Multi-interaction AI intelligent power bank management system

Abstract

The application discloses a multi-interaction AI intelligent power bank management system, relates to the technical field of portable intelligent power supply and artificial intelligence service scheduling, and is used for solving the problem that multi-interaction AI intelligent power banks are difficult to plan power supply, safely broadcast and multi-service collaborative control under the condition that battery energy and network resources are limited; the system comprises a power supply collaboration module, an audio link module, an AI service scheduling module and a somatosensory power consumption closed loop module. The power supply collaboration module completes protocol handshake and energy state determination and distributes power to generate a current upper limit according to a power scheduling period. The audio link module adjusts audio according to an energy state and an audio branch current upper limit. The AI service scheduling module sorts and degrades service requests in multiple types of budgets according to multi-dimensional loads and situation indexes under the constraint of the energy state. The somatosensory power consumption closed loop module links gesture peripherals according to the energy state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of portable intelligent power supply and artificial intelligence service scheduling, more specifically, the present application relates to a multi-interaction AI intelligent power bank management system. BACKGROUND

[0002] With the popularity of portable charging devices, intelligent sound boxes and translation terminals, users hope that a single device can simultaneously undertake multiple interactive functions such as emergency power supply, multi-lingual voice broadcast, artificial intelligence question and answer, and small games. Existing power banks only have simple power indication and fixed power output, and a small number of products add Bluetooth audio or voice assistant functions, but power supply management, audio processing and artificial intelligence services are often configured independently, lacking unified energy and resource scheduling strategies. When the battery power is insufficient or the external load is large, the common practice is to limit the output current according to a single threshold or directly shut down part of the business, which can easily cause problems such as voice broadcast being interrupted by sudden load, artificial intelligence service being unresponsive for a long time, and user interaction experience being high and low.

[0003] On the other hand, artificial intelligence services themselves have obvious resource fluctuation characteristics, and different lengths of text questions and answers, different levels of translation services, and different output modes have large differences in processing time, energy consumption and network traffic occupancy. Existing solutions usually use fixed answer length, fixed context depth and fixed broadcast strategy for each type of service, and only roughly degrade when the power or network is poor, which cannot sort and differentially control requests according to emergency level, degradability and safety contribution in a multi-service concurrent scenario. At the same time, most devices only provide simple key or touch operations, and lack systematic linkage design between body sensing input, small game load and peripheral power state, resulting in high power consumption in low power or long idle state, and failing to fully utilize energy state information to form a closed-loop regulation at the system level.

[0004] To solve the above problems, the present application provides a solution. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a multi-interaction AI intelligent power bank management system to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] In a preferred embodiment, it comprises: a power supply cooperation module, an audio link module, an AI service scheduling module, a body sensing power consumption closed loop module, and signal connections between the modules;

[0008] The power supply coordination module is used to calibrate the power scheduling period based on the prototype stage, drive the main control chip to periodically perform the fast charging protocol handshake, electric parameter collection and energy state determination, and generate energy state flags and current limit parameters for each branch according to the distributable power;

[0009] The audio link module is used to dynamically adjust decoding, equalization, noise reduction, echo and volume control according to the loudspeaker and power amplifier calibration parameters after reading the energy state and audio branch current limit, so that the audio branch power consumption is constrained by the current limit parameters generated by the power supply coordination module;

[0010] The AI service scheduling module is used to receive artificial intelligence service requests through Bluetooth double-channel communication, establish multi-dimensional load estimation of energy occupation, network occupation and audio occupation based on preset service attribute fields and voice duration, text length, calculate energy tension, network availability risk and emergency information lag risk to form a situation comprehensive index in combination with the remaining running time, network state and emergency request situation, and uniformly sort notification synchronization, translation, question and answer and copywriting update services under the triple constraints of energy budget, network budget and audio budget according to the comprehensive priority, dynamically determine the answer length, context depth, translation service level and output mode;

[0011] The somatosensory power consumption closed loop module is used to convert gestures recognized by the inertial sensor into mute, pause and wake-up control instructions under the constraints of the energy state and somatosensory branch power upper limit given by the power supply coordination module, and adjust the game frame rate and standby or wake-up strategies of audio, display and Bluetooth peripherals according to the energy state linkage.

[0012] In a preferred embodiment, in the prototype stage, a plurality of candidate power scheduling periods are selected, current fluctuations, protocol recovery time and processor occupancy are tested under typical working conditions of step load, terminal plugging and high-power audio start, and a suitable power scheduling period is determined by compromise and written into the parameter configuration table.

[0013] In a preferred embodiment, in each scheduling period, the main control chip completes handshake identification according to a plurality of standard fast charging protocols, determines the available output voltage range and protection current of the external terminal, collects the voltage and current of the battery and each service branch by using a series detection resistor and an analog-to-digital conversion channel, and calculates the current state of charge, remaining running time, total power, safe power upper limit and distributable power in combination with the battery state of charge curve, discharge test and aging and temperature rise test.

[0014] In a preferred embodiment, the energy state is divided according to the energy classification threshold and branch weight pre-calibrated by scene test and user experience, the distributable power is allocated to the external charging, audio, intelligent processing and somatosensory game branches according to the weight, and the current limit parameters of each branch are converted and written into the shared table;

[0015] The audio link dynamically adjusts decoding, equalization, noise reduction, echo and volume control in combination with the calibration results of the speaker and power amplifier after reading the energy state and audio branch current limit.

[0016] In a preferred embodiment, based on the energy state flag output by the power supply coordination module and the constraints of the artificial intelligence service branch power upper limit, a control data channel and an audio channel are established through Bluetooth dual-channel communication. First, the request type, load estimation, real-time level, degradable level and security contribution weight fields are uniformly defined for each artificial intelligence service request. Then, by using the statistical regression results of the relationship between speech duration, text length and processing time, energy consumption and traffic during the development stage, each request is divided into energy occupation, network occupation and audio occupation three-dimensional load estimation.

[0017] In a preferred embodiment, the scheduling end reads the current energy state, remaining running time and its trend from the power supply coordination module, and combines the network state sequence, network fluctuation, number of emergency requests and the last emergency broadcast interval reported by the supporting application, as well as the current audio power occupation and the time and duration of the next mandatory voice feedback from the audio link. Comprehensive calculation of energy tension, network availability risk and emergency information lag risk, construction of situation comprehensive index, and under the triple constraints of energy budget, network budget and audio budget, according to the situation index and the real-time, degradable, security weight and queuing time of each request, the comprehensive priority is calculated, and the notification synchronization, translation, question and answer and text update multi-class artificial intelligence business is uniformly sorted.

[0018] In a preferred embodiment, during specific scheduling, the combination of answer length, context depth, translation service level and output mode is tried for high-priority requests. When the multi-dimensional budget is insufficient, the answer is shortened, the context depth is reduced, the output mode is changed from voice plus text to only text, and the answer is postponed or discarded according to the preset degrading path. Combined with the prediction of future network available window and local multi-language emergency template library, high-level translation and question and answer are concentrated when the network condition is good, and local templates are used to cover rigid instructions when the network is poor for a long time. At the same time, according to the audio power and mandatory broadcast time window, the voice synthesis interlude is dynamically tightened, only the silent text output is reserved, and finally the instant display and broadcast are completed through the way of encrypted transmission and only processing text in volatile storage.

[0019] In a preferred embodiment, under the energy state given by the power supply coordination module and the upper limit of the body sense branch power, the body sense power consumption closed loop module establishes linkage control between the body sense input, small game load and peripheral power supply: the inertial sensor working frequency is selected through the comparison of recognition accuracy, response time delay and power consumption under multiple sampling frequencies, and in operation, posture data is collected at the inertial sensor working frequency and matched with the preset gesture template, and the hand waving, shaking and turning actions are converted into mute, pause and wake-up control instructions.

[0020] In a preferred embodiment, the small game frame rate, body sense sampling frequency and the switching and standby or wake-up strategy of audio, display and Bluetooth peripherals are adjusted dynamically according to the energy state, and high-power services are automatically degraded or closed when there is insufficient energy and no operation for a long time or high load, and after being woken up, the energy state is re-evaluated to selectively restore the services.

[0021] The technical effects and advantages of the multi-interaction AI intelligent power bank management system of the application are as follows:

[0022] The application introduces the power scheduling period, electrical parameter sampling and energy state grading into the power supply coordination module, unifies the external charging, audio, artificial intelligence processing and body sense game branch under the constraint of allocable power, provides a power supply boundary with a current upper limit for subsequent modules, and realizes the construction of a unified energy control benchmark from the hardware level. Based on this benchmark, the audio link module adjusts the decoding, equalization, noise reduction, echo and volume control in coordination after reading the energy state and the current limit of the audio branch, balances the sound quality and power consumption in the same framework, and reduces distortion and protection actions caused by sudden audio load.

[0023] More importantly, the AI service scheduling module maps the service request to a three-dimensional load of energy occupation, network occupation and audio occupation under the constraints of energy state and artificial intelligence service branch power upper limit, constructs a situation comprehensive index combining residual running time, network state evolution and emergency request accumulation, and performs unified sorting and hierarchical degradation under the triple constraints of energy budget, network budget and audio budget, realizes the multi-service collaborative control of "allocating resources according to the degree of urgency and degradability", and has stronger adaptability and interpretability compared with the single-dimensional current limiting according to only power or network. The body sense power consumption closed loop module further utilizes the energy state to drive the linkage of gesture recognition, small game frame rate and peripheral standby or wake-up strategy, forms a closed loop between the body sense interaction and the whole machine power consumption management, and thus significantly improves the endurance and system stability on the premise of ensuring key broadcasting and necessary interaction. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The module schematic diagram of the multi-interaction AI intelligent power bank management system of the application is shown in the figure.

[0025] Figure 2 This is a timing diagram of a multi-interactive AI smart power bank management system according to the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example:

[0028] This invention discloses a multi-interactive AI smart power bank management system, such as... Figure 1 As shown, it includes: a power supply coordination module, an audio link module, an AI service scheduling module, a motion sensing power consumption closed-loop module, and signal connections between the modules.

[0029] First, in the power supply coordination module, the main control chip performs cyclic control around the preset power scheduling cycle Ts. In each cycle, it sequentially completes protocol handshake status maintenance, electrical parameter sampling, energy state quantity calculation, branch power allocation, and power supply parameter broadcasting.

[0030] The power scheduling period Ts is not arbitrarily set, but is calibrated during the prototype stage based on the dynamic response speed of the battery cell and the computing power of the MCU. Specifically, for example, a set of candidate periods such as 10ms, 20ms, 50ms, and 100ms are selected. Under typical operating conditions such as load step, external terminal plugging and unplugging, and instantaneous high-power audio startup, the peak value of battery current fluctuation, protocol recovery time, and MCU utilization rate are measured. By comparing the trade-off between control response speed and computing load, a period value that can ensure that the output current fluctuation does not exceed the safety threshold and that the MCU utilization rate does not exceed the preset ratio is selected. Finally, this value is written into the parameter configuration table in the form of a configurable parameter and loaded as Ts during the firmware initialization stage.

[0031] When an external terminal is detected to be connected via a wired interface, the main control chip controls the multi-protocol fast charging chip to sequentially initiate standardized handshake sequences such as Apple 2.4A, USB BC, QC2.0 / QC3.0, and USB PD2.0 / PD3.0. Specifically, the main control chip controls the data pin level, voltage step, or CC pin resistor identification logic according to the timing specified in the chip datasheet, and reads the handshake result register of the fast charging chip through I²C or a dedicated status pin during the handshake process. When the handshake success flag for a certain protocol is set, the main control chip records the set of available output voltages corresponding to that protocol as follows: The protection current of this protocol given in the chip datasheet is denoted as... .

[0032] wherein, each voltage level and the specific value of the fast charging chip specification and the corresponding charging protocol standard, and in the product design stage, these data are fixed as parameters, and are not dynamically derived in the running process. Then the master control chip selects the target output voltage in the range of according to the external terminal category and the scene needs, and writes the control register of the fast charging chip to make the DC-DC converter output the matching voltage level in the fixed voltage interval.

[0033] In each power scheduling period Ts, the master control chip collects the battery terminal voltage , the battery terminal current , the external charging branch current , the audio branch current , the AI / Bluetooth / application processing branch current and the somatosensory and small game branch current in turn through the built-in ADC. The detection method of these currents is to connect a detection resistor in series in the corresponding branch, and then send the voltage drop across the detection resistor to the ADC for conversion to digital quantity. The resistance value of the detection resistor is selected according to the expected range and ADC resolution in the power supply design stage.

[0034] Then the master control chip constructs the SOC lookup table according to the SOC-voltage characteristic curve provided by the battery supplier, and modifies the lookup table in the prototype stage combined with the measured data of different discharge rates: through the constant current discharge experiment, the corresponding relationship between and the real SOC is recorded, and the manufacturer's curve is offset and scaled to make the lookup table accurately reflect the SOC in the working current range of the product. In the running time, the master control chip converts the lookup table to SOC.

[0035] In estimating the remaining running time , the master control chip is based on the nominal capacity of the battery and the current total discharge current , combined with the capacity attenuation under different discharge rates to fit the empirical coefficient , and introduces the minimum current constant to avoid division by zero , and constructs the following approximate relationship:

[0036] ;

[0037] wherein directly comes from the battery specification, through full discharge test under a variety of load currents in the prototype stage, the nominal capacity x SOC / The deviation between the actual discharge end time and the actual discharge end time was obtained by fitting using methods such as least squares. Constrained by both the ADC noise level and the accuracy of the sensing resistor, under static load conditions, by measuring the jitter range of the current reading, a value slightly larger than the jitter amplitude is selected so that the formula remains stable when the actual current is close to zero.

[0038] Subsequently, the main control chip also adjusts the sampling data in each cycle. and Calculate the current total output power The power of each branch is estimated based on the current of each branch, as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] in Provide power supply voltage for the audio amplifier. The power supply voltage for the system logic and AI processing unit is selected by the power engineer based on the chip's operating voltage range during the DC-DC power architecture design phase and confirmed through circuit simulation and temperature rise experiments. It is then used as a constant in the firmware calculation.

[0045] To ensure heat dissipation and safety, the main control chip... Reserve a safety margin power on the basis . By conducting long-term high-load aging tests on the prototype and measuring the junction temperature rise of the device under different ambient temperatures, the maximum average power allowed without exceeding the safe junction temperature is calculated. This average power is then rounded down by a certain safety factor to represent the upper limit of the safe total power allowed under the current battery operating conditions. Based on this, the main control chip calculates the available power that can be used for business load allocation before exceeding the thermal safety boundary. :

[0046] ;

[0047] In terms of state of energy determination, the main control chip simultaneously reads the battery temperature in each cycle. , with SOC and Together they are mapped to energy state flags .

[0048] in, The grading thresholds SOC1, SOC2, t1, t2, T1, T2 are not made out of thin air, but are determined through a series of experiments in the prototype stage: first, according to the recommended discharge cutoff SOC and the recommended safe power range in the battery specification book, the initial SOC1, SOC2 is given, and then in the constant power discharge and hybrid load scene, the user's endurance experience at different SOC and the timing of device protection action are recorded, and the initial SOC threshold is corrected; t1, t2 are determined by statistics of the remaining time that users can accept in typical use scenarios and the time point that must prompt power saving; T1, T2 are corrected according to the upper limit of the safety temperature in the battery and power device specification book, and combined with the shell temperature rise measured by the thermal imager, to ensure that long-term operation below T2 will not exceed the rated junction temperature of the chip and the cell. The final grading rule can be described as:

[0049] , when SOC≥SOC1 and and ;

[0050] , when SOC2≤SOC<SOC1 or and ;

[0051] , when SOC<SOC2 or and ;

[0052] , when or overvoltage / overcurrent fault condition occurs;

[0053] It should be noted that the above threshold values are stored in the parameter configuration table, and can be updated through one-time calibration before mass production according to different battery models.

[0054] In specific implementation, the main control chip determines in fixed priority order: first, it is determined whether is met, then it is determined whether is met on the premise that is not met, and then it is determined whether is met on the premise that and are not met, and the remaining cases not covered by the above conditions are uniformly determined as . Through the above determination order, it is ensured that any given combination of SOC, and corresponds to only one value in operation, avoiding ambiguity in state division.

[0055] After completion , the main control chip needs to allocate power to the external charging branch, audio branch, AI branch and somatosensory / gaming branch under the constraint . To this end, the main control chip sets weights , , , for each branch, which are not subjectively specified, but are calibrated through scene evaluation and user experience testing: in heavy charging scenes, heavy audio scenes, heavy AI interaction scenes, heavy gaming scenes and other typical working conditions, by adjusting the weight combination, recording the user's subjective score on charging speed, sound quality, AI response speed and endurance, and combining energy consumption data analysis, a set of weights with the best balance in most scenarios is selected. Finally, a set of weights corresponding to different is written into the parameter configuration table, and the main control chip selects the corresponding set of weights according to the current at runtime. In each cycle, the main control chip calculates the target power of each branch according to the weights in proportion :

[0056] ;

[0057] ;

[0058] ;

[0059] ;

[0060] Then convert the target power to the maximum current upper limit of each branch:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] Among them , and are the hardware current upper limit of each branch calculated according to the maximum working current given in the chip manual, the PCB wiring allowed current and the safety margin. These values are determined through schematic design and power simulation, and are fixed in the parameter table after being verified to be normal in the prototype current test. On the external charging branch, must not exceed , to ensure that the hardware protection capability of the fast charging chip is not exceeded.

[0066] After that, at the end of each scheduling cycle, the master chip will write the latest and , , power constraints parameters into the shared power supply parameter table, and open it to the query interface of the audio processing logic, Bluetooth and AI collaborative logic, and motion and small game logic.

[0067] The audio link module runs under the energy state and audio branch current limiting constraints provided by module one, and is responsible for completing the complete signal link from audio data reception, decoding, equalization to speaker output, as well as noise suppression and echo cancellation. After receiving the audio playback instruction from the Bluetooth classic audio channel or the matching application, the master chip first reads the current and from the shared power supply parameter table, and derives the allowed maximum output volume, voltage swing, and whether full-amount active noise reduction logic can be enabled according to and the rated power of the speaker. The rated power and frequency response range of the speaker are given by the acoustic test report of the speaker supplier, and are written into the acoustic parameter table through actual listening and power amplifier load test in the acoustic design stage; the power supply voltage is determined by the recommended working voltage range of the audio chip manual, and is fixed after THD+N test and efficiency test verification.

[0068] In the specific implementation process, the master chip writes the compressed audio data into the decoding chip input buffer according to the encoding format and sampling rate supported by the audio chip data manual, and triggers the decoding chip to perform decoding of MP3, WAV, FLAC, AAC, etc.

[0069] After that, the left and right channel digital audio streams output by the decoding enter the equalization processing path, and the frequency band division and initial gain coefficient of the equalizer are determined by the frequency scanning test data of the prototype in a semi-anechoic environment: the output sound pressure level of the speaker at different frequencies is measured by the engineering personnel, and the intrinsic response curve of the speaker is fitted, and according to the curve, each frequency band is allocated a compensation gain, so that the synthesized frequency response is as close as possible to the target listening curve.

[0070] The master chip makes dynamic correction to the gain parameters according to the current and : when is or and is large, the complete multi-band equalization curve is allowed to be used, and is opened to the user for fine tuning in the matching application; when is or the audio branch current is close to When the peak power is detected, the master chip automatically reduces the low frequency and overall gain amplitude to reduce the peak power, ensuring that the audio does not trigger the power supply protection. The balanced digital audio is converted into an analog signal in the digital-to-analog converter inside the decoding chip and sent to the dual-channel power amplifier circuit to drive the loudspeaker to output sound.

[0071] In the active noise reduction part, the master chip controls the microphone to collect environmental sound. The sensitivity, frequency response range and noise floor parameters of the microphone are given by the supplier's data manual and the prototype calibration experiment, which are recorded during hardware selection. According to the calibration experiment of white noise in a semi-anechoic chamber, the master chip constructs an environmental noise statistical characteristic model and generates a cancellation signal using an adaptive filter algorithm. The order, step factor and other parameters of the filter are determined by comparing the steady-state noise attenuation and convergence speed under different parameter combinations in the laboratory, and are stored as adjustable parameters in the firmware. After calibration, a noise attenuation of not less than 20dB can be achieved in the typical mid-frequency band, which is given as a performance indicator in the specification, but is still supported by the filter parameter curve in actual execution.

[0072] When hands-free calling, the master chip inputs the speaker output signal as the echo reference and the microphone sampling signal as the mixed signal containing echo, and configures the length and update step of the adaptive filter according to the impulse response characteristics of the echo path. These parameters also come from tests in different installation environments, by measuring the time delay and attenuation distribution of the acoustic path between the speaker and the microphone, the filter length covering the main echo energy is selected. After echo cancellation, the remaining speech signal is returned to the mobile terminal through the Bluetooth classic channel to ensure that the opposite end hears clear speech.

[0073] In the entire audio link, the master chip continuously monitors the audio branch current and When detecting Changes or Approaching , immediately reduce the output volume, simplify the equalization curve, and shut down part of the noise reduction frequency band to pull the audio instantaneous power back into the power budget allocated by the module.

[0074] It should be noted that the multi-interaction AI smart power bank of the present application is configured to be carried by the cross-regional high-speed train crew, and is used to provide emergency lighting, multi-lingual voice broadcast, real-time translation and line information query for the crew when the train breaks down in a long tunnel, there is no city power supply outside the train, and the fixed broadcast and guidance system in the train is partially disabled. However, when the train stops at a location inside a long tunnel in the mountains, the external cellular network signal is intermittently available due to tunnel attenuation, and the on-board emergency power supply needs to prioritize the operation of train control, signaling and a small number of emergency lighting devices. The multi-interaction AI smart power bank carried by the crew meets the railway carrying specifications in terms of capacity and total battery energy, and is used as a temporary multifunctional emergency terminal, which simultaneously bears the following loads: on the one hand, it supplements the power of intercoms, handheld terminals and other devices through wired interfaces to maintain the communication capability of the crew and the train dispatch center; on the other hand, it uses the built-in loudspeaker to conduct multi-lingual emergency broadcasting, and cyclically broadcasts evacuation notices and psychological soothing voices for passengers in different countries and regions in the train compartment; at the same time, it also provides the crew with question and answer capabilities based on GPT-type large models for quickly querying the current train operation plan, tunnel structure, location of the nearest emergency evacuation exit, and soothing words for different passenger groups, and for two-way communication through real-time translation function when facing passengers who do not speak the local language.

[0075] Since the train may have been running for a long time when the fault occurs, the power state of the multi-interaction AI smart power bank carried by the crew is not stable, and some devices have been used for music playing, gaming and daily charging before the incident, so the remaining available energy is limited. At the same time, the network condition in the tunnel fluctuates dramatically with the train position and the progress of the outside repair, and the bandwidth and latency of the mobile terminal when accessing the network through the on-board relay or the temporarily opened emergency base station differ significantly in different time windows. In order to ensure that the emergency broadcast in the train compartment does not interrupt, and to maintain the availability of translation and AI question and answer as much as possible under the conditions of insufficient power and poor network, simply relying on AI service configuration with fixed length answers, fixed context depth and fixed voice broadcast strategy may cause the following problems:

[0076] In some time windows, long text answers or large-scale context reasoning result in significant increase in reasoning time and power consumption, triggering the power bank to drop power too quickly or even triggering overcurrent protection, causing multi-lingual broadcast to be forced to interrupt; in another part of the time window, although the power is sufficient, the network condition is poor, and the cloud model interface is still called according to the high-level service parameters, causing request queuing or timeout failure, and passengers cannot obtain translation and question and answer feedback at critical moments.

[0077] In addition, in the above scenario, the noise level in the carriage is high, and the passengers' emotions are easily fluctuated. The multi-lingual voice broadcast and intercom communication often need to be alternately performed. When the power bank simultaneously undertakes the tasks of loudspeaker high-power broadcast, real-time translation voice synthesis, and power supply for external terminals, if the multi-lingual voice broadcast is simply superimposed without considering the current audio load and the power occupation of the power supply branch, the audio branch is likely to exceed the safe current upper limit, triggering the power amplifier protection or causing the output sound distortion, further intensifying the passengers' uneasy emotions. Therefore, in such a tunnel failure scenario, it is impossible to perform unified load estimation and service level scheduling on the translation, GPT type question and answer, and notification broadcast, etc. AI services, to comprehensively sort according to the real-time performance and degradability, and dynamically select different output modes such as text display, short voice broadcast, or only text prompt, so as to reliably maintain the key voice notification and the most necessary AI interaction.

[0078] Therefore, in the embodiment, as shown in Figure 2 , the AI service scheduling module takes the Bluetooth 5.0 dual-mode communication unit as the communication basis, and under the constraints of the energy state and the AI branch power upper limit , instead of simply limiting the service according to a single power or network condition, the energy state, the remaining running time, the network state evolution trend, the audio branch load, and the service type are comprehensively modeled in the special scenario of the long tunnel train failure, a situation comprehensive index reflecting the emergency degree of the scenario is constructed, and is taken as the core to perform multi-factor comprehensive analysis and joint scheduling on the AI services such as notification synchronization, real-time translation, GPT type question and answer, and multi-lingual emergency broadcast script updating, dynamically selecting the service level and the output form within the given multi-dimensional resource budget, so as to preferentially guarantee the continuous availability of the key emergency voice notification and the necessary AI interaction under the constraints of limited battery energy, fluctuating network state, and changing audio load.

[0079] Specifically, after the device completes the Bluetooth secure pairing and establishes the connection, the host chip allocates logical connections for the BLE low-power data channel and the classic Bluetooth audio channel according to the configuration process in the Bluetooth chip data manual. The former is used to transmit control instructions, notification texts, translation texts, and AI answer texts, and the latter is used to transmit music playing, call voice, and multi-lingual emergency broadcast audio stream. The host chip plans the BLE connection interval, MTU size, and other parameters according to the Bluetooth specification, which are determined through the throughput and power consumption relationship test in the development stage and stored in the firmware in the form of a configuration table.

[0080] Under the tunnel failure scenario, the host chip first determines the , the current remaining running time and its trend in the recent several cycles , the time interval since the last full-vehicle emergency broadcast , the number of requests marked as emergency in the service queue and type distribution, and the network status sequence periodically reported by the supporting application , to construct a comprehensive situation index . Specifically, and are mapped to the power tension component , the current value and short-term fluctuation amplitude of are mapped to the network uncertainty component , and and are mapped to the emergency information lag component , using the weight coefficients obtained through expert evaluation and simulation tuning in the development stage , , to construct:

[0081] ;

[0082] wherein the calculation methods of the above three components are: According to and are mapped to four levels of abundance, acceptable, tension, and extreme tension, and are further quantified as continuous values in the interval [0, 1]; According to the distribution of in the recent multiple service cycles, packet loss rate, and RTT jitter, network availability risk is quantified; According to and to assess whether there are long-time uncovered emergency languages or car areas, to reflect the lag risk of emergency information release. , , and , , The specific mapping relationship and weight value of are calibrated in the development stage by replaying multiple failure scenarios, simulating different broadcast and interaction strategies, and responding to the effects of the crew and passengers, and are fixed in the form of lookup table and parameters in the parameter configuration table.

[0083] In a specific implementation, different combinations of , can be mapped to segmented linear values in the range of 0~1 through a preset lookup table, for example, several corresponding to Interval mapping to 0.0~0.3, will be mapped to 0.3~0.6, will be mapped to 0.6~0.9, will be mapped to 0.9~1.0, will be close to the boundary. and Also can be mapped to a continuous value in the range of 0~1 by a similar lookup table according to , packet loss rate, RTT jitter and , , so that the component calculation logic of can be directly completed by table lookup value in engineering implementation.

[0084] When constructing a service request, the master chip defines a unified service request structure, each request containing request type , load estimation , real-time level , degradable level and security contribution weight and other fields, among which , and are mapped by the upper application according to the service type during configuration: for example, the evacuation path update and danger area warning facing the whole vehicle are marked as the highest and the highest , the lowest ; the soothing question facing a single passenger has a higher but can be moderately degraded; the general chat type question has a lower and a higher .

[0085] In terms of load estimation, it is still obtained by statistical regression on a large number of sample services: during the development stage, the engineering personnel record the average CPU occupation, delay and traffic consumption of different length texts, different length speeches triggering ASR, translation or large model inference on the mobile terminal side, establish the function relationship of inference time consumption and energy consumption relative to speech length , text word number , and obtain the coefficients α and β by linear or piecewise linear regression. At runtime, the load of each request is estimated by the following relationship:

[0086] ;

[0087] Among them is obtained by detecting the time difference between the user pressing the translation or AI button and stopping speaking, The specific values ​​of α and β are directly calculated from the output text length of the ASR engine or the user-input text length, and are stored in the load estimation parameter table, which can be iteratively updated based on subsequent operational statistics. To more precisely reflect the network and audio resource usage of different services, the main control chip further... Split into energy load Network load and audio load Three components: It is obtained from the above formula; Based on whether the request requires access to the cloud model, the maximum text length limit, and the current... Obtained through mapping; This is determined based on whether the request requires speech synthesis and playback, the expected speech duration, and the parameters in the current audio insertion strategy table.

[0088] Furthermore, within each power scheduling cycle or an integer multiple thereof, the main control chip reads the current power supply parameter from the power supply parameter table of module one. and , and according to Calculate the upper limit of instantaneous power available for the AI ​​branch. Based on the average power curve for the AI ​​inference process, the main control chip uses a pre-calibrated function. Obtain the energy budget for the current period ;

[0089] In one alternative implementation, function g is implemented using a lookup table, that is... Mapped to corresponding power ranges as predefined. The values ​​are assigned to each power range. The lookup table was obtained by replaying typical AI workload scenarios during the prototype stage, statistically analyzing the average power curve of the AI ​​inference process, and calibrating it in conjunction with the allowable temperature rise margin. This lookup table was written into the parameter configuration table before mass production and can be directly accessed during runtime. Obtained by looking up the table within the specified interval .

[0090] Supporting applications are based on the most recent several cycles The evolution trend predicts the available network capacity range over multiple future time windows and the available network budget for the current period. The metadata is sent to the device; the audio link module then calculates the audio budget available for AI voice insertion within the current business cycle based on the current emergency broadcast schedule, call status, and audio branch current occupancy. Therefore, the AI ​​business scheduling module is simultaneously subject to the energy budget during scheduling. Network budget and audio budget Triple constraints.

[0091] To perform the comprehensive ranking under the above multi-dimensional constraints, the master chip calculates a comprehensive priority score for each request in the service queue . The normalized real-time , degradability , security contribution , request waiting time and situation indicators jointly determine:

[0092] ;

[0093] wherein is the request waiting time normalization value in the queue, λ1~λ5 are the weight coefficients calibrated by simulating fault scenarios and evaluating the crew in the development stage. When is high, the last term significantly raises the priority of the emergency class request with high security contribution, which can obtain resources within a certain range even if is large; when is low, the system considers the trade-off between , and more, and takes into account general question and answer and soothing interaction.

[0094] Then, in the scheduling implementation, the master chip first ranks the requests from high to low according to in each service cycle, and then maintains the remaining budgets , and in three dimensions of , and respectively, and adopts a multi-dimensional resource-aware heuristic selection process: for each request after ranking, the master chip attempts to match a service level combination for it based on the current , and , including the translation service level , the maximum answer length , the context depth and the output mode , and thus obtains , , under the combination. If the combination satisfies , , , the request is confirmed to be selected with the level, and the corresponding resources are deducted from the budget; if it cannot be satisfied, the request is rejected according to Try downgrading the request from high to low, such as shortening the response length, reducing the context depth, or canceling voice prompts and only keeping the text prompts; if the budget still cannot be met within the maximum allowed downgrading range, postpone the request to the next business cycle or discard it if necessary.

[0095] in , and These represent the remaining available portions of the energy budget, network budget, and audio budget, respectively. , and These represent the estimated resource consumption of the request in the three dimensions of energy, network, and audio, respectively.

[0096] In specific translation work, Not only by With the present The decision also incorporates a future network window predictor. The supporting application establishes a periodic network availability model by statistically analyzing recent network availability records within the tunnel. When a brief window of good network availability is anticipated, it will... Attached to the request metadata.

[0097] AI business scheduling module higher and When the upcoming network access window appears, you can choose to temporarily use a medium-level translation service to quickly cover basic multilingual evacuation instructions. Then, when the network access window arrives, trigger a higher-level translation and GPT-type Q&A session in batches to obtain more detailed reassurance statements and route information. If the network is expected to be poor for an extended period, the system will prioritize calling the locally pre-built multilingual emergency template library and only trigger cloud-based inference for a very small number of critical issues. This will ensure multilingual coverage of basic information while suppressing network load and energy consumption.

[0098] For AI question-answering services, the main control chip still relies on [the specific technology / mechanism] when constructing the request. and The maximum answer length is calculated based on the mapping relationship. and context depth , , .

[0099] In one alternative implementation, and This can also be achieved using lookup tables or piecewise linear mappings, for example, based on... Classify battery status as , and Three gears, press again within each gear. The interval is Set different upper limits, and Only with The changes were determined during the prototype stage by simulating the impact of different response lengths and context depths on inference latency and energy consumption, and the content of the above lookup table was then fixed in the parameter configuration table.

[0100] In tunnel failure scenarios, the AI ​​business scheduling module further... Introduction and Selection logic: when When the request type is high and the question type is related to safety tips or evacuation routes, even if In It also allows for a moderate increase and To ensure complete interpretation of instructions; when Even when the request is low and is a typical chat-style Q&A, for The system also avoids consuming excessive resources on long conversations, instead reserving more resources for potential emergency broadcasts or translation requests. Through these methods, under the same battery level, AI question-answering behavior exhibits significantly different resource consumption patterns in different situations.

[0101] In terms of channel and audio resource coordination, the main control chip obtains the current audio decoding and playback load through the status interface of module 2, including whether multilingual emergency broadcasting is currently in progress, whether the crew intercom is in progress, and the current audio output power occupancy ratio. The audio link module writes relevant statistical data to the shared status area at the end of each audio frame.

[0102] Meanwhile, the audio link module estimates the latest start time for the next mandatory emergency voice broadcast based on the emergency broadcast script and the crew operation plan. and expected duration .

[0103] The AI ​​business scheduling module performs calculations. At that time, and Also consider: For AI responses requiring speech synthesis and broadcasting, only allow their estimated broadcast duration to not encroach on the time window of the next mandatory emergency broadcast; when the audio branch power is close to... When the allowed upper limit is too high or the time before the next mandatory broadcast is close, the AI ​​scheduling module will force all new requests to be processed. It is downgraded to text only, and the response text is transmitted via BLE to the mobile terminal or the small screen of the power bank for display without triggering any new voice interruption, thereby avoiding mutual obscuring or branch overload between emergency voice and AI voice.

[0104] In terms of security, the AI service scheduling module uses the aforementioned Bluetooth security pairing and AES-128 encryption and decryption mechanism. All notification texts, translation texts, and AI question and answer texts transmitted through the BLE channel are sent after being encrypted by the supporting application. The device uses the negotiated key to decrypt in the receiving path. The decrypted plaintext exists only in the RAM for immediate display and announcement and is not written into the flash memory or long-term log file.

[0105] The somatosensory power consumption closed-loop module relies on the and , a linkage relationship is established between somatosensory input, game load, and the power state of the whole machine. On the one hand, somatosensory triggering of the audio link module and the AI service scheduling module is achieved through inertial sensor sampling and gesture recognition. On the other hand, the game frame rate and sensor sampling frequency are dynamically adjusted according to the energy state Estate, and the external device is turned off and awakened according to Estate to achieve low-power operation of the whole machine.

[0106] In the somatosensory link, the sampling frequency of the inertial sensor is not fixed and written, but is determined during the prototype stage by comparing the relationship between gesture recognition accuracy, response delay, and power consumption at different sampling frequencies. Engineers collect a large number of hand waving, shaking, and turning gestures at sampling frequencies of 10Hz, 25Hz, 50Hz, 100Hz, etc. to train and evaluate the accuracy and response time of the gesture classification algorithm, record the current consumption curve of the inertial sensor at each frequency, and finally select the frequency with the lowest power consumption under the premise of meeting the experience requirements of recognition rate and response speed as the main sampling frequency, and record it in the parameter table.

[0107] The main control chip reads data from the three-axis acceleration sensor and three-axis gyroscope at this frequency during operation. The digital filter selected according to the noise spectrum characteristics during the design stage removes high-frequency noise and temperature drift. The gravity component separation algorithm is used to obtain a feature sequence reflecting dynamic posture changes, which is compared with gesture templates such as hand waving, shaking, and turning established during the algorithm training phase. The similarity threshold is determined by ROC curve analysis results on the training and validation sets to ensure that the false trigger rate and missed detection rate are within the expected range. When the similarity exceeds the threshold, the main control chip determines that the corresponding gesture is triggered, executes the corresponding animation or control action locally, and generates a somatosensory event to the module two or module three for silent, pause, or wake up AI, etc.

[0108] In the small game scheduling part, the basic frame rate of the lightweight 2D game engine is determined by the human visual persistence experiment data and game experience test: by letting users experience the same game at different frame rates, the subjective fluency score is counted, and 30 frames per second is selected as the lower limit to ensure that most users feel no stuttering. The main control chip adjusts the game frame rate according to the energy state Estate and the somatosensory sampling frequency of the inertial sensor, and controls the external device to be turned off and awakened according to Estate to achieve low-power operation of the whole machine. In the state, the game engine is allocated enough processing time slice and display refresh task to make the game run at the frame rate. In the state, the game engine is allocated enough processing time slice and display refresh task to make the game run at the frame rate. In the state, the game engine is allocated enough processing time slice and display refresh task to make the game run at the frame rate. In the state, the game engine is allocated enough processing time slice and display refresh task to make the game run at the frame rate.

[0109] In the low-power management of the whole machine, the main control chip periodically polls the state registers of the battery management chip, the Bluetooth chip, the audio decoding chip, the display driving chip and the inertial sensor through the I2C and SPI buses, and the polling period value is calibrated according to the compromise between the state change speed of the external device and the polling consumption in the development stage.

[0110] When the main control chip detects that there is no key or touch event within a certain time window, and there is no continuous audio stream or large amount of data transmission in the Bluetooth link, the standby process is triggered in the or state. The standby process includes: closing the power supply of the audio power amplifier and decoding chip according to the recommended low-power configuration in the chip manual, turning off the display screen or adjusting it to the lowest brightness, putting the Bluetooth chip into the low-power broadcast or sleep mode, switching the inertial sensor to the low-frequency or event-triggered sampling mode, and only retaining the RTC timer and a few wake-up interrupt lines.

[0111] Finally, when any peripheral device generates a wake-up event, such as user key, body gesture or external terminal initiated Bluetooth connection request, the main control chip first responds to the corresponding interrupt, and then immediately calls module one to execute energy state determination and power allocation once, to obtain new and upper limit of each branch power, and then decide whether to restore audio playback, AI conversation or game task according to the new energy state. If has been restored to or , the main control chip gradually powers up the related peripherals according to the recovery order and delay configuration in the parameter table of each module to prevent inrush caused by simultaneous power-on; if it is still in or , only the necessary basic interaction and alarm display are restored, and starting long-time and high-load business is prohibited.

[0112] The above formulas are dimensionless to calculate their numerical values, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0113] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product.

[0114] Those skilled in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and the constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0115] In addition, each functional module in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0116] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0117] Finally, the above is merely preferred embodiments of the present application, and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A multi-interactive AI smart power bank management system, characterized in that, include: Power supply coordination module, audio link module, AI service scheduling module, motion sensing power consumption closed-loop module, and signal connections between modules; The power supply coordination module is used to drive the main control chip to periodically perform fast charging protocol handshake, electrical parameter acquisition and energy status determination based on the power scheduling cycle calibrated in the prototype stage, and generate energy status flags and current limit parameters for each branch according to the allocable power. The audio link module is used to read the energy status and audio branch current limit, and then dynamically adjust the decoding, equalization, noise reduction, echo and volume control according to the speaker and power amplifier calibration parameters, so that the power consumption of the audio branch is constrained by the current limit parameters generated by the power supply coordination module. The AI ​​business scheduling module receives AI business requests via Bluetooth dual-channel communication. Based on preset business attribute fields and voice duration and text length, it establishes a multi-dimensional load estimate of energy consumption, network consumption, and audio consumption. Combining the remaining runnable time, network status, and emergency request status, it calculates energy stress, network availability risk, and emergency information lag risk to form a comprehensive situational indicator. Under the triple constraints of energy budget, network budget, and audio budget, it uniformly sorts notification synchronization, translation, Q&A, and text update services according to comprehensive priority, and dynamically determines the response length, context depth, translation service level, and output mode. The motion-sensing power consumption closed-loop module is used to convert gestures recognized by the inertial sensor into mute, pause, and wake-up control commands under the constraints of the energy status and the upper limit of the power of the motion-sensing branch given by the power supply coordination module, and to adjust the frame rate of the mini-game and the standby or wake-up strategies of audio, display, and Bluetooth peripherals according to the energy status.

2. The multi-interactive AI smart power bank management system according to claim 1, characterized in that: During the prototype stage, multiple candidate power scheduling cycles were selected, and current fluctuations, protocol recovery time, and processor utilization were tested under typical operating conditions such as step load, terminal plugging and unplugging, and high-power audio startup. The appropriate power scheduling cycle was then determined and written into the parameter configuration table.

3. The multi-interactive AI smart power bank management system according to claim 2, characterized in that: During each scheduling cycle, the main control chip completes handshake identification according to multiple standard fast charging protocols, determines the available output voltage level and protection current of the external terminal, and uses series detection resistors and analog-to-digital conversion channels to collect the voltage and current of the battery and each service branch. Combined with the battery state of charge curve, discharge test, and aging and temperature rise test, it calculates the current state of charge, remaining operating time, total power, safe power limit and allocable power.

4. The multi-interactive AI smart power bank management system according to claim 3, characterized in that: Based on the energy classification threshold and branch weight determined in advance through scenario testing and user experience calibration, the energy status is divided, and the allocable power is allocated to the external charging, audio, intelligent processing and motion-sensing game branches according to the weight, and converted into the current limit parameters of each branch and written into the shared table. After reading the energy status and audio branch current limits, the audio link dynamically adjusts the decoding, equalization, noise reduction, echo, and volume control based on the calibration results of the speaker and power amplifier.

5. The multi-interactive AI smart power bank management system according to claim 1, characterized in that: Based on the energy status flag output by the power supply coordination module and the power limit of the artificial intelligence service branch, a control data channel and an audio channel are established through Bluetooth dual-channel communication. First, the request type, estimated load, real-time level, degradability level and security contribution weight fields are uniformly defined for each artificial intelligence service request. Then, using the statistical regression results of the relationship between voice duration, text length and processing time, energy consumption and traffic during the development phase, each request is broken down into load estimates in three dimensions: energy consumption, network consumption and audio consumption.

6. The multi-interactive AI smart power bank management system according to claim 5, characterized in that: The dispatcher reads the current energy status, remaining uptime, and trends from the power supply coordination module. It then combines this information with the network status sequence, network fluctuations, number of emergency requests, and interval of the last emergency broadcast reported by supporting applications, as well as the current audio power usage and the time and duration of the next mandatory voice broadcast from the audio link. The dispatcher then calculates three components: energy stress, network availability risk, and emergency information lag risk. This constructs a comprehensive situation indicator. Under the triple constraints of energy budget, network budget, and audio budget, the dispatcher calculates a comprehensive priority based on the situation indicator and the real-time performance, degradability, security weight, and queuing time of each request. This priority is then used to uniformly sort various AI services, including notification synchronization, translation, Q&A, and text updates.

7. The multi-interactive AI smart power bank management system according to claim 6, characterized in that: In specific scheduling, for high-priority requests, we attempt to match combinations of response length, context depth, translation service level, and output mode. When the multidimensional budget is insufficient, we sequentially shorten the response, reduce the context depth, reduce from speech plus text to text only, and even postpone or discard it according to the preset degradation path. We also combine the prediction of future network availability windows and the local multilingual emergency template library. When network conditions improve, we focus on executing high-level translation and Q&A. When the network is poor for a long time, we use local templates to cover rigid instructions. At the same time, we dynamically tighten speech synthesis insertion according to audio power and mandatory playback time window, and only retain silent text output. Finally, we complete the real-time display and broadcast by encrypting the transmission and processing the text only in volatile storage.

8. The multi-interactive AI smart power bank management system according to claim 1, characterized in that: Under the constraints of the energy state and the upper limit of the power of the motion sensing branch given by the power supply coordination module, the motion sensing power consumption closed-loop module establishes linkage control between motion sensing input, mini-game load and peripheral power supply: by comparing the recognition accuracy, response delay and power consumption under multiple sampling frequencies, the inertial sensor operating frequency is selected. During operation, the inertial sensor operating frequency is used to collect posture data and match it with the preset gesture template, and the waving, shaking and flipping actions are converted into mute, pause and wake-up control commands.

9. A multi-interactive AI smart power bank management system according to claim 8, characterized in that: The system dynamically adjusts the frame rate of mini-games, motion sensing sampling frequency, and the on / off and standby or wake-up strategies of audio, display, and Bluetooth peripherals based on energy status. When energy is insufficient and there is no operation for a long time or high load, it automatically degrades or shuts down high-power services. After being woken up, it re-evaluates the energy status and selectively restores services.

Images