, SNR 63 dB, temperature drift greater than or equal to 0.08 mV / (Pa ℃), frequency offset greater than or equal to 32 Hz, resonance amplitude under process disturbance greater than or equal to 0.07 V, all indicators meet the requirements and fluctuate ≤5%, and local optimum is excluded.

[0042] Four, dynamic adjustment of optimization results

[0043] When the application scenario changes to underwater detection (requirements: water pressure resistance greater than or equal to 0.1-1 MPa, temperature greater than or equal to 0-30℃, frequency greater than or equal to 500-5000 Hz), without rebuilding the model and greater than or equal to DRL greater than or equal to the basic framework: adjust the optimization index weight (temperature drift weight decreases to greater than or equal to 0.05, and a new water pressure resistance weight greater than or equal to 0.2 is added), update the reward function to add a water pressure stability term , For water pressure stability, the target is greater than or equal to 0.9; based on history greater than or equal to DRL greater than or equal to model supplement greater than or equal to 1000 greater than or equal to sub-incremental training, output adaptive parameters (PZT greater than or equal to thin film greater than or equal to 15 mu m, cavity greater than or equal to 7*7*1.2 mm, shell greater than or equal to 1.2 mm, polytetrafluoroethylene acoustic matching layer greater than or equal to 20 mu m), simulation verification meets the underwater detection demand.

[0044] V. Implementation effect

[0045] Compared with the traditional "empirical formula + single lightweight design", the comprehensive performance of the optimized sensor of the present algorithm is significantly improved: the sound pressure sensitivity is improved from greater than or equal to 42 mV / Pa to greater than or equal to 58 mV / Pa (an increase of greater than or equal to 38.1%), the signal-to-noise ratio is improved from greater than or equal to 52 dB to greater than or equal to 63 dB (an increase of greater than or equal to 21.2%), the temperature drift is reduced from greater than or equal to 0.18 mV / (Pa ℃) to greater than or equal to 0.08 mV / (Pa ℃) (a decrease of greater than or equal to 55.6%), and the performance attenuation under process disturbance is reduced from greater than or equal to 22% to greater than or equal to 8% (a decrease of greater than or equal to 63.6%). It is fully adapted to the fault diagnosis of industrial equipment and the expansion demand in multiple scenes.

[0046] Figure 1 It is a schematic diagram of a pickup system based on a piezoelectric acoustic sensor. The pickup system includes a piezoelectric acoustic sensor and an amplification circuit. The piezoelectric acoustic sensor can perceive sound and convert the vibrated acoustic signal into an original electric signal. Since the original electric signal is weak and cannot be used directly, it is generally amplified by the amplification circuit. The amplified electric signal enters the audio system for processing.

[0047] Figure 2 It is a packaging structure diagram of a piezoelectric acoustic sensor. As shown in Figure 2 , the piezoelectric acoustic sensor includes electrodes and piezoelectric material. When the acoustic signal causes the piezoelectric material to vibrate, the region of stress concentration in the piezoelectric material will generate a potential difference between its upper and lower electrodes, so that the acoustic signal can be converted into an electric signal and extracted.

[0048] The DQN algorithm training is as shown in Figure 4 . The deep reinforcement learning algorithm replaces the Q-table of the traditional reinforcement learning algorithm with a neural network to estimate the value of each action, completes the transition from a discrete state space to a continuous state space, takes the state of the agent in the search space as the input of the Q greater than or equal to Network, and outputs the corresponding Q value of each action (the action in the present application), to obtain the action to be executed.

[0049] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify a subject or action, without necessarily requiring or implying any such actual relationship or order between such subjects or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations of the embodiments can be undertaken without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A multi-objective collaborative optimization algorithm for piezoelectric acoustic sensor packaging parameters, characterized in that: The specific steps of this optimization algorithm are as follows: S1: Determine the key design parameters of the piezo-acoustic sensor packaging structure. The key design parameters are those that have a direct or indirect impact on the sensor performance and have a priority in terms of impact. S2: Determine multiple optimization indicators for the piezo-acoustic sensor packaging structure. The optimization indicators aim to maximize the overall performance of the sensor, and there are inter-dependencies or constraints between the indicators. S3: A multiphysics coupling model is established using COMSOL software. This model is used to simulate the structural characteristics and actual working conditions of the piezo-acoustic sensor. S4: Based on deep reinforcement learning, the multi-objective optimization problem of the encapsulated parameters is transformed into a multi-objective decision problem. Through the interactive learning between the DRL agent and the optimization environment, the optimal solution set of the key design parameters is searched to achieve the synergistic balance of multiple optimization indicators.

2. The piezoelectric acoustic sensor package parameter multi-objective co-optimization algorithm according to claim 1, characterized in that: The key design parameters in S1 include the thickness of the thin-film piezoelectric material, the size of the internal cavity of the sensor, the electrode layout, the thickness and material of the outer shell, and the material and thickness of the acoustic matching layer; wherein, the thin-film piezoelectric material is selected from at least one of PVDF film, PZT film, ZnO film, and AlN film, and the outer shell material is a metal material with electromagnetic shielding and structural protection functions.

3. The piezoelectric acoustic sensor package parameter multi-objective co-optimization algorithm of claim 1, wherein: The optimization indicators in S2 include sound pressure sensitivity, signal-to-noise ratio, temperature drift, frequency shift, and process disturbance; wherein, the sound pressure sensitivity is defined as the ratio of the sensor's output electrical signal to the input sound pressure under a unit sound pressure, and the expression is S = output electrical signal / input sound pressure; The goal of these multiple optimization metrics is to simultaneously improve acoustic pressure sensitivity and signal-to-noise ratio, while reducing the impact of temperature drift, frequency shift, and process disturbances on sensor performance.

4. The piezoelectric acoustic sensor package parameter multi-objective co-optimization algorithm of claim 1, wherein: The multiphysics coupling model in S3 is constructed by adding acoustic, mechanical, electric, and thermal modules to COMSOL. During the simulation process, the acoustic impedance matching effect of the acoustic matching layer between the piezoelectric material and the object under test, the electromagnetic shielding effect of the shell, and the influence of the cavity on the vibration and stress distribution of the piezoelectric material must be considered. The model must also reproduce the vibration-stress-potential difference conversion process of the piezoelectric material under the action of acoustic signals.

5. The piezoelectric acoustic sensor package parameter multi-objective co-optimization algorithm of claim 1, wherein: The deep reinforcement learning in S4 is achieved by combining deep learning and reinforcement learning. Reinforcement learning optimizes behavioral strategies through trial-and-error interaction between the DRL agent and the optimization environment, with the goal of maximizing the long-term cumulative reward signal. Deep learning constructs a multi-layer neural network as a value function approximator or a policy approximator. The value function approximator is used to estimate the long-term cumulative reward value corresponding to different combinations of encapsulation parameters, and the policy approximator is used to directly generate adjustment strategies for encapsulation parameters. The neural network can extract features from sensor performance data to achieve end-to-end training and complex environment modeling.

6. The piezoelectric acoustic sensor package parameter multi-objective co-optimization algorithm of claim 1, wherein: The reward function of DRL in S4 is designed as a multi-objective reward function. The output value of the reward function is directly related to multiple optimization indicators in S2: when the sensor's sound pressure sensitivity is improved, the signal-to-noise ratio is improved, the temperature drift is reduced, the frequency offset is reduced, and the impact of process disturbances is reduced, the reward function outputs a positive reward; when the performance of the above optimization indicators degrades, a negative reward is output; and the reward function achieves a trade-off between different optimization indicators through a weight adjustment mechanism to avoid the overall performance imbalance caused by the optimization of a single indicator. The formula for calculating the reward function is as follows: One way to design the reward function is as follows: R = ω1(S p S min ) ω2∣Δf f target ∣ ω3A resonance ; where ω1, ω2, ω3 are weight coefficients for balancing the target priority, pre-optimized through experiments, S p is the simulated sound pressure sensitivity, S min is the minimum sound pressure sensitivity required by the task, Δf is the simulated frequency offset value, f target is the target frequency offset value, A resonance is the simulated parasitic resonance amplitude, which combines the generative adversarial network to simulate process disturbance, to improve the robustness of the algorithm, and the generation of simulation data depends on the pre-established multi-physical field coupling model; the output strategy is defined as: after convergence, the action a that maximizes is selected as the optimal parameter adjustment.

7. The multi-objective collaborative optimization algorithm for piezo-acoustic sensor packaging parameters according to claim 1, characterized in that: The construction of the optimized environment in S4 needs to incorporate the influence of process disturbances: the process disturbance scenarios are simulated by parasitic resonance amplitude analysis method and generative adversarial network. The process disturbance scenarios include thin film thickness deviation, cavity size error and electrode misalignment. The parasitic parameter changes and resonance characteristic degradation data of the sensor under different process disturbances are simulated and the data is used as the environmental state input of the DRL agent, so that the optimized packaging parameters have the ability to resist process disturbances.

8. The multi-objective collaborative optimization algorithm for piezo-acoustic sensor packaging parameters according to claim 1, characterized in that: The training process of the DRL agent in step S4 includes: S41: Initialize the feasible region of the packaging parameters, which is determined based on the physical constraints of the key design parameters in S1; S42: The DRL agent randomly selects an initial combination of encapsulation parameters from the feasible domain and inputs it into the COMSOL multiphysics coupling model constructed in S3 to obtain the corresponding sensor performance data; S43: Calculate the reward function value based on the performance data, and update the parameters of the neural network based on the reward value using the gradient descent algorithm; S44: Repeat steps S42-S43 until the reward function value converges to a stable range. At this point, the combination of the output encapsulated parameters constitutes the optimal solution set. S45: Select any combination of parameters from the set of optimal solutions, verify its performance stability under different working conditions using the COMSOL model, and determine it as the final optimization parameter after confirming that there is no local optimum.

9. The multi-objective collaborative optimization algorithm for piezo-acoustic sensor packaging parameters according to claim 1, characterized in that: The parameter optimization of the shell and acoustic matching layer in S4 specifically includes: adjusting the shell thickness and the material type and thickness of the acoustic matching layer through the DRL agent, so that the shell minimizes the attenuation of sound wave propagation while meeting the requirements of electromagnetic shielding and structural strength; at the same time, the acoustic matching layer achieves acoustic impedance matching between the piezoelectric material and the object under test, and improves the sound wave transmittance to above the preset threshold. Moreover, the adjustment process needs to be coordinated with the piezoelectric material thickness and cavity size parameters to avoid overall performance conflicts caused by the optimization of parameters of a single component.

10. The multi-objective collaborative optimization algorithm for piezo-acoustic sensor packaging parameters according to claim 1, characterized in that: The method also includes S5: dynamic adjustment and iteration of optimization results; when the application scenario of the piezo-acoustic sensor changes, the weights of the optimization index in step S2 and the parameters of the reward function in step S4 are readjusted, incremental training is performed based on the historically trained DRL model, and the optimal solution of the encapsulation parameters adapted to the new scenario is quickly output without rebuilding the COMSOL multiphysics coupling model and the DRL basic framework.

Images