Tea soup concentration prediction and self-adaptive soup output system based on machine learning

Abstract

The invention relates to the technical field of intelligent control systems, and discloses a tea soup concentration prediction and self-adaptive soup output system based on machine learning. The system comprises a multi-mode sensor array for collecting tea soup sensing signals in real time; the data processing and encoding module encodes the signal into a time sequence pulse sequence; the pulse neural network processing module constructs a storage and calculation integrated hardware circuit based on an analog memristor cross array, and outputs a concentration prediction pulse signal through event-driven calculation; and the self-adaptive soup outlet control module drives the soup outlet execution mechanism to start and stop according to the prediction signal. According to the method, real-time tea soup concentration prediction under milliwatt-level power consumption can be realized, and the problems of high power consumption and large delay when embedded equipment operates a complex model are solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control system technology, specifically to a tea infusion concentration prediction and adaptive infusion system based on machine learning. Background Technology

[0002] The tea beverage industry widely adopts automated equipment to improve brewing accuracy and user experience, with tea concentration monitoring being a core component. Current technologies generally use multimodal sensors to collect physical signals of the tea infusion, such as spectral and impedance data. These signals are then processed by a digital signal processor for feature extraction and pattern recognition, combined with a traditional artificial neural network model to predict concentration. Based on the prediction results, the dispensing mechanism is controlled to start and stop. This solution relies on a separate architecture for the central processing unit and memory, achieving end-to-end control through periodic sampling and batch data processing, and has become the mainstream technology for commercial smart teaware.

[0003] However, in existing technologies, the operation of complex deep learning models leads to a significant increase in device power consumption and prominent response latency, making it difficult to support the continuous and stable operation of miniaturized and portable devices. Summary of the Invention

[0004] This invention provides a tea infusion concentration prediction and adaptive infusion system based on machine learning, which can solve the technical problems of high power consumption and high latency on embedded devices.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a tea infusion concentration prediction and adaptive dispensing system based on machine learning, comprising a multimodal sensor array for real-time acquisition of at least two different types of sensor signals of tea infusion during brewing; a data processing and encoding module connected to the multimodal sensor array for preprocessing and encoding the sensor signals into a time-series pulse sequence; a spiking neural network processing module, the core of which is a memory-computing integrated hardware circuit built on an analog memristor cross array; this module receives the time-series pulse sequence, performs event-driven calculations on the input spatiotemporal pulse pattern through a spiking neural network, and directly outputs a prediction result pulse signal characterizing the real-time concentration or concentration change trend of the tea infusion; and an adaptive dispensing control module connected to the spiking neural network processing module and a dispensing actuator for generating control commands based on the prediction result pulse signal, driving the dispensing actuator to adaptively open or close the dispensing when the predicted concentration reaches a preset threshold; wherein, the spiking neural network processing module realizes local storage of neural network weights and analog domain multiplication and addition operations through an analog memristor cross array, completing the dynamic prediction of tea infusion concentration with ultra-low power consumption.

[0007] In one optional embodiment, the multimodal sensor array includes at least one miniature near-infrared spectral sensor unit and one multi-band impedance sensor unit; the system also includes a sensor fusion and state reconstruction module, which embeds a lightweight computational fluid dynamics-convolutional neural network hybrid model; the hybrid model takes local point data acquired by the front-end sensor as input, simulates the flow field and solute diffusion basic situation inside the pot in real time through the lightweight computational fluid dynamics model, and uses the convolutional neural network as a residual corrector to dynamically learn and compensate for the deviation between the actual physical parameters and the simulation model; finally, it outputs a dynamic cloud map of the tea concentration field in the three-dimensional space inside the pot, as spatial context enhancement information for the time-series pulse sequence received by the spiking neural network processing module.

[0008] In one optional embodiment, the data processing and encoding module employs a sparse encoding strategy inspired by the auditory nerve. This module converts the continuous analog signals from the multimodal sensor array into multi-channel features that conform to the cochlear frequency-position mapping principle. It also draws on the firing adaptive mechanism of auditory neurons to generate and transmit pulses only when the signal feature changes exceed a dynamic threshold, thereby generating a highly sparse and information-dense temporal pulse sequence.

[0009] In one optional embodiment, the sparse coding strategy of the bionic auditory nerve, whose dynamic threshold generation mechanism is implemented by a spiking recurrent neural network module, simulates the feedforward and lateral inhibition network of primary auditory cortex neurons. Through internal spiking activity, it dynamically adjusts the firing threshold of each channel to achieve adaptive focusing on the time-frequency domain salient features in the continuous signal, and simultaneously generates a modulated pulse stream characterizing the overall uncertainty level of the signal, which is fed to the spiking neural network processing module.

[0010] In one alternative embodiment, the sparse coding strategy of the bionic auditory nerve does not have fixed multi-channel frequency-position mapping parameters, but is dynamically configured by an online optimizer based on an evolutionary strategy. The online optimizer uses the accuracy and sparsity of the final concentration prediction of the spiking neural network processing module as a joint reward signal. Over several brewing cycles, it fine-tunes the center frequency and bandwidth of the filter bank of the cochlear mapping in a gradient-free manner, so that the entire coding-processing link co-evolves and adapts to the unique acoustic resonance characteristics of the teaware currently in use and the spectral characteristics of the specific tea varieties.

[0011] In one optional embodiment, the weight update mechanism of the simulated memristor cross array in the spiking neural network processing module is controlled by a meta-learning-driven synaptic plasticity rule. Based on the traditional pulse-time-dependent plasticity rule, a meta-parameter controller is introduced. This controller dynamically adjusts the amplitude and time constant of the learning window of the synaptic plasticity rule of the entire network according to the prior information of the current tea category and the prediction error feedback of the previous brewing cycles, thereby realizing the network's rapid online adaptive fine-tuning of new tea varieties at the hardware level.

[0012] In one alternative embodiment, the meta-parameter controller integrates a digital twin, which is a device-level behavioral model of the physical memristor array in the spiking neural network processing module. This digital twin can simulate the drift of memristor conductance caused by manufacturing deviations, electromigration, and fatigue effects. Based on the simulated state of the digital twin, the meta-parameter controller dynamically compensates for the distortion of network weight representation caused by hardware aging, thereby enabling the synaptic plasticity rules to maintain a stable and effective learning capability throughout the entire hardware lifecycle.

[0013] In one alternative embodiment, the meta-parameter controller regulates the synaptic plasticity rules under the global scheduling of an exploration-exploitation strategy module based on reinforcement learning. This exploration-exploitation strategy module treats each brewing cycle as an exploratory experiment, the action of which is to set a set of meta-parameters, and the reward is a weighted sum of prediction accuracy and energy consumption. Through multiple brewing cycles, the exploration-exploitation strategy module learns a meta-parameter scheduling strategy for the tea varieties most frequently consumed by the current user, thereby achieving an optimal balance between rapid adaptation and stable performance.

[0014] In one optional embodiment, the adaptive brewing control module incorporates a co-optimizer based on nonlinear model predictive control and pulse signal encoding. This co-optimizer decodes the predicted pulse signal output by the spiking neural network processing module into a concentration prediction trajectory within a future time window. Simultaneously, the co-optimizer integrates a simplified brewing kinetics model, using the brewing flow rate and water temperature fine-tuning as control variables, and aims to solve the optimal control problem in the finite time domain online with multiple objectives, including tracking the user's personalized concentration curve, maximizing the extraction efficiency of tea components, and minimizing energy consumption. It then outputs co-control commands to the brewing actuator and heating unit.

[0015] In one alternative embodiment, the multi-objective optimization problem in the collaborative optimizer is solved in real time by a distributed solver inspired by swarm intelligence. The solver runs on multiple parallel biomimetic agents, each representing a possible control trajectory, and follows simplified ant colony optimization or particle swarm optimization rules for pheromone exchange or velocity updates to quickly approach the Pareto optimal frontier. The final execution scheme is then selected from the optimal solution set based on real-time calculated user satisfaction predictions.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. This invention provides a tea infusion concentration prediction and adaptive brewing system based on machine learning. This scheme uses a multimodal sensor array to collect at least two different types of sensor signals from the tea infusion in real time during the brewing process, achieving multi-dimensional perception of the physical properties of the tea infusion and providing fundamental data support for concentration prediction. Based on the sensor signals from the multimodal sensor array, the data processing and encoding module preprocesses and encodes the signals into a time-series pulse sequence. Utilizing the sparse coding mechanism of a biomimetic auditory nerve, pulses are generated only when the signal characteristics change beyond a dynamic threshold, thereby significantly reducing data transmission volume and increasing information density.

[0018] 2. After receiving the time-series pulse sequence, the spiking neural network processing module performs event-driven calculations on the input spatiotemporal pulse pattern through an in-memory computing hardware circuit built on an analog memristor cross array. This module utilizes the analog memristor cross array to achieve local storage of neural network weights and analog domain multiplication and addition operations, avoiding frequent data transfer between the processor and memory, and completely eliminating the von Neumann bottleneck. Based on this, the spiking neural network directly outputs a predicted pulse signal characterizing the real-time concentration or concentration change trend of the tea infusion. Its event-driven characteristic activates the computing unit only when the signal changes, significantly reducing unnecessary energy consumption.

[0019] 3. The adaptive tea infusion control module generates control commands based on the predicted pulse signals, driving the infusion actuator to adaptively open or close the infusion when the predicted concentration reaches a preset threshold, achieving precise closed-loop control. Finally, through a hardware architecture that uses a simulated memristor cross array to perform multiplication and addition operations in the analog domain, the system achieves dynamic prediction of tea infusion concentration with ultra-low power consumption, effectively alleviating the power consumption and latency issues associated with complex models running on embedded devices.

[0020] 4. This invention not only improves computing efficiency through an in-memory computing architecture, but also reduces overall energy consumption by leveraging the sparse event-driven mechanism of spiking neural networks. This enables high-performance concentration prediction technology to adapt to the stringent power consumption constraints of portable and miniaturized tea beverage devices, providing a technical foundation for the miniaturization and long battery life of tea beverage devices. Attached Figure Description

[0021] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0022] Figure 2 This is a schematic diagram of the structure of the multimodal sensor array and sensor fusion module of the present invention;

[0023] Figure 3 This is a schematic diagram of the data processing and encoding module structure of the present invention;

[0024] Figure 4 This is a schematic diagram of the spiking neural network processing module of the present invention;

[0025] Figure 5 This is a schematic diagram of the adaptive soup dispensing control module of the present invention;

[0026] Figure 6 This is a schematic diagram of the meta-parameter controller and reinforcement learning scheduling process of the present invention. Detailed Implementation

[0027] Please refer to Figures 1 to 6 The present invention will now be described in further detail with reference to embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it.

[0028] Example 1: A tea infusion concentration prediction and adaptive dispensing system based on machine learning, comprising: a multimodal sensor array for real-time acquisition of at least two different types of sensor signals of tea infusion during brewing; a data processing and encoding module connected to the multimodal sensor array for preprocessing and encoding the sensor signals into a time-series pulse sequence; a spiking neural network processing module, the core of which is a memory-computing integrated hardware circuit built on an analog memristor cross array; this module is used to receive the time-series pulse sequence, perform event-driven calculations on the input spatiotemporal pulse pattern through a spiking neural network, and directly output a prediction result pulse signal characterizing the real-time concentration or concentration change trend of the tea infusion; and an adaptive dispensing control module connected to the spiking neural network processing module and a dispensing actuator for generating control commands based on the prediction result pulse signal, driving the dispensing actuator to adaptively open or close the dispensing when the predicted concentration reaches a preset threshold; wherein, the spiking neural network processing module realizes local storage of neural network weights and analog domain multiplication and addition operations through an analog memristor cross array to complete the dynamic prediction of tea infusion concentration with ultra-low power consumption.

[0029] A multimodal sensor array is a collection of multiple heterogeneous sensing units used to synchronously acquire information on the physicochemical state of tea infusion. The sensing type is any combination of two or more of the following: optical, electrical, acoustic, or thermal. For example, it is a combination of a near-infrared spectroscopy sensor and an impedance sensor, or a combination of a visible light image sensor and a temperature sensor, or a combination of a pH electrode and a turbidity sensor. Each sensor unit is integrated into the inner wall of the teapot, the embedded part of the lid, or a detachable probe. The installation position is adapted to the actual teaware structure and the brewing flow field distribution.

[0030] The data processing and encoding module is a set of digital circuits and firmware logic units that filter, normalize, time-align, and convert pulses in the raw analog sensing signal. Its preprocessing includes bandpass filtering to suppress power frequency interference, sliding window normalization to eliminate batch drift, and multi-channel timestamp synchronization to ensure spatiotemporal consistency. Its encoding method is based on threshold-triggered address event representation (AER), leakage current integral distribution (LIF) model based on integral ignition mechanism, or multi-scale pulse mapping based on wavelet packet decomposition. The encoding parameters are configured according to the extraction kinetics of different tea varieties and the sensor signal-to-noise ratio.

[0031] The spiking neural network processing module is an event-driven computational structure composed of neuronal nodes and synaptic connections. Its neuron model is a biologically interpretable model such as LIF, Izhikevich, or AdEx. The core hardware carrier of this module is a simulated memristor cross array, whose memristor units are oxygen vacancy migration devices based on a bilayer structure or conductive filament devices based on Ag / GeS thin films. The memristor conductance is continuously adjustable in the range of 0.1μS to 100μS, corresponding to the simulated expression of synaptic weights. The cross array size is set to 32×32, 64×64, or 128×128 according to system accuracy and power consumption constraints. The array is equipped with a low-noise analog front-end and pulse readout circuit. This module performs simulated domain vector-matrix multiplication directly on the memristor array and completes pulse integration and firing in combination with neuronal circuits. The weight update mechanism of the memristor array supports online synaptic plasticity rules or adopts a fixed weight deployment method.

[0032] The adaptive soup dispensing control module is an embedded control subsystem comprising a pulse decoding unit, a decision logic unit, and an execution drive interface. Its pulse decoding unit converts the predicted pulse signal into resolvable semantics such as concentration confidence level, change slope, or threshold time estimation. Its decision logic unit generates Boolean start / stop commands based on a preset concentration threshold, time safety margin, and user preference curve. Its execution drive interface is compatible with various soup dispensing actuators, including DC motor valves, shape memory alloy (SMA) actuators, and piezoelectric ceramic micropumps. The response delay of the soup dispensing actuator is 10ms to 500ms. The control module and the actuator are connected via PWM signals, I²C bus, or a dedicated pulse trigger line.

[0033] The system constructs a full-link pulse-native architecture for the dynamic extraction process of tea infusion, encompassing sensing, encoding, computation, and control: multimodal sensing captures multidimensional physical signals during the non-steady-state brewing process; biomimetic pulse encoding preserves key temporal features and compresses redundant information; memristor-based in-memory computing hardware supports an event-driven SNN, enabling low-power, high-concurrency spatiotemporal pattern recognition in the analog domain; and pulse semantics directly drives the closed-loop control, achieving a millisecond-level response closed loop from concentration perception to tea infusion dispensing.

[0034] The system's working process and principle are as follows: After brewing begins, a multimodal sensor array continuously collects at least two types of signals from the tea liquor, including optical absorption, conductivity, and temperature. The data processing and encoding module performs synchronous preprocessing on the signals from each channel and generates a sparse, asynchronous time-series pulse sequence with clear spatiotemporal semantics based on its dynamic change characteristics. This sequence is then fed into a spiking neural network processing module, where a memristor cross array performs synaptic weighted summation in the analog domain, followed by integration and dissemination by a neuron circuit, ultimately outputting a cluster of pulses representing the current concentration level or its rising / falling trend. The adaptive dispensing control module performs semantic decoding on this pulse cluster, determines whether it meets the preset concentration threshold condition, and immediately issues a start / stop command to the dispensing actuator when the condition is met. The entire process does not convert the pulse signal back into a traditional numerical tensor, avoiding the energy consumption and delay overhead caused by intermediate representation in the digital domain.

[0035] As an optional embodiment, the system is implemented as follows: During a single oolong tea brewing process, a miniature near-infrared spectral sensor monitors the absorbance attenuation in the 500–900nm band in real time, while a multi-band impedance sensing unit synchronously acquires the impedance phase shift in the 1kHz–1MHz band. After filtering and normalization by the data processing and encoding module, the two signals are mapped to a 4-channel frequency-encoded pulse stream and an 8-channel phase-encoded pulse stream, respectively. This 12-channel timing pulse sequence is input into a three-layer SNN constructed by a 64×64-scale memristor array. The network completes a forward inference within 20ms and outputs a concentration-compliant pulse cluster with a peak appearing at 15ms. After the adaptive dispensing control module identifies this pulse cluster, it drives the miniature DC motor valve to open the dispensing port within 25ms, achieving a high degree of synchronization between the dispensing action and the concentration inflection point.

[0036] The system achieves the following beneficial effects: It employs a multimodal sensor array to simultaneously acquire at least two types of physical signals from the tea infusion, overcoming the limitations of single-sensor methods that are susceptible to the influence of bubbles, suspended matter, or temperature disturbances, thus improving the robustness of concentration characterization; the data processing and encoding module converts continuous analog signals into sparse temporal pulse sequences, significantly reducing the data throughput and bandwidth requirements of subsequent calculations; the pulse neural network processing module implements an in-memory computing architecture based on an analog memristor cross array, eliminating the energy loss caused by data transfer in the traditional von Neumann architecture, keeping the power consumption for dynamic concentration prediction below 5mW; the adaptive dispensing control module directly analyzes the predicted pulse signal and drives the actuator, eliminating the intermediate steps of numerical decoding and logical judgment required by the MCU in traditional solutions, compressing the end-to-end response delay to the order of 50ms; the system solves the technical problem that existing tea infusion concentration detection methods rely on manual experience or high-power, complex models, making it difficult to achieve real-time, accurate, low-power automatic dispensing control, enabling the high-performance adaptive dispensing function to be deployed in resource-constrained devices such as portable travel tea sets, small office water dispensers, and smart teacups.

[0037] Example 2: In one embodiment, the multimodal sensor array includes at least one miniature near-infrared spectral sensor unit and one multi-band impedance sensor unit; the system also includes a sensor fusion and state reconstruction module, which embeds a lightweight computational fluid dynamics-convolutional neural network hybrid model; the hybrid model takes local point data acquired by the front-end sensor as input, simulates the flow field and solute diffusion basic situation inside the pot in real time through the lightweight computational fluid dynamics model, and uses the convolutional neural network as a residual corrector to dynamically learn and compensate for the deviation between the actual physical parameters and the simulation model; finally, it outputs a dynamic cloud map of the tea concentration field in the three-dimensional space inside the pot, as spatial context enhancement information for the time-series pulse sequence received by the spiking neural network processing module.

[0038] The miniature near-infrared spectral sensor unit refers to a miniaturized MEMS near-infrared spectral detection device integrated inside the lid or wall of a teapot. Its spectral response range is 700nm–1100nm, and it is used to non-invasively collect the absorption spectral characteristics of tea infusion in multiple narrow bands. The unit is 3mm×3mm×1.2mm in size, which is a miniature packaging form suitable for small teaware structures. Its sampling frequency is 1Hz–10Hz, which can be dynamically adjusted according to the brewing stage to balance the signal-to-noise ratio and time resolution.

[0039] A multi-band impedance sensing unit can refer to a contact impedance measurement module composed of two or more microelectrode arrays. Each set of electrodes operates at different excitation frequencies (1kHz, 10kHz, 100kHz, 1MHz) to synchronously acquire the changes in conductivity and dielectric constant of tea soup at different frequency bands. The unit is attached to the inner surface of the bottom of the pot or distributed circumferentially along the pot wall. The electrode material is gold, platinum or conductive polymer, and its geometric configuration is a ring, interdigitated or coplanar microstrip structure. The specific shape and size are set according to the internal spatial layout of the teaware and the signal penetration depth requirements.

[0040] The sensor fusion and state reconstruction module is a hardware-software co-processing subsystem deployed on the edge computing unit. Its hardware carrier is a low-power ARM Cortex-M7 microcontroller or a RISC-V architecture AI-accelerated SoC, and the software part runs a lightweight computational fluid dynamics-convolutional neural network hybrid model. This module is connected to the multimodal sensor array through the SPI / I²C interface and receives the conditioned raw spectral absorbance values ​​and multi-band complex impedance amplitude / phase data. Its output is a three-dimensional spatial gridded concentration field with a spatial resolution of 8×8×4 (axial × radial × vertical) and a time update period of 2s–5s, which can be adaptively adjusted according to the brewing process.

[0041] The lightweight computational fluid dynamics model is a simplified solver built on the incompressible Navier-Stokes equations and convection-diffusion equations. It is discretized using the finite volume method, and the number of grids is reduced to ≤5,000 control volumes while ensuring physical consistency. A steady-state initialization + transient correction strategy is used to reduce the computational cost per step. The input parameters of the model include: current water temperature, estimated tea bulk density, liquid level in the pot, and opening of the infusion spout. Some parameters are provided by the system's built-in temperature sensor and mechanical position encoder, while others are set empirically based on historical brewing data. Its outputs are the flow velocity field, pressure field, and initial solute concentration distribution field in the pot.

[0042] The convolutional neural network (CNN) serves as the residual corrector, employing a lightweight U-Net variant structure with a depth of 4 layers and ≤32 channels. Its input is the interpolation error map between the computational fluid dynamics-predicted concentration at its corresponding spatial location and the concentration measured at the multimodal sensor point. The output is the concentration compensation amount per grid. The weights of this CNN model are stored in on-chip SRAM using INT8 quantization format. The inference process is accelerated by dedicated SIMD instructions, with a single forward computation latency of less than 15ms. Its training data originates from high-fidelity computational fluid dynamics simulations and joint calibration experiments of real tea infusion impedance / spectroscopy, covering the diffusion behavior of three typical tea types—green tea, black tea, and oolong tea—under different water temperatures, tea quantities, and steeping times.

[0043] The three-dimensional tea infusion concentration field dynamic cloud map is a spatial data structure expressed in VTK or binary voxel array format. Each voxel stores a normalized concentration value and is mapped to the internal geometry of the physical teapot through a standardized coordinate system. After bilinear interpolation, the cloud map generates a spatial weighting coefficient matrix corresponding to the input node of the spiking neural network. This matrix is ​​used to perform topological sensing weighted fusion of temporal pulse sequences from sensor channels at different spatial locations, thereby achieving spatial context enhancement of the pulse input.

[0044] After receiving the local absorbance variation trend provided by the miniature near-infrared spectral sensor unit and the local conductivity gradient provided by the multi-band impedance sensor unit, the lightweight computational fluid dynamics model first deduces the solute transport path dominated by macroscopic convection within the vessel based on mass conservation and Fick's second law. Subsequently, the convolutional neural network residual corrector receives the difference image between the initial concentration field output by the computational fluid dynamics and the reference concentration field obtained by interpolation of the sensor's measured points, and extracts the spatial local non-uniformity features through the convolution kernel to generate a voxel-by-voxel compensation. The two are superimposed to form a dynamically updated three-dimensional concentration field cloud map. This cloud map is divided into several spatial regions, each corresponding to a set of input neurons in the spiking neural network processing module. The pulse firing intensity of each neuron is modulated according to the concentration gradient amplitude and rate of change in the region, thereby explicitly encoding the spatial distribution information into the temporal-spatial joint representation of the subsequent pulse sequence.

[0045] As an optional embodiment, the specific implementation of the present invention is as follows: During a single brewing of oolong tea, a miniature near-infrared spectral sensor unit collects the absorbance ratio of the tea soup in the central area below the teapot lid at wavelengths of 850nm and 940nm every 2 seconds; a multi-band impedance sensing unit simultaneously collects the impedance phase difference of the ring electrode pair at the bottom of the teapot under excitation at 10kHz and 100kHz; the two signals are sent to the sensor fusion and state reconstruction module after analog-to-digital conversion; a lightweight computational fluid dynamics model is used to calculate the initial concentration based on the current water temperature of 92℃, the estimated tea bulk density of 0.35g / cm³, and the liquid level of 65mm. The field shows that the solute mainly rises along the pot wall and forms an enrichment area near the lid. The convolutional neural network residual corrector compares the computational fluid dynamics prediction with the interpolation results of the measured points, identifies the concentration underestimation caused by vortex disturbances in the middle of the pot that were not modeled by computational fluid dynamics, and generates corresponding compensation. The fused three-dimensional concentration field cloud map shows that the concentration peak is located in the area below the lid and slightly to the right, and the vertical gradient is significant. This cloud map is mapped to 8 sets of spatial weights, which respectively control the firing threshold of 8 pulse coding channels, so that the sensor signals from the high gradient region generate more dense temporal pulses, thereby improving the response sensitivity of the pulse neural network to key dynamic regions.

[0046] Through the above technical solutions, this invention achieves the following: By introducing a miniature near-infrared spectral sensor unit and a multi-band impedance sensing unit to form a multimodal point-type sensing front end, it can simultaneously capture the optical absorption characteristics and electrical conduction characteristics of tea soup, overcoming the one-sidedness of single-mode response to component changes; By constructing a lightweight computational fluid dynamics-convolutional neural network hybrid model and outputting a three-dimensional concentration field dynamic cloud map, it can upgrade local measurements to spatial continuous distribution estimation, making up for the inherent defect of point sensors lacking spatial coverage; By feeding this cloud map as spatial context enhancement information into the spiking neural network processing module, it can guide the SNN to explicitly focus on regions with significant concentration gradients during the pulse coding and processing stages, improving its ability to model non-uniform diffusion processes; Finally, in actual brewing scenarios with severe water flow disturbances and uneven tea leaf unfurling, the system's prediction error for the overall concentration trend of tea soup is reduced by 42% compared to the single-sensor solution, and the response delay is shortened to within 3.1 seconds.

[0047] Example 3: In another optional embodiment, the present invention also provides a data processing and encoding module that adopts a sparse encoding strategy based on bionic auditory nerves; the module converts the continuous analog signals from the multimodal sensor array into multi-channel features that conform to the cochlear frequency-position mapping principle, and draws on the firing adaptive mechanism of auditory neurons to generate and transmit pulses only when the signal feature changes exceed the dynamic threshold, thereby generating a highly sparse and information-dense temporal pulse sequence.

[0048] Among them, the sparse coding strategy of bionic auditory nerve can refer to a signal coding method inspired by the mammalian auditory pathway. Its core lies in simulating the spatial decomposition characteristics of the cochlear basilar membrane in response to the sound wave spectrum and the event-driven response mechanism of primary auditory neurons. This strategy does not rely on a fixed sampling rate or uniform quantization precision, but achieves time-frequency joint characterization of multimodal sensing signals in a biologically interpretable way.

[0049] The cochlear frequency-position mapping principle involves decomposing the input continuous analog signal in the frequency domain through a set of bandpass filters. The distribution of the center frequency and bandwidth of each filter follows a logarithmic linear relationship described by the Becksy curve or Greenwood function, resulting in high resolution in the low-frequency band and wide coverage in the high-frequency band. The parameters of the filter set can be adapted according to the optical absorption spectrum of tea infusion, the impedance response frequency band, and the noise characteristics of the sensor. For example, it can be a 24-channel filter set with a center frequency range covering 10Hz–10kHz, or a reconfigurable filter structure that dynamically adjusts the number and distribution density of channels according to the actual brewing scenario. This embodiment of the invention does not impose any special limitations on this.

[0050] The multi-channel feature is at least one of the envelope signal, instantaneous phase shift, zero crossover rate, or short-time energy statistics output by each filter channel; each channel feature can be independently normalized and nonlinearly compressed to match the input dynamic range of the subsequent spiking neuron; its specific form can be set according to the sensor type and signal-to-noise ratio, for example, focusing on extracting the absorption peak shift response for near-infrared spectral signals, and focusing on extracting the phase hysteresis change for impedance signals. This embodiment of the invention does not impose any special limitations on this.

[0051] The adaptive firing mechanism of auditory neurons can refer to configuring a dynamically adjustable pulse firing threshold unit for each channel. This unit is estimated in real time based on the first / second difference of the local signal, the standard deviation of the sliding window, or the modulus maxima of wavelet coefficients, and is automatically updated according to the background noise level and signal activity. This mechanism can avoid false triggering of the constant threshold under low signal-to-noise ratio or missed triggering under strong steady-state signals. For example, the threshold can be updated by using an exponentially weighted moving average (EWMA) method, or by using a closed-loop adjustment method based on pulse feedback. This embodiment of the invention does not make any special limitations on this.

[0052] The dynamic threshold is a scalar or vector that changes over time, and its update cycle can be set according to the brewing stage: a more lenient threshold is set at the beginning of water injection to capture rapid dissolution response, and a more sensitive threshold is set during the stable immersion period to identify weak concentration gradient changes; the update rule of this threshold can be implemented by a hardware state machine or executed by a lightweight digital controller, and the specific implementation method can be selected according to the system power consumption budget and real-time requirements. This embodiment of the invention does not impose any special limitations on this.

[0053] A highly sparse, information-dense temporal pulse sequence is a discrete event stream in which each channel generates only 0–3 pulses per unit time. The pulse moments encode signal abrupt change points, and the pulse polarity or weight encodes the direction and amplitude of change. The temporal resolution of this sequence can reach the millisecond level, but the overall pulse density is less than 10% of that of traditional AER encoding, thereby significantly reducing on-chip communication bandwidth and activation energy consumption of downstream SNN modules. Its sparsity can be adjusted online according to the differences in tea type, water temperature and teapot material. For example, the pulse density can be increased for green tea with a high extraction rate, and further compressed for slow-release Pu'er tea. This embodiment of the invention does not impose any special limitations on this.

[0054] Specifically, during operation, this sparse coding strategy first synchronously sends the 700–950nm band transmitted light intensity signal acquired by the miniature near-infrared spectral sensor and the 1kHz–1MHz impedance phase signal acquired by the multi-band impedance sensing unit into the filter bank. After bandpass filtering, the envelope of each channel is extracted and the local variance is calculated to drive the dynamic threshold unit of the corresponding channel. When the change in the envelope of a certain channel exceeds the current threshold for two consecutive frames, a positive pulse is triggered, whose address encodes the channel number and the release timestamp is accurate to 10μs. All channel pulses are merged into a single time-series pulse stream in chronological order, which serves as the final output of the data processing and coding module and is fed to the spiking neural network processing module.

[0055] As an optional embodiment, the specific implementation of the present invention is as follows: During the brewing process of Longjing tea, within the first 30 seconds after water is poured, the near-infrared channel detects a strong absorption jump at 810nm, while the impedance channel exhibits a rapid negative phase bias in the 100kHz frequency band; accordingly, the encoding module generates two pulses with a time interval of 12ms in the corresponding channels, while the other channels remain silent because they have not reached the dynamic threshold; by the 45th second, the background noise of each channel increases, causing the threshold to automatically rise by 15%, suppressing the false pulses caused by steady-state drift; a total of 27 pulses are output in the entire initial dissolution stage, which reduces the amount of data by 98.3% compared with the traditional encoding method with an equivalent sampling rate of 1kHz, but fully retains the discriminative features such as the key dissolution start point, peak time, and attenuation slope.

[0056] Through the above technical solutions, this invention achieves the following: By employing a multi-channel filtering structure conforming to the cochlear frequency-position mapping principle, it can perform biologically plausible spectral decomposition of multimodal sensing signals, improving the physical interpretability of feature representation; by introducing an adaptive firing mechanism of auditory neurons and cooperating with a dynamic threshold unit, it can automatically adjust the pulse trigger sensitivity based on the real-time signal-to-noise ratio and signal activity, avoiding static noise interference and enhancing the ability to identify subtle concentration changes; and by generating pulses only when signal feature changes exceed the dynamic threshold, the generated highly sparse pulse sequence significantly reduces the data transmission load and the computational overhead of the subsequent spiking neural network, further reducing the overall power consumption of the system while ensuring prediction accuracy.

[0057] Example 4: In one embodiment, the present invention also provides a sparse coding strategy for bionic auditory nerves, the dynamic threshold generation mechanism of which is implemented by a pulse recurrent neural network module; the pulse recurrent neural network module simulates the feedforward and lateral inhibition network of primary auditory cortex neurons, dynamically adjusts the firing threshold of each channel through internal pulse activity, realizes adaptive focusing on the time-frequency domain significant features in continuous signals, and simultaneously generates a regulating pulse flow characterizing the overall uncertainty level of the signal, which is fed to the pulse neural network processing module.

[0058] The spiking recurrent neural network module is a recursive connection structure composed of multiple spiking neurons. Its input is the original pulse sequence from each channel of the data processing and encoding module, and its output is the dynamic firing threshold adjustment signal and the global uncertainty control pulse flow of the corresponding channel. The neuron connection mode of this module is a hybrid topology of feedforward connection and lateral inhibition connection. The feedforward connection is used to transmit the temporal characteristic response, and the lateral inhibition connection is used to apply competitive inhibition between adjacent frequency channels, thereby achieving selective enhancement of the energy concentration region and compression of the diffuse background response. The synaptic weights of the feedforward and lateral inhibition networks can be set to fixed values ​​or programmable parameters according to the actual hardware implementation requirements. For example, they can be analog weights based on memristor array configuration, or quantized weights loaded by digital controller. This embodiment of the invention does not make any special limitation on this.

[0059] The firing mechanism of each neuron in the pulsed recurrent neural network module is based on the membrane potential integral-firing-reset model. Its threshold is not constant, but evolves dynamically over time: when a channel continuously receives high-density pulse input, the membrane potential of its corresponding neuron rises faster, and at the same time, the excitability of neurons in neighboring channels is reduced through lateral inhibition connections, so that the firing behavior gradually converges to the dominant frequency band. This process simulates the cocktail party effect-like attention selection mechanism in the primary auditory cortex of organisms, thereby automatically focusing on time-frequency domain significant features such as acoustic resonance peaks, spectral absorption transition points or impedance phase inflection points that are strongly correlated with changes in tea concentration in multimodal sensing signals.

[0060] The process of dynamically adjusting the firing threshold of each channel is as follows: the neuron unit corresponding to each channel receives the historical pulse count of this channel, the pulse intensity difference of neighboring channels, and the global pulse synchronization index as state input in real time, and updates its dynamic threshold variable through the internal pulse-driven differential equation; the dynamic threshold variable is a one-dimensional real value, or a two-dimensional parameter set containing amplitude and time constant, and its update rule follows the plasticity form of exponential moving average, pulse-triggered step, or synaptic plasticity rule. The specific form is adapted according to the chip resource constraints and accuracy requirements. For example, it can be a lightweight nonlinear mapping based on on-chip LUT lookup table, or a continuous-time dynamic response implemented by small-scale analog circuits. The embodiments of the present invention do not make special limitations on this.

[0061] The synchronously generated control pulse stream, which characterizes the overall uncertainty level of the signal, is a global control signal independent of the characteristic pulses of each channel. Its pulse firing frequency is positively correlated with signal entropy, cross-channel pulse asynchrony, or residual fluctuation amplitude. This control pulse stream is fed to the spiking neural network processing module in an event-driven manner to adjust the gain coefficient or learning rate parameter of the critical path in its core memristor array, thereby reducing the prediction confidence weight during periods of high uncertainty and enhancing the characteristic response strength during periods of low uncertainty. The encoding method of this control pulse stream is frequency encoding, time interval encoding, or pulse group encoding, and its decoding method is integral window counting, delay matching detection, or pulse correlation analysis. The specific implementation is determined according to the downstream module interface protocol, and this embodiment of the invention does not impose any special limitations on this.

[0062] Specifically, during operation, the spiking recurrent neural network module first receives a multi-channel pulse sequence from the multimodal sensor array after preliminary encoding. Subsequently, each channel neuron dynamically updates its firing threshold and decides whether to generate an output pulse based on the current input pulse density and the suppression signals from other channels. During this process, if a high-frequency channel detects a rapidly enhancing pulse cluster with an absorption peak near 520 nm in the near-infrared spectrum, it will enhance its own response through a feedforward path, while suppressing the slowly drifting baseline pulse in the low-frequency impedance signal through a lateral suppression path. At the same time, the global state unit inside the module counts the dispersion and synchronization deviation of pulse firing in all channels. When inconsistencies in the responses of multiple sources are detected (such as a spectral abrupt change with no impedance response), the firing frequency of the uncertain pulse stream is increased. This controlled pulse stream is then sent to the spiking neural network processing module, triggering it to add a confidence attenuation factor to the current prediction result to avoid making premature exit decisions before the physical state is fully stable.

[0063] As an optional embodiment, the present invention is implemented as follows: During the brewing of a high-roasted oolong tea, a miniature near-infrared spectral sensor continuously captures the reflectivity change pulses of the tea liquor in the 650–900 nm band for 120 seconds, and a multi-band impedance sensing unit simultaneously outputs a phase angle pulse sequence in the 1 kHz–1 MHz band; the data processing and encoding module maps the above two signals into an 8-channel cochlear frequency band pulse stream; after receiving the 8-channel input, the pulse recurrent neural network module identifies the 4th–5th channel (corresponding to the 720–780 nm optical absorption). A strong synchronous pulse cluster appears at the impedance phase inflection point of 300–500kHz, while channels 1 and 2 exhibit low-frequency diffuse pulses. Based on this, the module increases the emission threshold sensitivity of channels 4–5, suppresses the response of channels 1 and 2, and generates a high-frequency modulation pulse stream. This modulation pulse stream causes the pulse neural network processing module to add a ±15% confidence interval mark to the concentration prediction result in the following 10 seconds, thereby delaying the triggering timing of the adaptive soup dispensing control module until the frequency of the uncertainty pulse stream falls back below the reference level, ensuring that the soup dispensing action occurs after the concentration trend is confirmed to be stable.

[0064] Through the above technical solutions, the present invention achieves the following: Since the pulse recurrent neural network module dynamically adjusts the emission threshold of each channel through feedforward and lateral inhibition networks, it can automatically focus on the time-frequency domain significant region that is strongly correlated with the change in tea concentration according to the real-time signal feature distribution, thereby improving the accuracy of key feature extraction; Since the module synchronously generates a control pulse stream that characterizes the overall uncertainty level of the signal and feeds it to the downstream SNN processing module, it can actively reduce the prediction confidence in the stage of signal state ambiguity, avoiding premature or delayed tea dispensing due to misjudgment; Since the entire dynamic threshold generation mechanism is entirely based on pulse event driving and does not require periodic sampling and floating-point operations, it works in conjunction with the bionic auditory nerve sparse coding strategy to form an end-to-end low-power signal representation link, supporting the system to complete highly robust concentration sensing at milliwatt-level power consumption.

[0065] Example 5: In another optional embodiment, the present invention also provides a sparse coding strategy for the bionic auditory nerve, wherein the multi-channel frequency-position mapping parameters are not fixed, but dynamically configured by an online optimizer based on an evolutionary strategy; the co-optimizer uses the accuracy and sparsity of the final concentration prediction of the spiking neural network processing module as a joint reward signal, and fine-tunes the center frequency and bandwidth of the cochlear mapping filter bank in a gradient-free manner over several brewing cycles, so that the entire coding-processing link co-evolves, adapting to the unique acoustic resonance characteristics of the currently used teaware and the spectral characteristics of the specific tea varieties, including:

[0066] Among them, the online optimizer based on the evolutionary strategy is a black-box optimizer that does not rely on gradient information and updates parameters through random sampling and population evolution mechanism. Its structure includes a population initialization module, a fitness evaluation module, a selection-mutation-crossover operation module, and a parameter update module. After each round of bubbling, the co-optimizer receives the prediction error index (e.g., root mean square error RMSE) and the pulse sequence sparsity index (e.g., average pulse firing rate per unit time or L0 norm normalized value) fed back from the spiking neural network processing module, and combines them in a weighted manner to form a joint reward signal. This joint reward signal is used to evaluate the merits of the current filter bank parameter configuration, thereby driving the population to evolve towards the high reward region.

[0067] The center frequency of the cochlear mapping filter bank is any frequency point in the range of 20Hz to 20kHz. Its specific value is set according to the dominant time-frequency response characteristics of different tea extract components in near-infrared and impedance signals. For example, it is a high-frequency bias group (center frequency covering 1kHz–8kHz) configured for the fast release characteristics of green tea, or a low-frequency enhancement group (center frequency covering 100Hz–2kHz) configured for the slow release characteristics of ripe Pu'er tea. The embodiments of the present invention do not make any special limitations on this.

[0068] The filter bank bandwidth is 5%–30% of the corresponding center frequency. The specific value is set according to the dynamic response rate of the changes in physical parameters of the tea soup during the brewing process. For example, in a Zisha teapot, where the heat conduction and solute diffusion are slow due to the wall thickness and porosity, a narrower bandwidth is used to enhance the time domain resolution. In a glass teapot, where the thermal response is rapid and the concentration gradient changes drastically, a wider bandwidth is used to improve the frequency domain coverage. This bandwidth parameter is set according to the actual situation, and the embodiments of the present invention do not impose any special limitations on it.

[0069] The population size of the online optimizer is configured according to the memory and computing resource constraints of the embedded device, for example, 8–64 individuals; each round of evolution iteration can be completed during the idle period after a single brewing, and the evolution generation is terminated early based on convergence judgment to avoid redundant calculations; this optimization process does not introduce additional training data or external annotations, but only relies on the prediction feedback and impulse statistics generated by the system itself, thus possessing fully self-supervised characteristics.

[0070] Specifically, the online optimizer uses preset universal cochlear filter bank parameters as the initial population during the first brewing. In each subsequent brewing round, based on the prediction results obtained after inputting the time-series pulse sequence generated in the previous round into the spiking neural network processing module, it calculates the error between the predicted result and the reference concentration value (e.g., the standard curve interpolation result obtained from offline calibration or the calibration point manually confirmed by the user), and simultaneously calculates the firing density and temporal distribution entropy of the pulse sequence. The error term and the sparsity term are weighted and fused to form a fitness function value, which is used to sort the candidate parameter groups in the current population. Then, selection, Gaussian perturbation mutation, and uniform crossover operations are performed to generate a new generation of parameter groups. The updated filter bank parameters are loaded into the data processing and encoding module at the beginning of the next brewing round to achieve a gradual adjustment of the cochlear mapping characteristics.

[0071] As an optional embodiment, the specific implementation of the present invention is as follows: When a user first uses a certain purple clay teapot to brew Dianhong black tea, the system activates the default filter group (the center frequency is distributed at logarithmic intervals between 200Hz and 5kHz, and the bandwidth is 20% of the corresponding center frequency); after five consecutive brewing rounds, the online optimizer observes that the predicted RMSE decreases from the initial 0.82% to 0.37%, while the pulse sparsity increases from 42 pulses per second to 28 pulses per second; the population evolution results show that the optimal parameter combination tends to shift the center frequency of the low-frequency band (200–800Hz) towards 650Hz and narrow the bandwidth to 12%, while the mid-to-high frequency band (2–5kHz) is broadened to 25%, indicating that the purple clay teapot is more sensitive to low-frequency impedance phase changes in the early stage of black tea brewing, and the high-frequency near-infrared absorption fluctuations are more discriminative; the adjustment results are automatically written to the local non-volatile storage area and directly loaded on the next startup, so that the system can maintain a high-precision and low-power operation state without repeated learning in subsequent use.

[0072] Through the above technical solutions, this invention achieves the following: Since the online optimizer uses prediction accuracy and pulse sparsity as joint reward signals, it can balance concentration prediction performance and hardware energy efficiency requirements; Since the collaborative optimizer adopts a gradient-free evolutionary strategy, it can bypass the backpropagation barrier caused by non-differentiable links in the spiking neural network (such as pulse firing function and memristor conductance jump behavior), thus achieving stable deployment on embedded edge devices; Since the optimization process relies on a real brewing cycle closed loop, the obtained filter parameters can effectively reflect the coupling relationship between the current acoustic characteristics of the teaware and the extraction spectrum of the tea leaves, enabling the entire encoding-processing link to have adaptive evolutionary capabilities for specific application scenarios.

[0073] Example 6: In an optional embodiment, the present invention also provides a weight update mechanism for a simulated memristor cross array in the spiking neural network processing module, which is controlled by a meta-learning-driven synaptic plasticity rule. Based on the traditional pulse-time-dependent plasticity rule, this synaptic plasticity rule introduces a meta-parameter controller. The controller dynamically adjusts the amplitude and time constant of the learning window of the synaptic plasticity rule of the entire network according to the prior information of the current tea category and the prediction error feedback of the previous brewing cycles, thereby realizing the network's rapid online adaptive fine-tuning of new tea varieties at the hardware level.

[0074] Meta-learning-driven synaptic plasticity rules are learning rules that extend the standard pulse-time dependent plasticity (synaptic plasticity rules) and possess cross-task generalization capabilities. The core of these rules lies in the fact that the learning dynamics parameters of the synaptic plasticity rules—including the learning window amplitude (i.e., the maximum change in synaptic weight) and the time constant (i.e., the time-sensitive range of causal / non-causal pulse pairs)—are no longer set to fixed values, but are driven in real time by externally adjustable variables. These synaptic plasticity rules provide dynamic scaling factors and time-gated signals for the conductance updates of each memristor cell in the simulated memristor cross array, thereby achieving overall regulation of the synaptic plasticity behavior of the entire network without changing the underlying physical structure.

[0075] The meta-parameter controller is a dedicated digital logic circuit module integrated within the spiking neural network processing module. Its input is connected to the tea category identification signal provided by the system's main control unit and the prediction error pulse stream from the output of the spiking neural network processing module. This controller is an on-chip coprocessor based on a finite state machine or a lightweight RISC-V core, used to execute the meta-parameter generation algorithm. Its output is connected to the analog bias circuit surrounding the memristor array via a low-latency parallel bus to configure the write voltage amplitude and timing control signals of each memristor unit in real time. The hardware implementation of this controller can be set according to the actual situation, for example, it can be an ASIC custom circuit or an FPGA programmable logic unit. This embodiment of the invention does not impose any special limitations on this.

[0076] The prior information of the tea category for brewing is obtained by the user through manual input, QR code recognition, or voice recognition. This includes, but is not limited to, major categories such as green tea, black tea, oolong tea, white tea, dark tea, and yellow tea, as well as their subcategories. This information is a discrete semantic symbol, which is encoded and input to the meta-parameter controller in the form of a multi-bit digital signal. The specific encoding format can be set according to the actual situation, such as one-hot encoding, binary encoding, or embedded vector quantization representation. This embodiment of the invention does not impose any special limitations on this.

[0077] The prediction error feedback in the first few brewing cycles is an error pulse sequence generated by the time alignment error between the prediction result pulse signal output by the pulse neural network processing module and the actual concentration reference value. This error pulse sequence is obtained by closed-loop calibration of the output of the standard concentration detection module (such as a miniature ultraviolet-visible spectral sensor) synchronously collected during the brewing stage. The frequency or amplitude of the error pulses represents the current prediction deviation. The feedback signal is an analog voltage signal, a digital counting signal, or a sparse pulse stream. Its acquisition and encoding method can be set according to the actual situation. This embodiment of the invention does not impose any special limitations on this.

[0078] The dynamic adjustment of the amplitude and time constant of the learning window for the synaptic plasticity rule of the entire network can refer to the following: the meta-parameter controller obtains the initial adjustment coefficient by looking up a table based on the prior information of tea category, and performs proportional-integral online correction in combination with real-time error feedback to generate a set of global control parameters; after digital-to-analog conversion, this set of parameters is applied to the write drive circuit of the memristor array, so that the change in conductance (amplitude) under the triggering of the same synaptic plasticity rule event is synchronously scaled with the width of the allowed triggering time window (time constant); the scaling relationship is a linear mapping or a piecewise nonlinear function, and its specific form is jointly calibrated according to the characteristics of the memristor device and the tea precipitation kinetic model. This embodiment of the invention does not impose any special limitations on this.

[0079] Specifically, when the system first identifies a new category of tea (such as ripe Pu'er tea), the meta-parameter controller presets a wider learning time constant and a higher amplitude based on the prior information of that category, enhancing the network's ability to capture long-term concentration evolution trends. Subsequently, during several rounds of brewing, if the prediction error is detected to be continuously decreasing, the time constant is gradually tightened and the amplitude is reduced, allowing the network to converge to a stable state. If the error rebounds, the control parameters are relaxed in reverse, restarting the exploration process. This closed-loop adjustment process is completed entirely at the hardware circuit level, without interrupting the spiking neural network inference process, and without relying on external host computer participation in the calculation.

[0080] As an optional embodiment, the specific implementation of the present invention is as follows: Before the start of a brewing cycle, the user selects the Pu'er ripe tea category via the device's touchscreen. The system sends the corresponding category ID to the meta-parameter controller in 4-bit binary code form; the controller looks up the initial parameter group: synaptic plasticity rule enhancement time constant. =120ms, suppression time constant =180ms, enhanced amplitude =0.15μS, suppression amplitude =−0.12μS; During the brewing process, the pulse neural network processing module continuously outputs concentration prediction pulses. Simultaneously, the miniature UV-Vis spectral sensor collects the actual absorbance value every 5 seconds at the dispensing spout and converts it into an error pulse stream input to the controller. The controller updates the control parameters every 30 seconds. After the first round, it will... The time was reduced to 110ms, and then further reduced to 95ms in the second round. The error gradually decreased from 0.15 μS to 0.09 μS; by the fourth brewing round, the prediction error stabilized within ±2%, the parameter entered a locked state, and the system completed the online adaptive fine-tuning of the tea variety.

[0081] Through the above technical solutions, this invention achieves the following: by introducing a meta-parameter controller to dynamically adjust the amplitude and time constant of the synaptic plasticity rule learning window, the spiking neural network processing module's response sensitivity to the extraction characteristics of different tea varieties is improved; because the adjustment basis includes prior information on tea categories and feedback on historical prediction errors, the problem of slow convergence or overfitting of traditional synaptic plasticity rules when facing new tea varieties is avoided; and because all adjustment logic is implemented in a closed loop at the hardware level, the parallelism and deterministic latency of the adaptive fine-tuning process and the real-time concentration prediction task are guaranteed, meeting the dual constraints of low power consumption and high real-time performance of embedded tea drinking devices.

[0082] Example 7: In one possible implementation, the present invention also provides a system in which a digital twin is integrated inside a meta-parameter controller. The digital twin is a device-level behavioral model of the physical memristor array in the spiking neural network processing module, capable of simulating the drift of memristor conductance caused by manufacturing deviations, electromigration, and fatigue effects. The meta-parameter controller dynamically compensates for the distortion of network weight representation caused by hardware aging based on the simulated state of the digital twin, thereby enabling the synaptic plasticity rule to maintain a stable and effective learning ability throughout the entire hardware lifecycle.

[0083] A digital twin can refer to a lightweight physical simulation model of a device embedded within a meta-parameter controller. Its inputs include the operating temperature of the memristor array, the cumulative number of pulses, the historical sequence of bias voltages, and the initial manufacturing parameter distribution. The output is the predicted deviation of the conductance value of each memristor cell at the current moment. This model is a reduced-order model constructed based on the finite element method or equivalent circuit modeling methods, such as using a dual-interface ion migration model or filamentary evolution equations, or employing a data-driven neural differential equation approximator. This embodiment of the invention does not impose any special limitations on this. The simulation accuracy of the digital twin depends on the actual chip package thermal characteristics and the memristor material system (e.g., ...). , Adaptive configurations are made for Ag-Ge-Se and other process nodes (such as 28nm or 40nm CMOS compatible processes).

[0084] The meta-parameter controller is a mixed-signal control unit integrated on the same chip. Its digital part performs state updates and error mapping calculations for the digital twin, while its analog part receives reference readout signals from the memristor array and completes closed-loop feedback. Based on the conductance drift prediction value output by the digital twin, the controller adjusts the learning rate gain coefficient in the synaptic plasticity rule. Time window width Alternatively, a proportional-integral compensation term can be applied to the synaptic weight update direction, which can be expressed as follows:

[0085] ;

[0086] in The learning rate gain coefficient after compensation. Represents the base learning rate gain coefficient. The error value predicted by the twin at the current time t is the weighted bias normalized error. For the integration variable The error value at the corresponding time point, and The compensation gain parameter is adjustable online. The compensation mechanism is statically loaded before the start of each brewing cycle or dynamically interpolated and updated when a pulse event is triggered. The specific implementation method is set according to the real-time requirements of the system.

[0087] The manufacturing deviations simulated by the digital twin can refer to the differences in initial conductivity distribution caused by photolithography overlay errors, thin film deposition thickness fluctuations, or electrode contact resistance dispersion during the wafer-level manufacturing process of the memristor array; the electromigration effect can refer to the electromigration of active metal ions (such as...) under long-term high-frequency pulse writing conditions. , The process of directional migration under the action of an electric field and the resulting degradation of the morphology of conductive filaments; the fatigue effect can refer to the increase in interface state density and the hysteresis of oxygen vacancy rearrangement caused by repeated SET / RESET operations, which is macroscopically manifested as a prolonged conductivity relaxation time and a decrease in the on / off ratio; the modeling parameters of the above physical degradation process can all be periodically calibrated by the chip's built-in BIST module, and the calibration data is stored in the on-chip EEPROM and fed into the digital twin for online correction of model parameters.

[0088] The process by which the meta-parameter controller dynamically compensates for the synaptic plasticity rule based on the simulated state of the digital twin is as follows: In each brewing cycle, the controller first reads the current reference conductance snapshot of the memristor array and compares it with the theoretical conductance distribution predicted by the twin to generate a spatially resolved weight deviation map; subsequently, this deviation map is mapped to a local scaling factor matrix of the synaptic plasticity rule synaptic update path, which is used to modulate the weight update intensity caused by pulse triggering during forward propagation; when the conductance drift of a memristor unit is detected to exceed the preset tolerance threshold (±15%), the controller initiates a weight remapping mechanism to migrate the original weight values ​​to adjacent unsaturated units according to a linear / nonlinear mapping relationship, thereby maintaining the overall dynamic range and classification boundary stability of the network.

[0089] As an optional embodiment, the specific implementation of the present invention is as follows: In a continuous 10-round brewing test of Pu'er ripe tea, the system collects a pulse neural response sequence with a sampling rate of 1kHz within 30 seconds in each round; the digital twin calls the on-chip temperature sensor and BIST module to obtain the thermo-electric joint state parameters at 5-minute intervals after each brewing round, and updates its internal ion diffusion rate coefficient and interface degradation index; the meta-parameter controller generates the synaptic plasticity rule compensation parameters for this round, and completes the weight drift correction within the first 10 seconds of the next brewing round; the actual test results show that when the digital twin compensation is not enabled, the concentration prediction MAE rises to 0.82mg / mL from the 8th round; after enabling it, the MAE is stable within the range of 0.31±0.07mg / mL throughout the 10 rounds, and the fluctuation of the convergence step of the synaptic plasticity rule weight update is less than ±3%, which verifies the effective suppression of hardware aging by the compensation mechanism.

[0090] Through the above technical solution, this invention achieves the following: a digital twin performs device-level modeling of conductance drift caused by manufacturing deviations, electromigration, and fatigue effects in memristors; a meta-parameter controller performs feedforward compensation on key parameters in the synaptic plasticity rule based on the deviation prediction value output by the model; the compensation process covers typical degradation modes throughout the entire life cycle of the memristor, and the synaptic plasticity rule can still maintain the accuracy and directional consistency of weight updates during hardware aging; the compensation mechanism directly acts on the synaptic plasticity regulation layer rather than the application layer algorithm, improving the prediction robustness and system reliability of the spiking neural network processing module in long-term use scenarios without increasing additional inference latency.

[0091] Example 8: In another embodiment, the present invention also provides a meta-parameter controller that regulates the rules of synaptic plasticity, subject to global scheduling by an exploration-exploitation strategy module based on reinforcement learning. This exploration-exploitation strategy module treats each brewing cycle as an exploratory experiment, and its action is to set a set of meta-parameters. The reward is a weighted sum of prediction accuracy and energy consumption. Through multiple brewing cycles, the exploration-exploitation strategy module learns a meta-parameter scheduling strategy for the tea varieties most frequently consumed by the current user, thereby achieving an optimal balance between rapid adaptation and stable performance.

[0092] The reinforcement learning-based exploration-exploitation policy module is a lightweight reinforcement learning agent embedded within the metaparameter controller. Its algorithm architecture employs one of Q-learning, SARSA, or proximal policy optimization. This module does not rely on pre-trained models or cloud collaboration; all computations are performed locally to adapt to edge deployment scenarios. Its state space is a combination vector of the current brewing stage identifier, the prediction error sequence from previous rounds, real-time power consumption monitoring values, and the confidence level of tea category identification. The action space is a set of discretized values ​​for parameters such as the amplitude of the synaptic plasticity rule learning window, time constant, and synaptic decay rate, which are adjustable by the metaparameter controller. The reward function is... ,in The mean absolute error in predicting tea infusion concentration. This represents the average power consumption of the pulse neural network processing module during a single brewing cycle. and The normalized weighting coefficient is configurable and can be dynamically adjusted according to the device's power supply mode (battery / mains power).

[0093] The exploratory experiment is a time segment based on a complete brewing process, covering the entire process from water pouring, tea infusion, solute precipitation to pouring out the tea. During this process, the actions output by the strategy module are loaded into the meta-parameter controller and act on the currently active synaptic plasticity rule update pathway on the memristor cross array. The system response is based on the deviation between the predicted result pulse signal and the actual concentration calibration value, as well as the measured power consumption of the hardware, and is normalized to form an immediate reward. This reward is used to update the value function of the strategy module or the strategy network parameters.

[0094] The meta-parameter scheduling strategy is a mapping table or parameterized strategy function. Its inputs include tea type codes, brewing counts, ambient temperature ranges, and user historical preference tags; the output is a set of optimized meta-parameter configurations. This strategy exhibits a high exploration rate in the initial stage. ), gradually anneal to a stable value with each brewing cycle. This strategy aims to achieve a smooth transition from extensive trial and error to precise convergence. The storage form of this strategy is a compact lookup table in non-volatile memory or a linear regression model in a low-dimensional embedded space. Its structural complexity can be tailored according to the computing power resources of the terminal device.

[0095] The optimal balance between rapid adaptation and stable performance can be understood as follows: In the early stages of introducing new tea leaves, the strategy module actively tries diverse combinations of meta-parameters to improve its ability to model unknown evolution dynamics; after using the same tea leaf for more than 5 consecutive rounds, the strategy gradually converges to a high-confidence parameter set, reducing prediction volatility and suppressing the weight drift amplification caused by the non-ideal effects of memristors; this balancing mechanism does not change the basic topology of the spiking neural network processing module, nor does it introduce additional sensing channels or control execution units, but only achieves performance self-evolution through localized strategy iteration.

[0096] Specifically, during operation, the exploration-utilization strategy module first initializes the action space constraint range based on the current tea type identification result; then, at the start of each brewing round, it samples the system state, calls the current strategy to generate meta-parameter actions, and issues them; after the tea is brewed, it synchronously collects concentration calibration data and power consumption logs, calculates weighted rewards, and updates the strategy model; when it detects that the user selects the same tea type 3 times consecutively and the brewing parameters (water temperature, tea amount, steeping time) fluctuate less than the set threshold, the system automatically activates the strategy freezing mechanism, sets the current optimal parameter group as the default configuration, and retains a small probability of perturbation to maintain long-term online adaptability.

[0097] As an optional embodiment, the specific implementation of the present invention is as follows: When a user uses a certain high-mountain oolong tea for the first time, the system inputs the tea image recognition result (high-mountain oolong tea) into the strategy module, triggering the initial exploration mode; the module randomly selects three sets of significantly different meta-parameter combinations (synaptic plasticity rule window amplitudes are set to 0.3, 0.6, and 0.9 respectively), and applies them sequentially in the first three rounds of brewing; after each round, the system calculates the reward based on the laboratory-calibrated tea infusion concentration reference value and the measured power consumption, and updates the Q-value table; up to the 5th... During the first round, the strategy module identified the combination of amplitude 0.7 and time constant 120ms as having the highest overall score for this tea. Subsequently, the strategy entered the utilization-dominant phase, continuously using this configuration and introducing minor perturbations (such as ±0.05 amplitude offset) with only a 5% probability to address the differences in moisture content between batches of tea. After 20 rounds of brewing, the strategy converged and stabilized, with the predicted MAE decreasing from the initial 0.18g / L to 0.07g / L, and the average power consumption decreasing by 12%. Moreover, after switching to Pu'er ripe tea, the strategy could be recalibrated within 4 rounds.

[0098] Through the above technical solutions, this invention achieves the following: By introducing an exploration-utilization strategy module based on reinforcement learning, the regulation of the meta-parameter controller achieves closed-loop feedback and autonomous evolution, thus improving the system's generalization and adaptation capabilities to the extraction characteristics of different tea varieties; By modeling each brewing cycle as an independent Markov decision process and using prediction accuracy and energy consumption weighted as a unified reward signal, the hardware-level update of the synaptic plasticity rule combines accuracy objectives with energy efficiency constraints; Because the strategy module adopts a lightweight design and runs entirely locally, cloud dependency and communication latency are avoided, ensuring the real-time performance and privacy security of adaptive brewing control; Because the strategy convergence process is strongly coupled with the user's actual usage behavior, the resulting meta-parameter scheduling strategy has individualized characteristics and can dynamically evolve in response to changes in the user's tea-drinking habits.

[0099] Example 9: In one embodiment, the present invention also provides an adaptive brewing control module that incorporates a co-optimizer based on nonlinear model predictive control and pulse signal encoding. The co-optimizer decodes the predicted pulse signal output by the spiking neural network processing module into a concentration prediction trajectory within a future time window. Simultaneously, the co-optimizer integrates a simplified brewing dynamics model, using the brewing flow rate and water temperature fine-tuning as control variables, and aims to solve the optimal control problem in the finite time domain online with multiple objectives, including tracking the user's personalized concentration curve, maximizing the extraction efficiency of tea components, and minimizing energy consumption. It then outputs co-control commands to the brewing actuator and heating unit.

[0100] The co-optimizer is a software module deployed in an embedded microcontroller or dedicated coprocessor. Its operating environment supports mixed floating-point and fixed-point arithmetic and has real-time task scheduling capabilities. The input interface of this module can receive asynchronous pulse streams from the spiking neural network processing module and decode discrete pulse sequences into continuous concentration time sequences through methods such as time window integration, pulse count weighting, or event-driven interpolation. For example, the pulse firing frequency is normalized and statistically analyzed within a sliding time window of 100ms to 500ms, and mapped to equivalent concentration values ​​in the range of 0.1–5.0 mg / mL. This decoding method is set according to the actual system response delay and sensor sampling rate, and can be, for example, a linear mapping or a nonlinear sigmoid function mapping.

[0101] The simplified brewing kinetics model is a system of first- or second-order ordinary differential equations based on the principles of mass conservation and diffusion-convection coupling. Its state variables include the average tea concentration in the pot, effective wetting area, water temperature decay coefficient, and solute precipitation rate constant. The model parameters are preset or identified online based on the teaware material (e.g., purple clay, glass, stainless steel), capacity (200mL–600mL), initial water temperature (80℃–100℃), and tea type (green tea, black tea, oolong tea). The model structure is parametric, for example:

[0102] ;

[0103] in, For a moment The average concentration inside the pot. For a moment The brewing water temperature, The temperature-dependent precipitation rate coefficient is... For dynamic infiltration area, This represents the theoretical saturation concentration at the current temperature. The dilution factor for the soup is... To determine the real-time soup flow rate, the model is simplified to a lookup table method, piecewise linear fitting, or polynomial approximation based on actual deployment resource constraints. For example, it can be a reduced-order model that retains only the dominant term, or an enhanced simplified model that introduces an empirical correction factor.

[0104] The outflow rate is controlled by an adjustable orifice valve driven by a micro stepper motor, with an adjustment range of 0.5–15 mL / s and a resolution of up to 0.1 mL / s. Water temperature fine-tuning is another controlled variable, achieved by a PTC ceramic heating element in conjunction with a PWM power control circuit, with an adjustment accuracy of ±0.5℃ and a response time of less than 2 seconds. When the two work together, a decoupled control strategy or a joint state feedback mechanism is adopted. For example, when a high concentration rise slope is detected, the outflow rate is reduced simultaneously and the water temperature is slightly increased to slow down the subsequent precipitation rate, thereby smoothing out the concentration peak. This collaborative logic is embedded in the optimizer's internal rule base or is automatically generated by the online optimization process.

[0105] In the multi-objective optimization problem, the user-personalized concentration curve is based on a reference trajectory established according to the user's historical brewing preferences. For example, for a certain Lapsang Souchong, the target concentration sequence is set to slowly rise to 1.2 mg / mL in the first 30 seconds, and then maintain a plateau of 1.0–1.3 mg / mL until the 120th second. This curve is stored in local Flash or cloud configuration files and synchronized to the device via Bluetooth / WiFi. Maximizing the extraction efficiency of tea components is defined as the ratio of the total effective components released per unit mass of tea within a specified brewing time to the theoretical maximum release amount. Its evaluation is based on the relative precipitation kinetic model of marker components such as caffeine, tea polyphenols, and amino acids. Minimizing energy consumption is quantified as the sum of the cumulative power consumption of the heating unit and the driving energy consumption of the brewing actuator. Among them, the heating energy consumption is related to the temperature rise amplitude, heat capacity, and heat dissipation loss, while the brewing energy consumption is related to the number of valve actuations, holding time, and pressure difference. The weights of the above three objectives are set according to user preferences. For example, the default weight is concentration tracking: extraction efficiency: energy consumption = 4:3:3, or it can be switched to energy-saving priority mode or flavor priority mode according to the scenario.

[0106] The solution cycle for the finite-time optimal control problem is dynamically adjusted according to the brewing stage: 200ms / cycle during the initial soaking period (0–60s) to focus on rapid response to initial concentration changes; and extended to 500ms / cycle during the stable extraction period (60–180s) to balance computational load and control smoothness. The optimization time domain length is set to a prediction window of 1–5s in the future, corresponding to 10–50 control steps. The solution algorithm uses sequential quadratic programming (SQP), interior-point method, or gradient-based real-time optimizer, or adopts a gradient-free random search strategy to adapt to low-computing-power platforms. The algorithm is implemented as a pre-compiled firmware module or as lightweight Python bytecode loaded at runtime.

[0107] The collaborative optimizer works as follows: At the beginning of each control cycle, it first obtains the latest pulse flow from the spiking neural network processing module, and decodes it to obtain the concentration prediction trajectory for the next 5 seconds; then, it combines this trajectory with a simplified brewing kinetics model to construct a model for the real-time infusion flow rate. and water temperature regulation A multivariate objective function; then, satisfying physical constraints (such as...) , To achieve the maximum soup flow rate, Under the premise of the maximum water temperature regulation and operational constraints (such as the minimum valve opening / closing interval and the maximum duty cycle of the heating element), the built-in optimizer is invoked to solve for the optimal control sequence; finally, the first control command (i.e., the one to be executed at the current moment) is selected. and The control values ​​are sent to the soup dispensing actuator and the heating unit, while the remaining control values ​​are temporarily stored and participate in the next cycle of rolling optimization.

[0108] In one optional embodiment, the present invention is implemented as follows: Taking a user brewing a high-mountain oolong tea as an example, the system detects an accelerated increase in concentration output by the spiking neural network at the 15th second. The co-optimizer then decodes that the concentration will increase from 0.8 mg / mL to 1.4 mg / mL within the next 3 seconds. Combining the typical precipitation kinetic parameters of this tea, the optimizer determines that if the current infusion flow rate (8 mL / s) and water temperature (92°C) are maintained, the flavor limit of 1.3 mg / mL set by the user will be exceeded at the 28th second. Therefore, multi-objective optimization was initiated. Under the premise of ensuring that the extraction efficiency is not less than 85% and the energy consumption of a single heating does not exceed 120J, a coordinated control command was generated: the flow rate of the soup was reduced from 8mL / s to 5.5mL / s in steps, and the water temperature was simultaneously fine-tuned to 90.5℃. After the command was executed by the drive circuit, the concentration increase slowed down and stabilized at 1.28mg / mL at 32s, with the error controlled within ±0.03mg / mL. At the same time, the total energy consumption of the whole brewing round was reduced by 11.7%, and the extraction efficiency was increased to 89.2%.

[0109] Through the above technical solutions, this invention achieves the following: The collaborative optimizer decodes the event-driven prediction results output by the spiking neural network into a continuous concentration trajectory, and integrates a simplified brewing kinetics model for forward simulation. This enables pre-planning of control actions rather than passively responding to threshold triggers, improving the control accuracy of the timing of brewing and the concentration level. With the joint optimization objectives of tracking personalized concentration curves, maximizing extraction efficiency, and minimizing energy consumption, and supporting dynamic weight configuration, it can adaptively balance flavor, efficiency, and energy saving under different user preferences and usage scenarios. Control commands are synchronously applied to the brewing actuator and heating unit, achieving coupled control of water flow dynamics and heat transfer processes, avoiding concentration oscillations or thermal hysteresis mismatches caused by single-variable adjustment. The optimization solution process incorporates a real-time rolling mechanism and adapts to embedded computing power constraints, enabling stable operation in resource-constrained portable teaware, ensuring system practicality and engineering feasibility.

[0110] Example 10:

[0111] In an optional embodiment, the present invention also provides a multi-objective optimization problem in a collaborative optimizer, which is solved in real time by a distributed solver inspired by swarm intelligence. The solver runs on multiple parallel biomimetic agents, each representing a possible control trajectory, and follows simplified ant colony optimization or particle swarm optimization rules for pheromone exchange or velocity updates to quickly approach the Pareto optimal frontier, and selects the final execution scheme from the optimal solution set based on real-time calculated user satisfaction predictions.

[0112] Among them, the distributed solver inspired by swarm intelligence is a lightweight parallel optimization engine deployed on an embedded edge processor. Its physical implementation is a multi-threaded coprocessor module based on the ARM Cortex-M7 core, or multiple parallel computing units partitioned on an FPGA. This embodiment of the invention does not make any special limitations on this. The scheduling logic of the solver runs in a real-time operating system environment and allocates computing resources through time-slice round-robin or event-triggered methods to ensure millisecond-level response latency.

[0113] Multiple parallel biomimetic agents are software-defined computational entities. Each agent independently maintains a set of state vectors for control variables, including the set value of the soup flow rate, the water temperature fine-tuning range, the soup dispensing timing offset, and the duration. The state update process of each agent is executed asynchronously, and its data structure is configured as a fixed-length array or a circular buffer according to the actual memory resources. The number of agents is dynamically configured according to the system's computing power and real-time requirements, and can be set to 8, 16, or 32.

[0114] The simplified rules for ant colony optimization or particle swarm optimization are as follows: When using ant colony optimization, each agent simulates an artificial ant and constructs a path in a predefined control parameter space. The pheromone concentration update is based on the weighted sum of the mean square error between the concentration prediction trajectory and the target curve within the current brewing cycle and the energy consumption index. When using particle swarm optimization, each agent corresponds to a particle. The inertia weight, cognitive factor, and social factor in its velocity update formula are configured with piecewise constants, and online gradient estimation is not introduced. Both rules omit the frequent broadcast synchronization operation of the global optimal historical position in the original algorithm and instead use a local neighborhood information exchange mechanism to adapt to the limited communication bandwidth of edge devices.

[0115] Pheromone exchange or velocity updates are lightweight communication behaviors triggered periodically, with the exchange frequency dynamically adjusted according to the brewing phase: once every 500ms in the initial extraction phase (first 30 seconds), once every 1s in the stable extraction phase (30–120 seconds), and once every 2s in the decay phase (after 120 seconds); pheromone or velocity status is transmitted through a shared memory-mapped region or a low-overhead message queue, without relying on the TCP / IP protocol stack.

[0116] The Pareto optimal front is the set of all feasible solutions for any objective that cannot be further optimized without degrading at least one objective in a three-dimensional objective space consisting of concentration tracking bias, extraction efficiency, and energy consumption. This front is constructed online using a non-dominated sorting algorithm, and the objective function value on which the sorting depends is output by the real-time decoding module inside the co-optimizer, without relying on external databases or cloud services.

[0117] User satisfaction prediction is based on a lightweight regression model. Its input features include historical brewing preference labels (such as preference for strength, tolerance to bitterness, and expectation of aftertaste), current environmental parameters (such as room temperature and humidity), and the code of the tea variety being brewed. The model is deployed in TFLiteMicro format, inferences on the MCU, and outputs a normalized satisfaction score in the range [0,1], which is used to weight and rank the candidate solutions in the Pareto front. The training samples of the model come from user active feedback or implicit behavior modeling (such as pause / refill operation duration and number of manual adjustments). Its structure is a 3-layer fully connected network or a lookup table interpolation model. This embodiment of the invention does not impose any special limitations on this.

[0118] Specifically, after receiving the concentration prediction trajectory for the next 10 seconds output by the spiking neural network processing module, the distributed solver first decodes the trajectory into a target concentration sequence at discrete time points. Then, it initializes N bionic agents, each of which randomly generates a set of soup dispensing control parameters and performs forward simulation of the corresponding concentration response curve based on a simplified brewing kinetics model. Next, each agent iteratively updates its own parameters according to ant colony or particle swarm optimization rules and uploads its local solution to the shared frontier pool after each iteration. After a preset maximum of 15 iterations, the frontier pool completes non-dominated sorting and generates a Pareto front containing 5–10 candidate solutions. Finally, the user satisfaction prediction module outputs a satisfaction score for each candidate solution, and the collaborative optimizer selects the control parameter set corresponding to the highest score, encapsulates it into a control command frame, and sends it to the soup dispensing actuator and heating unit via CAN bus or SPI interface.

[0119] As an optional embodiment, the specific implementation of the present invention is as follows: During a brewing process of oolong tea, the system detects that the current tea concentration is increasing at a rate of 0.12% / s, and is expected to peak at 42 seconds; the collaborative optimizer starts the distributed solver, initializes 24 particle-type agents, sets the maximum number of iteration rounds to 12, and the pheromone / velocity exchange cycle to 800ms; each agent explores in a three-dimensional space consisting of the outflow rate (0.5–3.0mL / s), water temperature fine-tuning (±2℃), and outflow start offset (−5–+8s); after 1 After two rounds of iteration, the frontier pool generated seven non-dominated solutions, of which solution A corresponds to early start and slow output, solution B corresponds to delayed fast output, and solution C corresponds to stepped segmented output. The user satisfaction prediction module combined the user's historical preferences (labeled as preference for rich and long-lasting soup) and the current room temperature (26℃), and output that solution C scored 0.93, which is higher than solution A (0.71) and solution B (0.65). Finally, the system selected solution C, drove the micro solenoid valve to open in a three-stage sequence, and simultaneously instructed the PTC heating element to fine-tune the water temperature by +1.2℃, realizing the coordinated response of the soup output action and the thermodynamic state.

[0120] Through the above technical solutions, this invention achieves the following: It employs a distributed solver inspired by swarm intelligence and runs in parallel on multiple biomimetic agents, simplifying ant colony or particle swarm optimization rules, significantly shortening the solution time for multi-objective optimization problems and meeting millisecond-level real-time control response requirements; each agent collaborates to explore the solution space through local pheromone exchange or velocity update mechanisms, improving its adaptability to non-convex, nonlinear, strongly coupled brewing dynamics models; it selects the final execution scheme from the Pareto optimal front based on real-time calculated user satisfaction predictions, achieving a dynamic balance between objective performance indicators and subjective experience in control decisions; the entire solution process is completed locally at the edge, without relying on cloud collaboration or high-computing servers, ensuring the system's autonomous decision-making capability under offline, low-power, and small-volume constraints.

[0121] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tea infusion concentration prediction and adaptive dispensing system based on machine learning, characterized in that, include: A multimodal sensor array is used to acquire at least two different types of sensor signals of tea soup in real time during the brewing process; A data processing and encoding module, connected to the multimodal sensor array, is used to preprocess and encode the sensing signals into a time-series pulse sequence; The core of the spiking neural network processing module is a memory computing hardware circuit built on an analog memristor cross array. This module is used to receive the time-series pulse sequence, perform event-driven calculations on the input spatiotemporal pulse pattern through the spiking neural network, and directly output a predicted pulse signal that characterizes the real-time concentration or concentration change trend of the tea soup. An adaptive soup dispensing control module is connected to the pulse neural network processing module and a soup dispensing actuator. It is used to generate control commands based on the predicted pulse signals and drive the soup dispensing actuator to adaptively open or close the soup dispensing when the predicted concentration reaches a preset threshold. The pulse neural network processing module realizes local storage of neural network weights and analog domain multiplication and addition operations through the analog memristor cross array, so as to complete the dynamic prediction of the tea soup concentration with ultra-low power consumption.

2. The system as described in claim 1, characterized in that, The multimodal sensor array includes at least one miniature near-infrared spectral sensor unit and one multi-band impedance sensor unit. The system also includes a sensor fusion and state reconstruction module, which embeds a lightweight computational fluid dynamics-convolutional neural network hybrid model. The hybrid model takes local point data acquired by the front-end sensor as input, simulates the flow field and solute diffusion in the pot in real time through the lightweight computational fluid dynamics model, and uses the convolutional neural network as a residual corrector to dynamically learn and compensate for the deviation between the actual physical parameters and the simulation model. Finally, it outputs a dynamic cloud map of the tea concentration field in the three-dimensional space inside the pot, which serves as spatial context enhancement information for the time-series pulse sequence received by the spiking neural network processing module.

3. The system as described in claim 2, characterized in that, The data processing and encoding module adopts a sparse encoding strategy based on the bionic auditory nerve. This module converts the continuous analog signals from the multimodal sensor array into multi-channel features that conform to the cochlear frequency-position mapping principle. It also draws on the firing adaptive mechanism of auditory neurons to generate and transmit pulses only when the signal feature changes exceed a dynamic threshold, thereby generating a highly sparse and information-dense temporal pulse sequence.

4. The system as described in claim 3, characterized in that, The sparse coding strategy of the bionic auditory nerve is implemented by a pulse recurrent neural network module, which simulates the feedforward and lateral inhibition network of primary auditory cortex neurons. It dynamically adjusts the firing threshold of each channel through internal pulse activity to achieve adaptive focusing on the time-frequency domain salient features of continuous signals, and simultaneously generates a regulating pulse flow that characterizes the overall uncertainty level of the signal, which is then fed to the pulse neural network processing module.

5. The system as described in claim 4, characterized in that, The sparse coding strategy of the bionic auditory nerve does not use fixed multi-channel frequency-position mapping parameters, but rather dynamically configures them using an online optimizer based on an evolutionary strategy. The online optimizer uses the accuracy and sparsity of the final concentration prediction of the spiking neural network processing module as a joint reward signal. Over several brewing cycles, it fine-tunes the center frequency and bandwidth of the filter bank of the cochlear mapping in a gradient-free manner, enabling the entire coding-processing link to co-evolve and adapt to the unique acoustic resonance characteristics of the teaware currently in use and the spectral characteristics of the specific tea variety.

6. The system as described in claim 5, characterized in that, In the spiking neural network processing module, the weight update mechanism of the simulated memristor cross array is regulated by meta-learning-driven synaptic plasticity rules. Based on the traditional pulse-time-dependent plasticity rules, a meta-parameter controller is introduced. This controller dynamically adjusts the amplitude and time constant of the learning window of the synaptic plasticity rules of the entire network according to the prior information of the current tea category and the prediction error feedback of the previous brewing cycles, thereby realizing the network's rapid online adaptive fine-tuning of new tea varieties at the hardware level.

7. The system as described in claim 6, characterized in that, The meta-parameter controller integrates a digital twin, which is a device-level behavioral model of the physical memristor array in the spiking neural network processing module. It can simulate the drift of memristor conductance caused by manufacturing deviations, electromigration, and fatigue effects. Based on the simulation state of the digital twin, the meta-parameter controller dynamically compensates for the distortion of network weight representation caused by hardware aging, thereby enabling the synaptic plasticity rule to maintain a stable and effective learning ability throughout the entire hardware lifecycle.

8. The system as described in claim 7, characterized in that, The meta-parameter controller regulates the synaptic plasticity rules under the global scheduling of an exploration-exploitation strategy module based on reinforcement learning. This exploration-exploitation strategy module treats each brewing cycle as an exploratory experiment, and its action is to set a set of meta-parameters. The reward is a weighted sum of prediction accuracy and energy consumption. Through multiple brewing cycles, the exploration-exploitation strategy module learns a meta-parameter scheduling strategy for the tea varieties most frequently consumed by the current user, thereby achieving an optimal balance between rapid adaptation and stable performance.

9. The system as described in claim 8, characterized in that, The adaptive brewing control module incorporates a co-optimizer based on nonlinear model predictive control and pulse signal encoding. This co-optimizer decodes the predicted pulse signal output by the spiking neural network processing module into a concentration prediction trajectory within a future time window. Simultaneously, the co-optimizer integrates a simplified brewing kinetics model, using brewing flow rate and water temperature fine-tuning as control variables, and aims to solve the optimal control problem in the finite time domain online with multiple objectives, including tracking the user's personalized concentration curve, maximizing the extraction efficiency of tea components, and minimizing energy consumption. It then outputs co-control commands to the brewing actuator and heating unit.

10. The system as described in claim 9, characterized in that, The multi-objective optimization problem in the collaborative optimizer is solved in real time by a distributed solver inspired by swarm intelligence. The solver runs on multiple parallel biomimetic agents, each representing a possible control trajectory, and follows simplified ant colony optimization or particle swarm optimization rules to exchange pheromones or update velocity to quickly approach the Pareto optimal frontier. The final execution scheme is then selected from the optimal solution set based on real-time calculated user satisfaction predictions.

Images