Reinforcement learning driven AI voice Internet of Things intelligent control optimization method

Abstract

The invention discloses a reinforcement learning driven AI voice Internet of Things intelligent control optimization method, and relates to the technical field of voice control. The method comprises the following steps: acquiring multi-mode signals such as voice and environment through a microphone array and multiple sensors, and forming an input vector through space-time alignment and compression processing; analyzing the voice instruction by using an attention mechanism and a knowledge graph, and converting the voice instruction into an equipment executable parameter; dynamic impedance matching is realized by means of reinforcement learning, and the system energy efficiency is improved; multi-device phase synchronization is completed based on a phase-locked loop technology; predicting equipment faults through vibration modal analysis and transfer learning; meanwhile, the system has the functions of scene perception, energy recovery and safety protection, and can automatically adapt to scenes to recover energy. The control performance of the AI voice Internet of Things is improved, the voice recognition accuracy is improved, the system efficiency is high, the synchronization error of multiple devices is reduced, faults can be early warned 50 hours in advance, the energy recovery efficiency is doubled, the standby power consumption is reduced, and the interactivity, stability and energy saving performance of the devices are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of voice control technology, and in particular to an AI voice Internet of Things intelligent control optimization method driven by reinforcement learning. Background Art

[0002] With the popularization of Internet of Things technology, AI voice control devices are increasingly used in scenarios such as smart homes, outdoor equipment, and industrial Internet of Things. However, existing technologies still have significant bottlenecks in terms of interaction reliability, energy efficiency management, and multi-device collaboration in complex environments.

[0003] At the voice interaction level, traditional methods rely on a single microphone to pick up sound, which is easily affected by environmental noise. This significantly reduces speech recognition accuracy, especially in noisy scenarios such as outdoor camping and factory workshops. At the same time, existing semantic parsing models lack a deep understanding of context and scenarios, making it difficult to handle ambiguous commands (such as "dim the lights" when the device target is unclear), leading to control errors. Furthermore, cross-device identity authentication and privacy protection mechanisms are imperfect, posing a risk of leakage of user voice data and device status information.

[0004] In terms of energy efficiency and control optimization, traditional IoT devices often use fixed-parameter control strategies, which are unable to dynamically adapt to load changes. For example, when the light source ages or the ambient temperature changes, the LED driver circuit will experience significant impedance mismatch, resulting in decreased system efficiency and increased energy consumption. In multi-device collaboration scenarios, traditional synchronization protocols (such as NTP) lack accuracy (millisecond-level errors), making it difficult to achieve high-precision requirements such as light and shadow synchronization. They also lack the ability to proactively predict device failures, often requiring post-failure maintenance, leading to downtime losses and wasted resources.

[0005] In terms of technology integration and innovation, existing solutions often design voice recognition, control algorithms, energy efficiency optimization, and other modules independently, lacking a cross-layer collaborative intelligent optimization framework. For example, the voice interaction module is not linked to the device's energy consumption status and cannot automatically adjust the interaction strategy in low-battery scenarios. Multi-device collaborative control does not incorporate distributed technologies such as blockchain, making it difficult to ensure the consistency and immutability of command execution. In addition, energy recovery technology is limited to a single energy type (such as only recovering ripple energy) and lacks an intelligent energy allocation strategy, resulting in low recovery efficiency. Summary of the Invention

[0006] The present invention proposes a reinforcement learning-driven AI voice IoT intelligent control optimization method to solve the problems mentioned in the above-mentioned prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a reinforcement learning-driven AI voice IoT intelligent control optimization method, comprising:

[0008] Signal spatiotemporal alignment acquisition steps: Stereo sound pickup is achieved through a microphone array. Combined with the Bluetooth module's phase-synchronized clock signal, the generalized cross-correlation delay estimation (GCC-PHAT) algorithm is used to localize the sound source. Adaptive sliding window filtering is applied to the supply voltage, with the window size dynamically adjusted according to voltage fluctuations to suppress ripple noise.

[0009] Semantic enhancement and parsing of voice commands: For complex keystrokes, a hierarchical state transition network (HSTN) is designed to construct a spatiotemporal correlation matrix between keystroke sequences and voice commands. A voice enhancement algorithm is introduced to improve the signal-to-noise ratio of voice signals in environments with strong background noise, and an attention mechanism is used to parse the ambiguous semantics of commands.

[0010] Reinforcement learning dynamic impedance matching steps: Based on the unique frequency characteristics of the LC resonant circuit, an impedance adaptive adjustment model is constructed; the reinforcement learning algorithm is used to optimize the impedance matching between the driving circuit and the light source load; the driving circuit impedance is set to Z d , the light source load impedance is Z l , define the matching function as: Where Re represents the real part operation, It's Z l The conjugate complex number of the output PWM pulse width modulation frequency adjustment instruction is adjusted with a step size of 0.1KHz during the reinforcement learning process. Assuming the initial PWM frequency is f0 and the adjusted frequency is f, it is calculated through continuous iteration: f = f0 + n × 0.1KHz, where n is the number of adjustment steps, which is dynamically determined by the feedback of the matching function M through the reinforcement learning algorithm. At the same time, in order to suppress electromagnetic interference EMI, an adaptive notch filter is introduced; the center frequency fc of the filter tracks the PWM base frequency f in real time. pwm , f c =f pwm ;Suppress electromagnetic interference through adaptive tracking mechanism to ensure stable operation of the system;

[0011] Phase synchronization control steps for each device: A synchronization pulse signal is injected through the hardware interface. The master device broadcasts the phase reference, and the slave device locks the PWM output phase through the phase-locked loop (PLL) circuit to achieve synchronization of the flashing frequency of the light sources of each device. In the camping light networking scenario, the light and shadow synchronization matrix is ​​constructed by dynamically adjusting the gradient change of the PWM duty cycle of each device.

[0012] Fault prediction steps based on vibration modal analysis: collect equipment triaxial vibration data through the reserved interface; decompose the signal into intrinsic mode function (IMF) through empirical mode decomposition (EMD) i (t), where i represents the i-th intrinsic mode function; it is input into the convolutional neural network CNN, and the fault is identified based on the K mutation of the IMF3 kurtosis value; is the normal mean, and ΔK is the set threshold. When the predicted remaining life L≤50 hours, a pre-maintenance strategy is generated through reinforcement learning (RL) to ensure equipment operation.

[0013] Furthermore, it also includes:

[0014] Voice interaction scene perception steps: The integrated ambient light sensor and power supply voltage fluctuation characteristics are used to identify user camping, emergency, and maintenance scenarios. In camping scenarios, the scene-specific voice command set is automatically activated to trigger the RGBW light mixing algorithm, achieving 16 million dynamic color gradients. At the same time, reinforcement learning is used to optimize the voice wake-up threshold to improve the wake-up success rate in noisy environments.

[0015] Energy recovery reinforcement learning steps: Utilizing the bidirectional energy flow characteristics of the buck-boost converter, when the device is in standby mode, the MOSFET conduction timing is controlled through reinforcement learning to recover the ripple energy of the light source driving circuit to the energy storage capacitor; an energy management strategy based on stochastic gradient descent (SGD) is designed to reduce standby power consumption, while maintaining the energy storage capacitor voltage through adaptive duty cycle adjustment.

[0016] Furthermore, in the signal time-space alignment acquisition step, the compressed sensing CS technology is introduced to downsample the vibration data, and the sampling rate is reduced to the Nyquist rate. The signal is reconstructed using the orthogonal matching pursuit (OMP) algorithm under the condition of .

[0017] Furthermore, in the step of semantic enhancement analysis of voice commands, a command reasoning model based on the knowledge graph is constructed, a mapping relationship is established between the professional terms of the floodlight mode and the hardware parameters, the light intensity distribution after the command execution is predicted through the graph neural network GNN, and the control parameters are preloaded to the slave device cache in advance.

[0018] Furthermore, in the reinforcement learning dynamic impedance matching step, a dual-loop control architecture is designed. The inner loop adjusts the PWM duty cycle in real time through a PID controller to maintain a constant output current; the outer loop optimizes the PID parameters through reinforcement learning, and automatically readjusts the proportional coefficient, integral time, and differential time when the internal resistance changes due to aging of the light source.

[0019] Furthermore, in the phase synchronization control step of each device, a device status consensus mechanism is adopted. Every 100 control instructions generated by the master device are packaged into a timestamp and instruction hash value, written into the device blockchain through a specific hardware interface, and the validity of the block is verified by the slave device to achieve synchronization of the instruction execution order.

[0020] Furthermore, in the fault prediction step based on vibration modal analysis, a transfer learning framework is introduced to migrate the fault model trained in the laboratory environment to the outdoor environment. The convolution layer parameters are adjusted through domain adaptation (DA) technology to achieve the required fault recognition rate in the field scenario. At the same time, federated learning (FL) is used to update the model at the edge node to protect the privacy of user device data.

[0021] Furthermore, in the voice interaction scene perception step, an adaptive equalizer EQ circuit is designed to automatically switch the filtering mode according to the scene recognition result: the low-frequency enhancement mode is enabled in the camping scene to enhance the human voice frequency band in the ambient sound; the high-frequency enhancement mode is enabled in the emergency scene to highlight the penetration of the alarm command, and the hardware filtering parameter switching is realized through the resistance combination.

[0022] Furthermore, in the energy recovery reinforcement learning step, a multi-objective optimization model is constructed to minimize standby power consumption and maximize energy storage efficiency at the same time, and an ε-greedy algorithm is used to balance exploration and utilization to achieve the required energy storage capacitor charge retention rate in the standby test.

[0023] Compared with the existing technology, the beneficial effects of the present invention are:

[0024] In the field of voice interaction, this application uses a five-microphone array and a multi-scale denoising algorithm to increase the signal-to-noise ratio of voice signals to more than 85dB. Combined with the four-level attention mechanism and knowledge graph parsing technology, the voice recognition accuracy reaches 80% even in complex environments with noise intensity as high as 40dB, significantly surpassing traditional methods. This application method supports multiple rounds of dialogue and can understand various scenario-based instructions. For example, after receiving the "night camping mode" instruction, the system can automatically adjust the brightness and color temperature of the device to provide users with a comfortable experience and significantly improve interaction efficiency. At the same time, the introduction of speaker recognition and differential privacy technology can effectively protect user data privacy while achieving personalized interaction. The misrecognition rate is controlled at an extremely low level, and the risk of data leakage is greatly reduced.

[0025] In terms of energy efficiency optimization and fault prediction, an innovative four-level reinforcement learning architecture achieves dynamic impedance matching, increasing system efficiency to 93.5% and reducing standby power consumption to 3.2mW. The multi-port energy recovery system fully utilizes various energy sources, including ripple, temperature differences, and vibration, achieving a recovery efficiency of 72%. In typical scenarios, the device's self-powered capacity is significantly enhanced, and the battery replacement cycle is extended to 9 months. A fault prediction model based on vibration modal analysis and transfer learning can accurately warn of faults such as bearing pitting and motor imbalance 24 to 72 hours in advance. With high warning accuracy, it successfully transforms post-maintenance mode into predictive maintenance, significantly reducing losses caused by unplanned downtime.

[0026] In terms of multi-device collaboration, the blockchain-based distributed synchronization protocol achieves microsecond-level time synchronization, with an error controlled within 5μs. In a scenario where 100 devices are collaborating, the light and shadow consistency error is only 4%. This application method supports complex light and shadow narratives and implements edge computing load sharing, with command response latency as low as 50ms, meeting the needs of real-time collaboration. The application of zero-knowledge proof identity authentication and abnormal behavior detection mechanisms increases the security vulnerability detection rate to 95%, effectively ensuring the legitimacy of device identity and the security of data transmission.

[0027] In terms of scene adaptability and technological innovation, the hidden Markov model is used to predict user behavior trends, and it can automatically adapt to more than 10 different scenes such as camping, emergency, and plateau, with high scene recognition accuracy. The quantum-inspired reinforcement learning algorithm accelerates strategy convergence, and combines the electrothermal coupling model with the multi-objective optimization framework to achieve balanced optimization of energy efficiency, lifespan, and stability. The "perception-decision-execution-optimization" intelligent closed loop constructed by this application method provides an innovative technical paradigm of cross-layer collaboration for AIoT devices. It has broad application prospects in many fields such as smart homes, outdoor equipment, and industrial Internet of Things, and will effectively promote the industry to develop in the direction of efficiency, intelligence, and greenness. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic block diagram of the reinforcement learning-driven AI voice IoT intelligent control optimization method proposed in the present invention;

[0029] Figure 2 This is a schematic diagram comparing the multimodal signal denoising effects of a reinforcement learning-driven AI voice IoT intelligent control optimization method proposed in the present invention;

[0030] Figure 3 This is a schematic diagram comparing the reinforcement learning convergence speed of the reinforcement learning-driven AI voice IoT intelligent control optimization method proposed in this invention. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0033] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0034] Reference Figure 1 and Figure 3 :A specific implementation of a reinforcement learning-driven AI voice IoT intelligent control optimization method

[0035] 1. Steps for temporal and spatial alignment of multimodal signals

[0036] At the hardware level, the microphone array uses five Knowles SPM1423 high-sensitivity MEMS microphones, which have a sensitivity of -38dBFS and a frequency response range of 20Hz-20kHz. They are arranged in a regular pentagonal pattern, and the spacing between adjacent microphones is precisely set to 8cm, which can effectively achieve 360-degree full-range sound capture. Each microphone is connected to an independent TLV320ADC3101 preamplifier, and its gain is adjusted by I 2The C interface can be fine-tuned in 0.5dB steps within the range of 0-60dB, ensuring that both faint human whispers and loud ambient sounds can be clearly and accurately captured. The Bluetooth module uses the nRF52840 chip, which supports the IEEE802.1AS standard and can output a phase-synchronized clock signal with an accuracy of up to 1ns, providing an extremely accurate time reference for sound source positioning. During the sound source positioning process, the generalized cross-correlation delay estimation (GCC-PHAT) algorithm is used to process the signals collected by the microphone array. Through more than 10,000 calibrations of measured data in different environments (such as indoors, outdoors, and noisy streets), the positioning error is finally stably controlled within 0.3 meters, meeting the positioning requirements of most practical application scenarios.

[0037] For power supply voltage acquisition, a voltage divider circuit consisting of two high-precision, low-temperature drift resistors (TCR ≤ 5ppm / °C) converts the 3.0-4.2V lithium battery voltage into a 0-3.3V signal suitable for acquisition by the STM32H7 series microcontroller ADC. An adaptive sliding window filtering algorithm is implemented in the microcontroller through a hardware acceleration module. The window size is determined by the standard deviation of the voltage fluctuations. When the standard deviation exceeds 0.2V, the window is rapidly reduced to 10ms to quickly respond to drastic voltage changes. When the standard deviation is less than 0.1V, the window is expanded to 100ms to enhance filtering effectiveness. Ultimately, ripple noise is suppressed to within ±50mV, ensuring accurate power supply parameter acquisition. For vibration data acquisition, the InvenSense ICM-20948 triaxial accelerometer was selected, with a measurement range of ±16g, a resolution of 0.001g, and a sampling rate of up to 8kHz. Compressed sensing (CS) technology was introduced, downsampling the original 10kHz vibration data to 2.5kHz using a random Gaussian matrix. The signal was then reconstructed using the orthogonal matching pursuit (OMP) algorithm. To further improve reconstruction accuracy, an iterative optimization mechanism was incorporated into the algorithm. After multiple iterations, the reconstruction error was kept within 3%, significantly saving 40% of wireless transmission bandwidth and effectively reducing data transmission pressure.

[0038] In the signal processing flow, all collected voice, power supply, and vibration signals first enter the preprocessing module. In addition to standard DC removal and normalization, voice signal preprocessing also includes endpoint detection. This method combines dual-threshold energy detection with zero-crossing rate to accurately identify the start and end points of speech, remove ineffective silence, and improve subsequent processing efficiency. After filtering, the power supply signal undergoes outlier detection, using the 3σ principle to determine whether the voltage data is abnormal. If an anomaly is detected, median filtering is used to correct it. After downsampling and reconstruction, the vibration signal undergoes frequency domain analysis. Fast Fourier Transform (FFT) is used to convert the time domain signal into the frequency domain, extracting the signal's frequency characteristics and providing data support for subsequent fault diagnosis.

[0039] 2. Steps for semantic enhancement of voice command analysis

[0040] The Hierarchical State Transition Network (HSTN) uses a three-level state design: initial state, key recognition state, and instruction execution state. Taking a compound key operation as an example, when the device's power button and mode button are pressed simultaneously, the hardware circuit generates a specific level combination signal, which is input into the microcontroller's GPIO pin. The microcontroller captures this signal through its internal interrupt service routine and enters the key recognition state. In the key recognition state, the system determines the function corresponding to the key combination by querying a pre-set key combination table. At the same time, to prevent false triggers, the system will judge parameters such as the duration and interval of key presses. Only key operations that meet specific conditions will be considered valid operations.

[0041] For speech enhancement, spectral subtraction based on a generative adversarial network (GAN) runs on a dedicated ADISHARC series digital signal processor (DSP). The generator uses a U-Net architecture consisting of 8 downsampling layers and 8 upsampling layers. Each layer uses the ReLU activation function, which can effectively learn noise characteristics and generate clean speech. The discriminator consists of a 5-layer convolutional neural network and uses the LeakyReLU activation function. By distinguishing the difference between the generated speech and the real clean speech, it guides the generator to optimize. After training in a variety of noise environments (such as factory noise, traffic noise, wind and rain noise, etc.), using more than 500 hours of noisy speech data, the signal-to-noise ratio of the speech signal can be improved by more than 12dB in a strong background noise environment of 75dB, significantly improving the speech quality.

[0042] In the implementation of the attention mechanism, temporal attention extracts temporal features from speech signals using a bidirectional LSTM network. This bidirectional LSTM network consists of two layers, each with 256 hidden units. It fully captures the dependencies between the preceding and following speech signals and extracts key temporal features. Frequency-domain attention utilizes a channel-wise attention mechanism on the mel-spectrogram to enhance the weights of semantically relevant frequency bands. Specifically, the mel-spectrogram is first subjected to global average pooling and global max pooling to obtain two feature vectors. These two feature vectors are then concatenated and passed through two fully connected layers to obtain per-channel weights. Finally, these weights are multiplied with the original mel-spectrogram to enhance the features of important frequency bands. Contextual attention, based on the Transformer architecture, combines historical interaction records to understand the context of commands. The system maintains an interaction history buffer of size 10. Whenever a new command is received, it is fed into the Transformer model along with the previous commands in the buffer. A multi-head attention mechanism calculates the associations between different commands, thereby accurately understanding the contextual meaning of the current command.

[0043] The instruction reasoning model is constructed through the knowledge graph. The knowledge graph is stored and managed using the Neo4j graph database. It contains more than 10,000 entities and 50,000 relationships, covering various device parameters, operating instructions and scenario information. For example, when the "flood mode" instruction is received, the graph neural network (GNN) first searches for nodes related to "flood mode" in the knowledge graph to obtain its corresponding hardware parameter information, such as PWM frequency 15KHz, duty cycle 80%, etc. Then, the message passing mechanism of GNN is used to predict the light intensity distribution after the instruction is executed. Through multiple iterative optimizations, the prediction error is controlled within 2.5%. Finally, the system preloads these parameters to the slave device cache in advance to ensure that the instruction can be executed quickly and accurately.

[0044] 3. Reinforcement Learning Dynamic Impedance Matching Steps

[0045] Based on the unique frequency characteristics of the LC resonant circuit, an impedance adaptive adjustment model is constructed; the reinforcement learning algorithm is used to optimize the impedance matching between the driving circuit and the light source load; the driving circuit impedance is set to Zd, the light source load impedance is Zl, and the matching function is defined as:, where Re represents the real part operation and is the conjugate complex number of Zl; during the reinforcement learning process, the output PWM pulse width modulation frequency adjustment instruction is adjusted with a step size of 0.1KHz; the initial PWM frequency is set to f0, and the adjusted frequency is f, and the iterative calculation is continuous:, where n is the number of adjustment steps, which is dynamically determined by the reinforcement learning algorithm based on the feedback of the matching function M; at the same time, in order to suppress electromagnetic interference EMI, an adaptive notch filter is introduced; the center frequency fc of the filter tracks the PWM base frequency in real time; the adaptive tracking mechanism is used to suppress electromagnetic interference to ensure stable operation of the system.

[0046] The LC resonant circuit uses a TDK inductor (model: NLFC523222T2R2MN) with an inductance of 2.2μH and a Murata capacitor (model: GRM31CR61E226ME15L) with a capacitance of 22μF. The resonant frequency can be optimized by fine-tuning the capacitor or inductor parameters according to actual needs. The reinforcement learning algorithm runs on an STM32H750 series 32-bit microcontroller, using a modified version of the Deep Deterministic Policy Gradient (DDPG) algorithm, TD3 (TwinDelayedDDPG). The state space contains 12 parameters, including real-time load impedance, supply voltage, ambient temperature, output current, PWM frequency, and duty cycle. Each parameter is normalized and mapped to the range [-1, 1]. The action space is a combination of PWM frequency (adjustable range 10-20kHz, 0.1kHz step size) and duty cycle (0-100%, 0.5% step size). The reward function comprehensively considers factors such as system efficiency, stability, and lifespan loss. Specifically, it is expressed as: R = 0.6 × efficiency - 0.3 × |current fluctuation| - 0.1 × temperature rise. Through continuous training and optimization, the matching degree between the driver circuit impedance and the light source load impedance is stabilized at above 95%.

[0047] To further optimize control performance, a dual-loop control architecture was designed. The inner loop uses a PID controller with a 1ms adjustment cycle, adjusting the PWM duty cycle in real time to maintain a constant output current. The PID controller parameters are initially tuned using the Ziegler-Nichols method and then optimized using reinforcement learning. In actual operation, the current control error is within ±1%, ensuring output current stability. The outer loop uses a reinforcement learning algorithm to automatically re-tune the PID parameters based on changes in internal resistance (within a ±20% range) caused by factors such as light source aging. The proportional coefficient Kp is adjustable between 0.8 and 1.2, the integral time Ti is variable between 50ms and 200ms, and the derivative time Td is adjustable between 1ms and 5ms. Furthermore, an adaptive notch filter is employed, whose center frequency closely tracks the PWM baseband frequency. The filter coefficients are automatically adjusted using the LMS (least mean square) algorithm, effectively suppressing electromagnetic interference (EMI) and reducing radiated noise by 15dBμV / m. When tested under different load conditions (such as 20% load, 50% load, and 100% load), the system can quickly adapt to load changes and maintain efficient and stable operation.

[0048] 4. Multi-device phase synchronization control steps

[0049] Synchronization pulse signals are transmitted between the master and slave devices via a dedicated synchronization signal line. This signal line uses low-latency, highly interference-resistant shielded cable (such as Belden 1694A) to reduce signal attenuation and interference during transmission. The master device uses a powerful Raspberry Pi 4B equipped with a quad-core 64-bit ARM Cortex-A72 processor. It generates a block containing a timestamp and control instructions every 500ms, with block size limited to 128 bytes. Timestamps are obtained using a high-precision real-time clock (RTC) chip (such as the DS3231), with an accuracy of ±2ppm. The slave devices verify block validity using a consensus algorithm based on Byzantine Fault Tolerance (PBFT). This verification process is implemented on an STM32F4 series microcontroller and takes no more than 10ms. To improve verification efficiency, the system uses parallel computing technology to simultaneously verify the hash values ​​and signatures of multiple blocks.

[0050] The phase-locked loop (PLL) circuit is integrated into the microcontroller of the slave device and can quickly lock the phase reference broadcast by the master device (with an accuracy of 1μs). During the locking process, a frequency-phase dual-loop control strategy is adopted. The frequency loop is first used to quickly capture the frequency of the master device, and then the phase loop is used to accurately adjust the phase to achieve synchronization of the flashing frequency of multiple device light sources, with the phase difference controlled within 5°. In the camping light networking scenario, the light and shadow synchronization effect is achieved by dynamically adjusting the PWM duty cycle of each device. For example, when the PWM duty cycle of the master light is set to 100%, the duty cycle of slave light 1, which is closer to the master light, is set to 80%, and the duty cycle of slave light 2 is set to 60%. To ensure the accuracy of synchronization, the system calculates the signal transmission delay and device response time each time a control command is sent, and advances or delays the command sending time based on the calculation results to ensure that the synchronization delay does not exceed 100μs.

[0051] Furthermore, a blockchain-based device status consensus mechanism is employed, with each block containing a timestamp, command hash value, device status information, and more. Slave devices receive and verify blocks via specific hardware interfaces (such as SPI), ensuring the immutability and synchronization of command execution order. To prevent blockchain data from becoming excessively large, the system employs a pruning strategy, regularly deleting older, verified blocks that have been executed, retaining only the most recent 100 blocks. In the event of network topology changes (such as device joining or leaving), the system can quickly renegotiate consensus to ensure stable operation of the entire network.

[0052] 5. Fault prediction steps based on vibration modal analysis

[0053] The three-axis vibration data of the equipment is collected through the reserved interface; the signal is decomposed into the intrinsic mode function IMF through empirical mode decomposition (EMD) i(t), where i represents the i-th intrinsic mode function; it is input into the convolutional neural network CNN, and the fault is identified based on the K mutation of the IMF3 kurtosis value; is the normal mean, and ΔK is the set threshold. When the predicted remaining life L≤50 hours, a pre-maintenance strategy is generated through reinforcement learning (RL) to ensure equipment operation.

[0054] The vibration sensor, a TE Connectivity 333AH model, connects to the device's mainboard via a dedicated interface. The acquisition frequency is set to 10kHz to capture subtle vibration changes during device operation. The sensor interface uses the SPI protocol for data transmission, with a transmission rate of up to 10Mbps, ensuring real-time data. The empirical mode decomposition (EMD) algorithm is implemented in the microcontroller using optimized C language code. To improve computational efficiency, parallel computing technology is used to distribute signal decomposition tasks across multiple cores. After decomposing the collected vibration signal into multiple intrinsic mode functions (IMFs), a Hilbert transform is performed on each IMF component to obtain its instantaneous frequency and amplitude information.

[0055] The convolutional neural network (CNN) model consists of three convolutional layers, two pooling layers, and two fully connected layers. The convolutional layers use a 3×3 convolution kernel with a stride of 1 and padding of 1 to extract local features from the vibration signal. The pooling layers use max pooling with a 2×2 kernel size and a stride of 2 to reduce the dimensionality of the features. Training in a laboratory environment using over 100 hours of vibration data, including normal equipment vibration data, bearing failure vibration data, and fan failure vibration data, enables accurate identification of early-stage fault modes in components such as bearings and fans. For example, when a rolling element fails, the kurtosis value of its IMF3 component undergoes a significant mutation. The CNN model can use this information to determine the fault type, achieving a fault identification accuracy of 92%.

[0056] When the remaining life of a device is predicted to be less than 50 hours, the system automatically triggers reinforcement learning to generate a preemptive maintenance strategy. This preemptive maintenance strategy includes preemptively switching to a backup light source module and adjusting the device's duty cycle (e.g., 30 minutes of operation followed by 5 minutes of downtime). To adapt to fault prediction needs in diverse environments, a transfer learning framework is introduced to transfer fault models trained in laboratory environments to complex outdoor environments. Domain adaptation (DA) techniques are used to adjust convolutional layer parameters, specifically employing an adversarial domain adaptation approach. During training, a domain classifier is trained simultaneously to ensure that the feature distributions of the source domain (laboratory environment) and the target domain (outdoor environment) are as close as possible. Federated learning (FL) is also used to update the model at the edge nodes. Each edge node trains the model locally using its own dataset and then uploads the trained model parameters to the server. The server then updates the global model using a weighted average, continuously improving fault prediction accuracy while protecting user device data privacy. In actual outdoor testing, after transfer learning and federated learning optimization, fault identification accuracy remained above 88%.

[0057] 6. Voice Interaction Scenario Perception Steps

[0058] The ambient light sensor and microphone are integrated into the same module. The BH1750FVI digital ambient light sensor is selected, which has a measurement range of 0-100klux and an accuracy of ±3%. By analyzing the intensity and spectral distribution of the ambient light and combining the supply voltage fluctuation characteristics, the hidden Markov model (HMM) is used to identify the user's scene. The HMM model contains 5 states (indoor, outdoor, camping, emergency, and maintenance). By collecting a large amount of data in different scenarios (more than 1,000 sets of data for each scenario) for training, the model's initial state probability, state transition probability, and observation probability matrix are determined. When the light intensity is detected to be less than 300lux and the supply voltage is greater than 3.6V, it is judged to be a camping scene, and the system automatically activates the scenario-based voice command set.

[0059] In "Starry Sky Mode", the RGBW light mixing algorithm achieves a dynamic gradient effect of 16 million colors by adjusting the brightness ratio of the four color LEDs: red, green, blue, and white. Specifically, PWM dimming technology is used to control four independent PWM channels respectively, and the brightness of the LED is adjusted by changing the duty cycle of the PWM signal. To achieve a smooth gradient effect, a linear interpolation algorithm is used. Within each dimming cycle, the brightness value at the current moment is calculated based on the set gradient speed and target brightness. At the same time, to prevent the LED from overheating, the system will monitor the temperature of the LED in real time. When the temperature exceeds the set threshold (such as 60°C), the brightness will be automatically reduced or the speed of the cooling fan will be increased.

[0060] In terms of voice wake-up, the voice wake-up threshold is optimized through reinforcement learning, and it is increased from the initial -42dBFS to -38dBFS. During the training process, the wake-up success rate and false wake-up rate are used as the main indicators of the reward function. Positive rewards are given when the wake-up success rate improves, and negative rewards are given when the false wake-up rate increases. Tests were conducted in noisy environments (such as markets, stations, etc.), and the wake-up success rate increased to more than 95%. At the same time, an adaptive equalizer (EQ) circuit is designed to automatically switch the filtering mode according to the scene recognition results. In the camping scene, the low-frequency boost mode is enabled, and the sound gain of the 20Hz-200Hz frequency band is increased by 6dB to enhance the human voice frequency band in the ambient sound; in the emergency scene, the high-frequency boost mode is enabled, and the gain of the 2KHz-8KHz frequency band is increased by 4dB to highlight the penetration of the alarm command. The EQ circuit is implemented using a digital signal processing chip (such as CS4272), and is implemented through I 2 The S interface transmits data with the microcontroller and realizes fast switching of hardware filter parameters through a specific resistor combination, with the switching time not exceeding 10ms.

[0061] 7. Energy Recycling Reinforcement Learning Steps

[0062] The buck-boost converter uses the Texas Instruments LM5160 chip, which features bidirectional energy flow, a conversion efficiency exceeding 95%, and supports a wide input voltage range (2.9V-60V) and output voltage range (1.2V-60V). When the device is in standby mode, a reinforcement learning algorithm controls the conduction timing of the MOSFET (Infineon IRFS4110TRPBF) to recover the ripple energy (power ≥ 50mW) generated by the light source driver circuit into the energy storage capacitor. The energy storage capacitor uses an AVX tantalum capacitor (model: TPSD180M010R0100) with a capacity of 1000μF and a voltage rating of 10V, which can effectively store the recovered energy.

[0063] An energy management strategy based on stochastic gradient descent (SGD) runs on the microcontroller. By continuously adjusting the adaptive duty cycle (ranging from 0.1% to 5%), it reduces standby power consumption to below 5mW while maintaining the energy storage capacitor voltage at least 3.0V. During the energy management process, the system monitors the energy storage capacitor voltage, the device's standby power consumption, and the status of the external power supply in real time. When the energy storage capacitor voltage reaches 4.2V (fully charged), energy recovery automatically stops. When an external power source is connected, it prioritizes powering the device and charging the energy storage capacitor.

[0064] To achieve more efficient energy management, a multi-objective optimization model (Pareto optimality) was constructed to simultaneously minimize standby power consumption and maximize energy storage efficiency. An ε-greedy algorithm was used to balance exploration and utilization, with the exploration rate ε dynamically adjusted between 0.1 and 0.3. During training, a large amount of training data was generated by simulating different usage scenarios and power supply conditions to optimize the model. During a 30-day standby test, the energy storage capacitor charge retention rate increased from 65% using traditional methods to 89%, significantly extending the equipment's maintenance-free period.

[0065] Comparison of Fault Prediction Model Accuracy

[0066] Data Type Traditional CNN accuracy Transfer learning + DANN accuracy Laboratory data 92％ 93％ Field data (no migration) 78％ 90％ Plateau data (after migration) 75％ 88％ High temperature data (after migration) 70％ 85％ Mixed environment data 65％ 82％

[0067] Traditional CNN and transfer learning + DANN perform differently for different data types. The difference is small for laboratory data. However, transfer learning + DANN achieves higher accuracy than traditional CNN in field data (no transfer), various post-transfer data (high altitude, high temperature, etc.), and mixed environment data, showing a clear advantage.

[0068] Comparison of energy recovery system endurance improvement

[0069]

[0070] In different working modes such as continuous lighting, intermittent work, standby, mixed mode (lighting + communication), extreme environment (low power), etc., the battery life of this application method is improved compared with the traditional method, with the improvement ratio ranging from +50% to +200%, especially the standby mode is doubled.

[0071] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A reinforcement learning driven AI voice IoT intelligent control optimization method, characterized in that: include: Signal spatiotemporal alignment acquisition steps: Stereo sound pickup is achieved through a microphone array. Combined with the Bluetooth module's phase-synchronized clock signal, the generalized cross-correlation delay estimation (GCC-PHAT) algorithm is used to localize the sound source. Adaptive sliding window filtering is applied to the supply voltage, with the window size dynamically adjusted according to voltage fluctuations to suppress ripple noise. Semantic enhancement and parsing of voice commands: For complex keystrokes, a hierarchical state transition network (HSTN) is designed to construct a spatiotemporal correlation matrix between keystroke sequences and voice commands. A voice enhancement algorithm is introduced to improve the signal-to-noise ratio of voice signals in strong background noise environments, and an attention mechanism is used to parse the fuzzy semantics of commands. Reinforcement learning dynamic impedance matching steps: Based on the unique frequency characteristics of the LC resonant circuit, an impedance adaptive adjustment model is constructed; the reinforcement learning algorithm is used to optimize the impedance matching between the driving circuit and the light source load; the driving circuit impedance is set to Z d , the light source load impedance is Z l , define the matching function as: Where Re represents the real part operation, It's Z l The complex conjugate of During the reinforcement learning process, the output PWM pulse width modulation frequency adjustment instruction is adjusted with a step size of 0.1KHz; the initial PWM frequency is set to f0, and the adjusted frequency is f. Through continuous iterative calculation: f = f0 + n × 0.1KHz, where n is the number of adjustment steps, which is dynamically determined by the feedback of the matching function M through the reinforcement learning algorithm; at the same time, in order to suppress electromagnetic interference EMI, an adaptive notch filter is introduced; the center frequency fc of the filter tracks the PWM base frequency f in real time pwm , f c =f pwm ;Suppress electromagnetic interference through adaptive tracking mechanism to ensure stable operation of the system; Fault prediction steps based on vibration modal analysis: collect equipment triaxial vibration data through the reserved interface; decompose the signal into intrinsic mode function (IMF) through empirical mode decomposition (EMD) i (t), where i represents the i-th intrinsic mode function; it is input into the convolutional neural network CNN, and the fault is identified based on the K mutation of the IMF3 kurtosis value; is the normal mean, and ΔK is the set threshold. When the predicted remaining life L≤50 hours, a pre-maintenance strategy is generated through reinforcement learning (RL) to ensure equipment operation.

2. The reinforcement learning-driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: Also includes: Phase synchronization control steps for each device: A synchronization pulse signal is injected through the hardware interface. The master device broadcasts the phase reference, and the slave device locks the PWM output phase through the phase-locked loop (PLL) circuit to achieve synchronization of the flashing frequency of the light sources of each device. In the camping light networking scenario, the light and shadow synchronization matrix is ​​constructed by dynamically adjusting the gradient change of the PWM duty cycle of each device. Voice interaction scene perception steps: Through the integrated ambient light sensor and power supply voltage fluctuation characteristics, the user's camping, emergency, and maintenance scenarios are identified; in the camping scenario, the scenario-based voice command set is automatically activated to trigger the RGBW mixing algorithm to achieve 16 million color dynamic gradients. At the same time, the voice wake-up threshold is optimized through reinforcement learning to improve the wake-up success rate in the usage environment.

3. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: Also includes: Energy recovery reinforcement learning steps: Utilizing the bidirectional energy flow characteristics of the buck-boost converter, when the device is in standby mode, the MOSFET conduction timing is controlled through reinforcement learning to recover the ripple energy of the light source driving circuit to the energy storage capacitor; an energy management strategy based on stochastic gradient descent (SGD) is designed to reduce standby power consumption, while maintaining the energy storage capacitor voltage through adaptive duty cycle adjustment.

4. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the signal time-space alignment acquisition step, the compressed sensing CS technology is introduced to downsample the vibration data. When the sampling rate is reduced to the Nyquist rate, The signal is reconstructed using the orthogonal matching pursuit (OMP) algorithm under the condition of .

5. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the voice command semantic enhancement parsing step, a command reasoning model based on the knowledge graph is constructed, a mapping relationship is established between the professional terminology of the floodlight mode and the hardware parameters, the light intensity distribution after the command execution is predicted through the graph neural network GNN, and the control parameters are preloaded into the device cache in advance.

6. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the reinforcement learning dynamic impedance matching step, a dual-loop control architecture is designed. The inner loop adjusts the PWM duty cycle in real time through a PID controller to maintain a constant output current. The outer loop optimizes the PID parameters through reinforcement learning and automatically readjusts the proportional coefficient, integral time, and differential time when the internal resistance changes due to aging of the light source.

7. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the phase synchronization control step of each device, a device status consensus mechanism is adopted. Every 100 control instructions generated by the master device are packaged into a timestamp and instruction hash value, and written into the device blockchain through a predetermined hardware interface. The slave device verifies the validity of the block to achieve synchronization of the instruction execution order.

8. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the fault prediction step based on vibration modal analysis, a transfer learning framework is introduced to migrate the fault model trained in the laboratory environment to the outdoor environment. The convolutional layer parameters are adjusted through domain adaptation (DA) technology to achieve the required fault recognition rate in the field scenario. At the same time, federated learning (FL) is used to update the model at the edge node to protect the privacy of user device data.

9. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the voice interaction scene perception step, an adaptive equalizer EQ circuit is designed to automatically switch the filtering mode according to the scene recognition result. In the camping scene, the low-frequency boost mode is enabled to enhance the human voice frequency band in the ambient sound; In emergency scenarios, high-frequency enhancement mode is enabled to highlight the penetration of alarm commands, and hardware filtering parameter switching is achieved through resistor combination.

10. The reinforcement learning driven AI voice IoT intelligent control optimization method according to claim 1 is characterized in that: In the energy recovery reinforcement learning step, a multi-objective optimization model is constructed to simultaneously minimize standby power consumption and maximize energy storage efficiency. An ε-greedy algorithm is used to balance exploration and utilization to ensure that the energy storage capacitor charge retention rate in the standby test meets the requirements.

Images