Intelligent acoustic impedance matching method and system for water turbine based on deep learning

Abstract

The present application relates to the field of intelligent operation and maintenance of hydroelectric power generation and active acoustic control, and discloses a method and system for intelligent acoustic impedance matching of a hydraulic turbine based on deep learning. The method comprises: collecting full-flow multi-source heterogeneous data to construct a sensor topology graph; using a spatiotemporal graph neural network for feedforward prediction of flow field excitation characteristics; using a multi-agent deep reinforcement learning algorithm with integrated physical constraint learning to make decisions on optimal configuration instructions; actively canceling intrinsic capacitance through a negative impedance converter, reconstructing the equivalent mechanical stiffness of the superstructure, and realizing real-time matching and noise reduction of acoustic impedance; and performing online fine-tuning based on residual noise feedback. The system comprises perception, reasoning, decision-making, execution and feedback optimization modules. The present application breaks through the physical limitations such as the Pashen limit through a non-Foster circuit, significantly enhances the suppression effect of ultra-low frequency below 100 Hz, effectively solves the control lag problem through feedforward logic, improves the stability of complex working conditions, and ensures the adaptability and robustness throughout the life cycle by combining physical constraints and incremental learning.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and active acoustic control technology for hydropower generation, specifically to a method and system for intelligent acoustic impedance matching of hydro turbines based on deep learning. Background Technology

[0002] As the core power conversion equipment in a hydropower station, the operating status of the turbine directly affects the safety of the power grid and the quality of power. However, in actual operation, the turbine is constantly under a complex fluid dynamic environment. The intense interaction between the water flow and the runner blades, guide vanes, and volute induces severe fluid-structure interaction vibrations, which in turn generate complex nonlinear, broadband noise. Among these, the runner blade passing frequency (BPF) and its higher-order harmonic noise are the main components of turbine noise. This noise not only deteriorates the working environment of the power station personnel, but more seriously, the long-term acoustic fatigue effect can cause the propagation of micro-cracks in the turbine's metal structure, and even lead to fatigue damage to structural components, severely shortening the unit's service life and increasing maintenance costs.

[0003] In existing hydro turbine noise reduction engineering practices, passive noise reduction technology is the most widely used method, mainly relying on laying physical sound-absorbing materials on the flow channel walls or arranging resonant cavities with specific geometric structures. However, passive noise reduction schemes face insurmountable physical limitations. Since the physical geometric parameters of the sound-absorbing materials and resonant cavities are fixed after installation, their sound absorption frequency characteristics exhibit narrow band characteristics, making it impossible to track the frequency drift caused by load changes, speed fluctuations, or guide vane opening adjustments in the hydro turbine. Under variable operating conditions, passive noise reduction devices often deviate from their optimal operating point, resulting in a significant decrease in noise reduction effect. Furthermore, acoustic causality limitations dictate that absorbing low-frequency noise usually requires a large physical volume or thickness, while the internal space of the hydro turbine flow channel is extremely compact and requires maintaining the flatness of the flow surface. This makes controlling ultra-low frequency noise below 100Hz within a limited space a recognized technical bottleneck in the industry.

[0004] To overcome the limitations of passive noise reduction, active noise control (ANC) technology has been gradually introduced into the field of hydraulic machinery. Traditional active control systems mostly employ feedback-based control algorithms, such as proportional-integral-derivative (PID) control or least mean square (LMS) adaptive filtering algorithms. However, in the highly turbulent, non-stationary, and transiently nonlinear acoustic environment of a water turbine, the performance of traditional algorithms is unsatisfactory. Due to the extremely rapid changes in the water flow field, the traditional "sensing-feedback-regulation" mode exhibits significant control lag, and the convergence speed of the algorithm often cannot keep up with the abrupt changes in the noise signal, leading to system instability or even acoustic feedback oscillations under highly dynamic transient conditions. Furthermore, existing technologies lack effective means for the coordinated control of non-uniformly distributed multi-point sound sources, making it difficult to cope with the multidimensional sound pressure characteristics brought about by complex flow fields.

[0005] In summary, existing noise reduction technologies have significant shortcomings in adapting to all operating conditions, achieving ultra-low frequency noise suppression, and providing rapid real-time response. Therefore, how to break through the physical and algorithmic limitations of traditional acoustic control and develop an adaptive intelligent noise reduction system capable of real-time prediction of complex flow fields in hydro turbines, possessing wide-bandwidth (especially ultra-low frequency) suppression capabilities, and achieving efficient multi-point coordination has become a key technical problem urgently needing to be solved in the field of intelligent operation and maintenance and active acoustic control in hydropower generation. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based intelligent acoustic impedance matching method and system for water turbines, in order to solve the problems of narrow bandwidth and large size of passive noise reduction in existing water turbine noise reduction technologies, as well as the problems of control lag, instability and difficulty in dealing with non-uniform multidimensional sound fields in traditional active control algorithms under high dynamic transient conditions.

[0007] To achieve the above objectives, this invention provides a deep learning-based intelligent acoustic impedance matching method for hydro turbines, the method comprising:

[0008] Acquire multi-source heterogeneous data across the entire flow channel and construct a sensor topology map. By deploying multi-source sensing modules at key locations in the turbine flow channel (including the volute inlet, guide vane area, runner chamber, and tailrace pipe), heterogeneous data reflecting the unit's operating conditions and dynamic characteristics of the acoustic field are collected synchronously, and a spatiotemporal topology map of the entire flow channel sensors is constructed by combining the physical location coordinates of the sensors.

[0009] The flow field excitation characteristics are predicted based on the spatiotemporal graph neural network (STGNN). The heterogeneous data is input into the pre-trained STGNN feedforward model to capture the evolution of water pressure fluctuations in the flow channel space and their temporal correlation. The dominant frequency, phase and amplitude characteristics of the flow field excitation are predicted in advance 5-10 ms before the noise actually reaches the surface of the acoustic superstructure.

[0010] The circuit configuration instructions are determined using a multi-agent deep reinforcement learning (MARL) algorithm. Using the aforementioned flow field excitation characteristics as state input, the optimal non-Foster circuit configuration instructions are determined using a multi-agent deep reinforcement learning algorithm with integrated physical constraint learning (PINN) mechanism. It addresses non-uniform sound pressure fields through communication and collaboration among intelligent agents.

[0011] Real-time acoustic impedance reconstruction is achieved through non-Foster active reactance compensation. The negative impedance converter (NIC) integrated in the intelligent acoustic superstructure is adjusted in real time according to the configuration command. The negative capacitance and negative inductance characteristics are generated by the operational amplifier feedback loop to actively cancel the intrinsic capacitance of the piezoelectric material, reconstruct the equivalent mechanical stiffness of the superstructure, and achieve acoustic impedance matching and broadband noise suppression.

[0012] Online fine-tuning and self-evolution management are performed based on residual noise feedback. The residual noise signal after noise reduction is monitored in real time and fed back as a reward value to the reinforcement learning framework for policy iteration, so as to realize online performance fine-tuning and long-term self-evolution management of the system under all operating conditions.

[0013] Preferably, acquiring multi-source heterogeneous data across the entire flow channel includes: synchronously acquiring the turbine's rotational speed via a distributed sensor array. Guide vane opening Pressure pulsation at key flow points and residual noise in the anechoic area Construct data state vector .

[0014] Preferably, the spatiotemporal graphical neural network (STGNN) feedforward model performs feedforward prediction specifically including:

[0015] Spatial feature extraction: Graph Convolutional Network (GCN) is used to process the sensor topology map to identify the physical path and interaction of pressure pulsations in the flow channel space.

[0016] By capturing temporal patterns and combining them with Temporal Convolutional Networks (TCN) or Long Short-Term Memory Networks (LSTM) to extract the temporal dependence of pressure pulsations, the accuracy of prediction results under highly dynamic transient conditions is ensured.

[0017] The feedforward logic implementation leverages the parallel computing capabilities of the model to complete inference before the noise signal reaches the superstructure surface, thus upgrading the system control logic from the traditional "hysteresis feedback" to "real-time prediction".

[0018] Preferably, the multi-agent deep reinforcement learning (MARL) algorithm includes:

[0019] Agent-independent modeling abstracts each resonant unit on the acoustic superstructure surface into an independent agent. Each agent independently calculates circuit parameters based on local perception data and global prediction features.

[0020] Distributed cooperative control, where neighboring intelligent agents exchange state information through communication protocols, utilizes a centralized training-decentralized execution (CTDE) architecture to achieve spatial cooperative hedging and solve the problem of non-uniform sound pressure field governance;

[0021] Physical consistency constraints are implemented by explicitly introducing the residuals of the Navier-Stokes equations and acoustic wave equations into the reward function. Physical constraint learning (PINN) ensures that the circuit parameters output by the algorithm conform to the physical laws of fluid-structure interaction.

[0022] Preferably, the non-Foster active reactance compensation and reconstructed equivalent mechanical stiffness specifically includes:

[0023] The NIC circuit transforms the load impedance using the positive feedback mechanism of the operational amplifier, thus generating a negative capacitance. ;

[0024] Intrinsic capacitance cancellation, precise adjustment The value is made similar to the intrinsic capacitance of the piezoelectric material. They tend to cancel each other out, eliminating the capacitive component of the system and breaking the Paschen limit's restriction on the physical thickness of the superstructure;

[0025] Inverse piezoelectric coupling utilizes the inverse piezoelectric effect to couple the adjusted electrical properties into the mechanical structure in real time, dynamically changing the mechanical response of the superstructure so that its acoustic impedance matches the incident sound wave in real time in the ultra-low frequency band (below 100Hz).

[0026] Preferably, the reward value update based on residual noise feedback specifically includes:

[0027] The core optimization metric, the reward function, is based on the reduction in the total sound power level of the target frequency band. It drives the model convergence by maximizing the negative reward (i.e. minimizing the sound power level).

[0028] Online parameter fine-tuning: When the system detects long-cycle characteristic drift caused by turbine aging and wear, it automatically triggers an online incremental learning process to update the neural network weights.

[0029] In a second aspect of the present invention, a deep learning-based intelligent acoustic impedance matching system for a hydro turbine is provided, the system comprising:

[0030] The sensing module is used to collect heterogeneous data of the entire flow channel of the water turbine through a multi-source sensor array and to construct a sensor topology map based on physical coordinates.

[0031] The inference module, which embeds the STGNN prediction model, is used to extract the spatiotemporal features of the flow field in real time and feedforward to predict the noise frequency and amplitude at future moments.

[0032] The decision module runs the MARL algorithm and integrates PINN constraints to generate optimal NIC configuration instructions based on predicted features and agent communication information.

[0033] The execution module includes an intelligent acoustic superstructure with an integrated negative impedance converter (NIC) circuit, which is used to change the structural stiffness in real time and suppress noise through active reactance compensation;

[0034] The feedback optimization module is used to evaluate the system performance based on the residual noise signal after noise reduction and to drive the online self-evolution management of the model.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] This invention achieves ultra-wideband noise suppression by integrating non-Foster circuitry (NIC) to generate negative capacitance and inductance characteristics, actively canceling the intrinsic capacitance of piezoelectric materials and thus breaking the acoustic causality limit (Paschen limit). This allows the acoustic superstructure to achieve near-perfect absorption of broadband noise while maintaining an extremely thin physical thickness, particularly significantly enhancing the suppression of ultra-low frequency pulsations below 100Hz, thus solving the technical bottleneck between limited space in turbine flow channels and effective low-frequency noise control.

[0037] With extremely low latency and real-time prediction and response capabilities, this invention employs a Spatiotemporal Graphical Neural Network (STGNN) to construct feedforward control logic, enabling feature extraction and inference to be completed 5-10 ms before the noise wave actually reaches the anechoic surface. Compared to the traditional lag-based "sensing-feedback-adjustment" model, this system can effectively predict sudden noise changes, significantly improving the stability and response accuracy of the system under complex transient conditions such as turbine startup, load shedding, and rapid opening and closing of guide vanes, while avoiding the risk of acoustic feedback oscillation.

[0038] This invention achieves highly intelligent distributed collaboration and physical consistency by employing a multi-agent deep reinforcement learning (MARL) algorithm. It treats several resonant units on the superstructure surface as a collaborative whole capable of communication. This distributed collaboration mechanism can accurately address the extremely complex non-uniform sound pressure field distribution inside the turbine. Simultaneously, the introduction of Physically Constrained Learning (PINN) embeds physical principles such as the Navier-Stokes equations into the AI ​​algorithm, ensuring that the generated control strategy is not only data-driven but also fully conforms to the physical logic of fluid-structure interaction, significantly improving the system's robustness and reliability under various extreme conditions.

[0039] The system features adaptive and self-evolving management throughout its entire lifecycle. Its feedback optimization module, combined with an online incremental learning mechanism, continuously optimizes strategies based on residual noise feedback. This enables the system not only to cope with short-term load fluctuations but also to adaptively compensate for acoustic characteristic drift caused by mechanical wear and component aging during the turbine's operation over several years. This achieves closed-loop optimization of noise reduction performance throughout its entire lifecycle, reducing subsequent manual maintenance costs. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the process of a deep learning-based intelligent acoustic impedance method for water turbines proposed in this invention.

[0041] Figure 2 This is a schematic diagram illustrating the principle of active reactance compensation for non-Foster circuits (NICs) in this invention.

[0042] Figure 3 This is a schematic diagram illustrating the interaction logic between the STGNN feedforward model and the MARL decision algorithm in this invention.

[0043] Figure 4 This is a structural diagram of a deep learning-based intelligent acoustic impedance matching system for water turbines proposed in this invention.

[0044] Figure 5 This is a schematic diagram of the electronic device in this invention. Detailed Implementation

[0045] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0046] Example 1

[0047] Please refer to Figure 1 and Figure 3 As shown, in a first aspect of the present invention, a deep learning-based intelligent acoustic impedance matching method for water turbines is provided. This method achieves precise suppression of noise throughout the entire flow path of the water turbine through a closed-loop logic of perception, prediction, decision-making, execution, and feedback. Specifically, it includes the following steps:

[0048] S101, Full-channel multi-source heterogeneous data sensing and sensor topology construction

[0049] This embodiment first uses a distributed multi-source sensor array deployed at key locations in the turbine flow channel (including but not limited to the volute inlet, guide vane area, runner chamber, and tailrace) to simultaneously collect heterogeneous data reflecting the real-time operating conditions and dynamic characteristics of the acoustic field of the unit. Specifically, the data collected by the sensing module includes the rotational speed, which reflects the macroscopic operating state of the unit. With guide vane opening And pressure pulsation at key nodes of the flow channel, reflecting the dynamic characteristics of the micro-flow field. In addition, the system will monitor the residual noise signal behind the anechoic area in real time. The aforementioned heterogeneous parameters are constructed into a multidimensional data state vector:

[0050]

[0051] In this step, the sensor does not operate in isolation, but rather is associated with its physical location coordinates (3D spatial coordinates) within the entire flow channel to construct a spatiotemporal topology map of the sensor. Those skilled in the art will understand that by constructing such a topology map with spatial attributes, geometric constraints can be provided for subsequent identification of the physical propagation path of noise within the flow channel, thereby more accurately characterizing the evolution logic of the sound field on complex flow surfaces.

[0052] S102. Perform feedforward prediction of flow field excitation characteristics based on spatiotemporal graph neural network (STGNN).

[0053] The system inputs the state vector and spatiotemporal topology map obtained in step S101 into a pre-trained spatiotemporal graph neural network (STGNN) feedforward model. This prediction process is divided into two dimensions: spatial feature extraction and temporal pattern capture. First, the topology map is processed using a graph convolutional network (GCN) to identify the interaction and propagation path of pressure pulsations in the flow channel space. Then, a temporal convolutional network (TCN) or a long short-term memory network (LSTM) is combined to capture the temporal dependence of pressure pulsations.

[0054] The core technological innovation of this embodiment lies in performing "feedforward prediction" rather than "hysteresis feedback." Utilizing the model's strong parallel computing capabilities, the system can complete inference 5-10 ms before the noise wave actually reaches the acoustic superstructure surface, predicting the dominant frequency, phase, and amplitude characteristics of the flow field excitation in advance. This feedforward mechanism effectively solves the control lag problem caused by rapid changes in the water flow field, ensuring that the system maintains high-precision response even under highly dynamic transient conditions (such as load shedding or rapid adjustment of guide vane opening), avoiding the instability common in traditional ANC systems.

[0055] S103. Configuring instructions for decision circuits using Multi-Agent Deep Reinforcement Learning (MARL).

[0056] Based on the flow field excitation characteristics predicted in step S102, a multi-agent deep reinforcement learning (MARL) algorithm using the integrated physical constraint learning (PINN) mechanism is used to determine the optimal circuit configuration instructions. In this embodiment, each resonant unit on the acoustic superstructure surface is abstracted as an independent agent, and each agent independently and collaboratively calculates the circuit parameters based on its locally perceived pressure state and global prediction characteristics.

[0057] To cope with the extremely complex and non-uniform sound pressure field inside the turbine, adjacent intelligent agents exchange state information through a communication protocol, utilizing a centralized training-decentralized execution (CTDE) architecture to achieve distributed cooperative hedging in the spatial domain. Simultaneously, to ensure that the instructions generated by the AI ​​algorithm conform to the physical laws of fluid-structure interaction, this embodiment explicitly introduces the residuals of the Navier-Stokes equations and the acoustic wave equations as physical constraints (PINN) into the reward function. This prevents extreme parameters from causing circuit instability due to model output, greatly improving the system's robustness under various extreme conditions. The final output configuration instructions... Including negative capacitors Adjusting the inductance and adjusting resistor .

[0058] S104. Real-time acoustic impedance reconstruction achieved through non-Foster active reactance compensation.

[0059] The execution module adjusts the negative impedance converter (NIC) circuit integrated in the intelligent acoustic superstructure in real time according to the configuration instructions generated in step S103. The negative capacitance is generated using the positive feedback mechanism of the operational amplifier. Its key characteristic lies in actively canceling the intrinsic capacitance inherent in piezoelectric materials. .

[0060] Specifically, through precise control and By tending to cancel each other out and eliminating the capacitive component of the system, the constraints of acoustic causality limitations (Paschen limit) on the physical thickness of the superstructure are broken, enabling even extremely thin superstructures to achieve acoustic impedance matching in the ultra-low frequency band below 100Hz. The adjusted electrical characteristics are coupled into the mechanical structure using the inverse piezoelectric effect, dynamically changing the equivalent mechanical stiffness of the superstructure, so that its acoustic impedance matches the incident sound wave in real time across the entire frequency band, achieving a near-perfect noise cancellation effect.

[0061] S105. Online fine-tuning and self-evolution management based on residual noise feedback:

[0062] The system monitors the residual noise signal after noise reduction in real time through a feedback optimization module. And use this to construct the reward value for reinforcement learning. If the residual noise is significantly reduced, the model is given a positive reward to drive model convergence; otherwise, a penalty logic is triggered to optimize the prediction and decision-making strategy.

[0063] Those skilled in the art should understand that this step is not only to correct the instantaneous deviation of the prediction model in S102, but more importantly, to achieve self-evolutionary management throughout the entire life cycle. When the turbine experiences long-term acoustic characteristic drift due to mechanical wear and component aging after several years of operation, the system will automatically trigger an online incremental learning process to automatically update the neural network weights to compensate for the aging effects, ensuring that the system remains at its optimal noise reduction performance point throughout its entire life cycle, significantly reducing subsequent manual maintenance costs.

[0064] Example 2

[0065] This embodiment elaborates in detail the specific implementation process of active reactance compensation and its reconstructed equivalent mechanical stiffness based on non-Foster circuit (NIC) as described in claims 4 and 8. Its core logic lies in breaking the physical limitations of traditional acoustic control through the negative impedance transformation of electrical characteristics.

[0066] Regarding the physical structure and circuit transformation principle of the non-Foster active reactance compensation unit, the execution module in this embodiment integrates a negative impedance converter (NIC) consisting of a high-performance operational amplifier (Op-Amp) and its feedback loop. Specifically, the NIC circuit mainly consists of an operational amplifier and passive components such as a precision digital potentiometer and a variable capacitor diode connected in its feedback loop. Those skilled in the art will understand that, utilizing the positive feedback mechanism of the operational amplifier, the NIC can transform the load impedance into its opposite value according to a specific ratio, thereby generating a "negative capacitance" that does not exist in nature at the electrical level. And the characteristic of "negative inductance".

[0067] In engineering practice for noise control in hydraulic machinery, traditional piezoelectric shunt damping techniques or passive noise reduction methods typically only adjust the positive inductance and resistance. This limits the system's ability to handle low-frequency noise below 100Hz due to acoustic causality constraints, i.e., the Paschen limit, resulting in extremely narrow sound absorption bandwidth and large physical dimensions, making installation difficult within the compact flow channels of water turbines. This embodiment achieves a qualitative improvement in the performance of piezoelectric materials through the introduction of a NIC circuit. The specific steps are as follows:

[0068] The active cancellation of intrinsic reactance means that piezoelectric ceramic materials, when used as transducers, inherently possess intrinsic capacitance. This manifests as parasitic parameters in traditional circuits, severely limiting the system's energy transfer bandwidth. In this embodiment, the control system, based on decision commands... Precisely adjust the digital potentiometer to achieve the negative capacitance generated by the NIC. In terms of magnitude, it is similar to the intrinsic capacitance of a piezoelectric element. They tend to cancel each other out, that is, they satisfy the total reactance. The physical conditions.

[0069] The dynamic reconstruction of equivalent mechanical stiffness, this electrical-level cancellation effect, is coupled in real time to the mechanical body of the superstructure through the inverse piezoelectric effect, directly altering the mechanical response characteristics of the acoustic superstructure. By eliminating the capacitive component in the system reactance, the system can overcome the Paschen limit's strict limitation on the physical thickness of the superstructure, enabling it to achieve real-time acoustic impedance matching with incident sound waves in the ultra-low frequency band (below 100Hz) even with an extremely thin thickness.

[0070] The multi-parameter real-time adjustment logic executes configuration commands received by the module that include not only negative capacitance parameters. To counteract capacitance, it also includes inductance parameters for precisely adjusting the resonant frequency. and resistance parameters used to optimize sound absorption bandwidth and adjust damping characteristics. .

[0071] Through the aforementioned non-Foster active reactance compensation technology, this embodiment enables the acoustic superstructure within the turbine flow channel to maintain extremely high control accuracy and flexibility under complex, highly turbulent, and non-stationary acoustic field environments. For example, when the turbine's runner blade passing frequency (BPF) drifts due to load fluctuations, the NIC circuit can reconstruct the impedance characteristics according to instructions within milliseconds, ensuring continuous broadband noise suppression. This method not only resolves the physical contradiction between the limited internal space of the turbine flow channel and the effectiveness of low-frequency noise reduction, but also significantly enhances the system's robustness under all operating conditions, laying a solid hardware foundation for achieving adaptive intelligent noise reduction throughout the entire life cycle of the hydroelectric generator set.

[0072] Example 3

[0073] Please refer to Figure 3 As shown in the figure, this embodiment elaborates in detail the specific application logic of the multi-agent deep reinforcement learning (MARL) algorithm and its integrated physical constraint learning (PINN) mechanism described in claims 5, 6 and 7 in the intelligent acoustic impedance matching process of a water turbine.

[0074] S301, Agent-independent modeling and local perception initialization

[0075] In this embodiment, each resonant unit installed on the intelligent acoustic superstructure on the surface of the turbine flow channel is abstracted as an independent intelligent agent. During the initialization phase, each intelligent agent is assigned a unique spatial coordinate identifier and configured with independent computing resources to process local sensing data. Those skilled in the art will understand that the sound field inside a turbine is extremely complex and exhibits non-uniform distribution characteristics, making it difficult for traditional single-point control strategies to achieve a comprehensive noise reduction effect. By decomposing the complex superstructure surface into several collaboratively working intelligent agents, each agent can independently calculate the circuit parameters that meet the acoustic impedance matching requirements of its region based on feedback from local pressure sensors at its location, combined with the global feedforward prediction features output by a spatiotemporal graph neural network (STGNN).

[0076] S302, Distributed Cooperative Control and Space Hedging Based on CTDE Architecture

[0077] To address the challenge of managing the non-uniform sound pressure field induced by strong turbulence during turbine operation, this embodiment employs a centralized training-decentralized execution (CTDE) architecture to achieve distributed collaborative control among multiple agents. In practice, adjacent agents exchange their state information and predicted actions in real time via a pre-defined low-latency communication protocol. During the training phase, the algorithm uses global information to perform a centralized evaluation of the strategy to capture the spatial coupling correlation between resonant units at different locations. During the execution phase, each agent can quickly make decisions based solely on local observations and communication characteristics. This collaborative mechanism effectively performs "cooperative hedging" in the spatial domain, preventing sound power rebound in local areas due to independent adjustments by individual units, and ensuring the global optimality of the system control strategy under complex flow field conditions.

[0078] S303, Embedding and Physical Consistency Guarantee of Physically Constrained Learning (PINN) Mechanism

[0079] To overcome the problem that pure data-driven deep learning algorithms may cause prediction instability or outputs that violate physical laws under extreme conditions, this embodiment explicitly introduces the Physically Constrained Learning (PINN) mechanism into the reinforcement learning framework.

[0080] Residual term introduction: In addition to the conventional control error, the loss function of the neural network also integrates the residual terms of the Navier-Stokes equations describing fluid motion and the residual terms of the sound wave equations describing sound wave propagation.

[0081] Physical Constraints: During the iteration process, the algorithm continuously evaluates whether the circuit parameters output by the current control command conform to the physical logic of fluid-structure interaction. If the agent's output action leads to an increase in the residuals of the physical equations, the system will determine that the action is illegal or high-risk and impose severe penalties. Those skilled in the art will understand that this physical consistency constraint ensures that the intelligent system's generated control logic strictly adheres to physical laws even when facing sudden and severe transient conditions such as turbine startup and load shedding, greatly improving the robustness and traceability of the control strategy.

[0082] S304. Multi-objective optimization reward function design and dynamic strategy iteration

[0083] This embodiment designs a composite reward function that balances noise reduction effectiveness and physical consistency, with its core optimization metric being the reduction in total sound power level in the target frequency band. The specific reward value calculation model is expressed as follows:

[0084]

[0085] In the formula, It represents the change in total sound power level within the target noise reduction frequency band. The lower the value, the better the noise reduction effect, and the larger the corresponding negative bonus value. This is the sum of the residual terms of the aforementioned physical equations; These are preset weighting coefficients used to balance noise reduction performance with the strength of physical constraints.

[0086] By maximizing the composite reward value, the system can drive the model to autonomously converge to the optimal configuration strategy under complex and ever-changing working conditions.

[0087] S305, All-condition Online Fine-tuning and Long-cycle Adaptive Self-evolution

[0088] Based on the aforementioned MARL framework, the system exhibits significant long-term self-evolutionary management capabilities. In actual operation, the system continuously monitors the residual noise signal after noise reduction through edge computing units and calculates reward values ​​in real time to drive policy iteration. Those skilled in the art will understand that, over a service life spanning several years, the acoustic characteristics of a hydro turbine will inevitably drift slowly due to mechanical wear, guide vane corrosion, or component aging.

[0089] This embodiment uses an online incremental learning mechanism to capture these tiny performance degradation paths and automatically update the neural network weights, enabling the control system to have full lifecycle adaptive capabilities and effectively reducing subsequent manual operation and maintenance costs.

[0090] Example 4

[0091] Reference Figure 4As shown, the present invention provides a deep learning-based intelligent acoustic impedance matching system for hydro turbines, which is used to implement the deep learning-based intelligent acoustic impedance matching method for hydro turbines described in embodiments 1-3 above.

[0092] Those skilled in the art will understand that this system aims to construct a closed-loop noise reduction architecture with real-time perception, feature prediction, intelligent decision-making, and precise execution capabilities through the deep integration of hardware modules and software algorithms. The system specifically includes the following modules:

[0093] (1) Perception module

[0094] The sensing module is configured to collect heterogeneous data of the entire turbine flow channel in real time through a distributed multi-source sensor array deployed at key locations in the turbine flow channel.

[0095] Hardware configuration: The hardware of this module includes a dynamic pressure sensor array installed in areas such as the turbine volute inlet, guide vane area, runner chamber, and tailrace pipe to monitor flow field pressure pulsations; it also includes a high-precision tachometer and guide vane opening sensor installed on the turbine main shaft to obtain the real-time operating conditions of the unit; in addition, error microphones (microphone arrays) are distributed downstream of the flow surface in the silencing area to collect residual noise signals.

[0096] Functionality: The sensing module is responsible for preprocessing and synchronizing the acquired heterogeneous data, and constructing a spatiotemporal topology map of the sensors based on the three-dimensional physical coordinates of the sensors, reflecting the spatial correlation of the flow field, and generating data state vectors. The basis for subsequent processing.

[0097] (2) Reasoning Module

[0098] The inference module is embedded with a pre-trained spatiotemporal graph neural network (STGNN) feedforward model, which aims to achieve real-time extraction and advanced prediction of flow field excitation characteristics.

[0099] Computing architecture: This module is typically deployed on edge computing servers or embedded high-performance controllers at hydropower plant sites to meet the low-latency requirements of industrial-grade control.

[0100] Prediction Logic: The inference module receives spatiotemporal topology data output by the perception module, extracts spatial interaction features using a graph convolutional network (GCN), and captures the time dependence of pressure pulsations using a temporal convolutional network (TCN). The key is to predict the dominant frequency, phase, and amplitude of the noise 5-10 ms before it actually reaches the anechoic surface, thus upgrading the system's control logic from traditional "hysteresis compensation" to "real-time prediction."

[0101] (3) Decision Module

[0102] The decision module is configured to run a multi-agent deep reinforcement learning (MARL) algorithm and integrate a physical constraint learning (PINN) mechanism to generate optimal configuration instructions.

[0103] Collaborative Mechanism: The decision-making module treats each resonant unit on the superstructure surface as an independent intelligent agent. Based on the characteristics predicted by the inference module, it responds to the non-uniformly distributed sound pressure field through communication and collaboration between the intelligent agents.

[0104] Command Generation: This module is constrained by physical laws during the decision-making process, ensuring that the generated control commands conform to the fluid-structure interaction (FSI) physical logic. The decision module ultimately outputs the optimal non-Foster circuit (NIC) configuration command. This is used to guide the execution module in performing impedance reconstruction.

[0105] (4) Execution module

[0106] The execution module consists of a negative impedance converter (NIC) circuit and a piezoelectric transducer unit (such as a PZT piezoelectric ceramic sheet) integrated into the intelligent acoustic superstructure body.

[0107] Hardware execution: Each NIC circuit unit corresponds to a piezoelectric transducer node. The module receives digital configuration instructions from the decision module and converts them into high-precision analog control signals.

[0108] Physical effect: The execution module utilizes the positive feedback loop of the operational amplifier to generate negative reactance characteristics, actively canceling the intrinsic capacitance of the piezoelectric material and reconstructing the equivalent mechanical stiffness of the superstructure. Through the inverse piezoelectric effect, the electrical characteristics are coupled to the mechanical structure in real time, enabling the acoustic impedance of the superstructure surface to match the incident sound wave in real time in the ultra-low frequency band (below 100Hz), achieving physical absorption of broadband noise.

[0109] (5) Feedback optimization module

[0110] The feedback optimization module is used to build a closed-loop management system, which drives the online evolution of the model through real-time evaluation of noise reduction performance.

[0111] Performance evaluation: This module monitors the residual noise signal after silencing using an error microphone and converts it into a reward value in the reinforcement learning framework.

[0112] Self-evolution capability: The feedback optimization module incorporates an online incremental learning mechanism. When the system detects mechanical wear, component aging, or long-term acoustic characteristic drift caused by the long-term operation of the turbine, it automatically triggers the model fine-tuning process to update the neural network weights, ensuring that the system maintains optimal noise reduction stability throughout its entire life cycle, effectively reducing manual maintenance costs.

[0113] Those skilled in the art will understand that the system module division described in this embodiment is merely exemplary. In practical applications, the functions of the above modules can be flexibly integrated or split according to the topology of the computing hardware without departing from the technical principles of this invention.

[0114] Example 5

[0115] Reference Figure 5 As shown, this invention provides an electronic device that serves as the core computing and control carrier in the intelligent acoustic impedance matching system for water turbines described in this invention. Those skilled in the art will understand that this electronic device is intended to represent various forms of digital computers, including but not limited to industrial control computers (IPCs), edge computing servers, high-performance workstations, servers, mainframe computers, and other suitable computer combinations.

[0116] Furthermore, this electronic device can also represent various forms of mobile devices or embedded high-performance controllers, such as dedicated artificial intelligence computing chips, smart terminals, field-programmable gate array (FPGA) controllers, and similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0117] Electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded from storage unit 608 into random access memory (RAM) 603. RAM 603 can also store various programs and data required for the operation of electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0118] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as a pressure pulsation sensor, error microphone, tachometer, and keyboard; output unit 607, such as various displays, speakers, and NIC circuit control interfaces; storage unit 608, such as a hard disk or optical disk; and communication unit 609, such as a network interface card, modem, or wireless transceiver. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0119] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips (such as computing cards for accelerating spatiotemporal graph neural network inference), various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as methods For example, in some embodiments, the method It can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 608.

[0120] In this embodiment of the invention, part or all of the computer program may be loaded into and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by computing unit 601, one or more steps of the intelligent acoustic impedance matching method for water turbines described above may be performed. Alternatively, in other embodiments, computing unit 601 may be configured to execute the method by any other suitable means (e.g., by means of firmware). .

[0121] Program code used to implement the methods of the present invention may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowchart to be performed. In the context of this document, a machine-readable medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium, including but not limited to electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.

[0122] Furthermore, the systems and techniques described herein can be implemented in computing systems that include back-end components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include front-end components (e.g., a user computer with a graphical user interface or a web browser), or computing systems that include any combination of such back-end components, middleware components, or front-end components.

[0123] System components can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet. Computer systems can include clients and servers, which are generally geographically separated and typically interact via communication networks.

Claims

1. A deep learning-based intelligent acoustic impedance matching method for hydro turbines, characterized in that, include: Heterogeneous data from the entire flow channel of the water turbine are collected by a multi-source sensing module to construct a sensor topology map of the entire flow channel; The heterogeneous data is input into the feedforward model of the spatiotemporal graph neural network (STGNN) to capture the evolution law of water pressure fluctuations in the flow channel space and their temporal correlation, and to predict the flow field excitation characteristics at future moments. Based on the aforementioned flow field excitation characteristics, the optimal non-Foster circuit configuration instruction is determined using a multi-agent deep reinforcement learning (MARL) algorithm. According to the configuration instructions, the negative impedance transformer (NIC) integrated in the intelligent acoustic superstructure is adjusted, and the equivalent mechanical stiffness of the superstructure is reconstructed in real time through non-Foster active reactance compensation to achieve acoustic impedance matching and noise suppression. Reward values ​​are updated based on residual noise feedback, enabling online fine-tuning and self-evolution management under all operating conditions.

2. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The heterogeneous data acquisition specifically includes: The turbine's rotational speed, guide vane opening, flow channel pressure pulsation, and residual noise are simultaneously collected, and the data state vector is represented as follows: .

3. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The steps for the feedforward prediction performed by the spatiotemporal graphical neural network (STGNN) feedforward model include: The dominant frequency and amplitude of noise are predicted 5-10ms before it reaches the surface of the intelligent acoustic superstructure, thus improving the control logic from hysteresis feedback to real-time prediction.

4. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The non-Foster active reactance compensation specifically includes: By utilizing the feedback loop of an operational amplifier to generate negative capacitance and negative inductance, the inherent intrinsic capacitance of the piezoelectric material is actively canceled out, thereby breaking the Paschen limit and suppressing low-frequency pulsations below 100Hz.

5. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The multi-agent deep reinforcement learning (MARL) algorithm includes: Each resonant unit distributed on the surface of the superstructure is regarded as an independent agent. The units communicate with each other to carry out distributed spatial cooperative counterbalancing control against the non-uniform sound pressure field.

6. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 5, characterized in that, The Multi-Agent Deep Reinforcement Learning (MARL) algorithm employs a Physically Constrained Learning (PINN) mechanism, and its steps include: The residuals of the Navier-Stokes equations and the acoustic wave equations are explicitly introduced into the reward function to ensure that the circuit parameters output by the algorithm conform to the physical fluid-structure interaction law.

7. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 6, characterized in that, The optimization metric for the reward function is based on the reduction in the total sound power level of the target frequency band.

8. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The optimal non-Foster circuit configuration instruction is: in This represents the negative capacitance parameter. and These represent the adjusted inductance and resistance parameters, respectively.

9. The intelligent acoustic impedance matching method for hydro turbines based on deep learning according to claim 1, characterized in that, The reconstructed equivalent mechanical stiffness specifically includes: The execution unit adjusts the NIC circuit according to the configuration instructions, and uses the inverse piezoelectric effect to change the dynamic response characteristics of the superstructure body in real time.

10. A deep learning-based intelligent acoustic impedance matching system for hydro turbines, characterized in that, include: Sensing module: used to collect heterogeneous data from the entire flow path of the water turbine and construct a sensor topology map; Inference module: Embedded spatiotemporal graph neural network (STGNN) for predicting flow field excitation characteristics at future moments; Decision module: Utilizes multi-agent deep reinforcement learning (MARL) to compute the optimal non-Foster circuit configuration instructions; Execution module: includes a negative impedance converter (NIC) circuit for adjusting the equivalent mechanical stiffness of the smart acoustic superstructure according to instructions via non-Foster active reactance compensation; Feedback optimization module: Used to update the reward value of the decision module based on the feedback of residual noise.

11. An electronic device, comprising at least one processor; and a memory communicatively connected to said at least one processor; characterized in that, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.

Images