A device, system and method for identifying based on social interaction of young children

Abstract

The application discloses a kind of recognition device, system and method based on social interaction of infant, including badge type shell, for fixed in the chest position of wearer;Voice acquisition module, fixed in the badge type shell, for collecting voice data to form recording file;Low-frequency communication module, fixed in the badge type shell;Main control module is respectively with the voice acquisition module and the low-frequency communication module electric connection and with main control module electric connection data storage module and wireless transmission module, the application solves the problem of strong noise interference and voiceprint recognition failure, near-field ranging accuracy and "wall misjudgment", high-frequency high-precision synchronization leads to battery endurance disaster, recording and positioning event split, 100% near-field event capture is realized, from physical layer ensures the accuracy of "social interaction occurs", ensures that every audio file uploaded to the cloud is highly condensed "language interaction dry goods", and greatly prolongs the battery endurance of hardware device, with very low hardware cost, perfect two-way double-channel clear dialogue restoration is realized.

Description

Technical Field

[0001] This invention relates to the technical field of early childhood education, specifically to a recognition device, system, and method based on early childhood social interaction. Background Technology

[0002] In a preschool education setting, data on children's daily social behaviors, conversations, and teacher-child interactions are crucial for assessing their language development, social preferences, and mental health. However, existing technologies have the following significant limitations when implemented in actual kindergarten environments: (1) Strong environmental noise interference and voiceprint recognition failure: The kindergarten environment is extremely noisy, and the children's speech is discontinuous and incomplete. It is extremely difficult to accurately identify the children's identities and separate their voices by relying solely on traditional software voiceprint recognition algorithms. (2) Poor near-field ranging accuracy and "misjudgment through walls": Traditional Bluetooth RSSI or WiFi positioning has serious multipath effects and large signal drift, making it impossible to accurately determine face-to-face social interactions within 1 meter. Ultra-wideband (UWB) positioning has problems such as excessive power consumption, high cost, and inability to link with recording. (3) Battery life disaster caused by high frequency and high precision synchronization: If distributed wearable devices are to maintain unnecessary millisecond-level (such as within ±10ms) absolute time synchronization in the classroom, the devices must frequently perform wireless communication handshakes with the gateway, which will generate extremely high power consumption, causing the batteries of micro wearable devices to be quickly depleted and unable to meet the needs of all-day battery life. (4) Separation of recording and positioning events: Traditional positioning systems and recording systems operate independently. Due to the lack of effective end-side instantaneous linkage and cloud tolerance compensation mechanisms, the workload of later audio track alignment is catastrophically amplified. Summary of the Invention

[0003] In view of the shortcomings of the prior art, the present invention provides a recognition device, system and method based on early childhood social interaction, which solves the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A device for recognizing early childhood social interactions, comprising: The badge-style casing is designed to be attached to the wearer's chest. The voice acquisition module, fixed inside the badge-shaped housing, is used to acquire voice data to form an audio recording file. It includes a first microphone and a second microphone. When worn, the first microphone faces upward toward the wearer's mouth, and the second microphone faces downward toward the wearer's mouth. The low-frequency communication module, fixed inside the badge-shaped housing, includes a miniature high-permeability ferrite wire-wound antenna for transmitting and receiving near-field low-frequency broadcast signals within a set local spatial sensing range. The transmitting module supports drive level control to achieve adjustment of different sensing ranges. The main control module is electrically connected to the voice acquisition module and the low-frequency communication module respectively. The main control module has a built-in drive level control program and a voice activity detection (VAD) algorithm. The main control module triggers an effective speaking state based on the difference in acoustic physical parameters of the first microphone and the second microphone. The drive level control program adjusts the drive strength of the low-frequency communication module to control the local spatial sensing range within the set default distance. At the same time, it controls the low-frequency communication module to transmit a near-field low-frequency broadcast signal containing the device ID and routinely listens to the low-frequency broadcast signals transmitted by other nearby identification devices to generate low-frequency interactive events. The data storage module is electrically connected to the main control module and is used to locally store the recording files and the low-frequency interaction events. The wireless transmission module is electrically connected to the main control module and is used to upload locally stored data to a remote cloud platform.

[0005] Because preschool children are small in size, this device adopts a lightweight chest badge design. To achieve accurate voice separation within a limited volume, the first and second microphones are arranged vertically within the badge-shaped housing, with the physical center distance between them limited to 2.5cm - 3.5cm.

[0006] Preferably, when detecting a valid speaking state, the main control module calculates the real-time signal energy ratio between the first microphone and the second microphone. ,in This represents the real-time signal energy value of the first microphone. The energy value of the second microphone is the real-time signal energy value; it is determined to be the wearer's own valid speech if and only if the energy ratio R is greater than a preset absolute energy threshold θ. In actual wear, because the badge is located on the chest, the physical distance between the upper first microphone and the child's mouth is significantly smaller than that between the lower second microphone. When the child speaks, the energy of the sound waves reaching the first microphone is significantly higher than that of the second microphone, and there is a slight time difference of arrival (TDOA).

[0007] Preferably, when the main control module determines that there is a valid speaking state, it controls the data storage module to automatically start encoding and recording; when it detects that there is no valid speaking sound for more than 3 consecutive seconds, it controls the data storage module to automatically close the current recording file and save it, and then recreate a new recording file locally and restart the voice detection process.

[0008] Preferably, the low-frequency communication module has a built-in anti-collision mechanism; the length of the low-frequency broadcast signal data packets it transmits is limited to a preset millisecond-level time window, and the anti-collision mechanism is configured such that after a valid speaking state is triggered, the low-frequency communication module transmits the low-frequency broadcast signal by randomly delaying and retransmitting it a certain number of times within a fixed time period, and the total period of the random delay and retransmission is... This is much smaller than the 3-second silent slice time of the audio file.

[0009] Preferably, the low-frequency communication module transmits and receives a near-field low-frequency broadcast signal of 125kHz, and the local spatial sensing range is set to 1 meter by default. Traditional Bluetooth RSSI ranging suffers from signal instability in indoor environments due to electromagnetic multipath effects, often resulting in false alarms such as "it can still be sensed through a thin wall or screen".

[0010] This invention also provides a system for recognizing social interactions in young children, comprising: Multiple identification devices for early childhood social interaction as described in any one of claims 1-7 are worn by children and teachers in the class, and the device ID of each identification device is bound to the corresponding child's student ID or teacher's employee ID in the system; The IoT gateway, as the reference clock source within the local area network, is used to periodically send clock synchronization frames to each of the identification devices over a long period of time; each of the identification devices synchronizes its time with the IoT gateway through its wireless transmission module, keeping the global clock deviation of all identification devices within ±100ms during the observation period. The remote cloud platform, including a processor and memory, communicates with the wireless transmission module via a wireless network to receive audio files and low-frequency interaction events uploaded by each identification device, and performs bidirectional intersection matching of device ID and timestamp on the same global timeline to identify social interaction behaviors.

[0011] This invention also provides a method for recognizing social interactions in young children, comprising the following steps: S1: Low-power local clock alignment: After each identification device is powered on, it performs low-frequency local network clock synchronization with the IoT gateway through the wireless transmission module to maintain the global clock deviation among all devices in the class not exceeding ±100ms. S2: Dual-microphone energy comparison VAD-triggered recording: During device operation, the main control module uses the real-time signal energy ratio R of the vertically placed dual microphones to determine whether the wearer is speaking; when valid speech is detected, encoding recording is started and the recording file with device ID and transmission timestamp is stored locally. This timestamp can be set as the start time of valid speech. S3: Near-field low-frequency anti-collision broadcast: At the moment VAD triggers recording, the device broadcasts a local near-field signal to the outside world through a 125KHz low-frequency communication module and a miniature ferrite antenna. The signal is randomly retransmitted within a fixed time period to avoid air collisions between signals from multiple devices. The signal contains a transmission timestamp (the start time of the effective speech), the device ID, and the low-frequency sensing range. S4: Near-field event bidirectional capture: A neighboring device within 1 meter sensing range of this device captures the signal through its normally monitored low-frequency receiver, and packages the device ID, receiving timestamp, transmitting device ID, transmitting timestamp, and low-frequency sensing range into a low-frequency interactive event and stores it locally; S5: Cloud-based elastic spatiotemporal alignment and multimodal big data analysis: The remote cloud platform receives data uploaded by each device, and according to the bidirectional ID and timestamp in the event information, slices and splices the recording files of two devices that have interactive conflicts or intersections on the global timeline within a tolerance window of ±100ms. Based on the audio waveform or timestamp offset, elastic alignment compensation is performed to restore the multi-channel bidirectional complete dialogue audio track. Based on the aligned audio track and events, multi-dimensional social topology and teacher-child interaction indicators are calculated. S6: Cloud-based NLP semantic-based social metric algorithm calculation. This metric not only helps understand the social situation of young children, but also allows for simultaneous understanding of teacher-child interaction metrics. The rigorous total social time of target i within the observation period T Defined as: in: The total number of low-frequency near-field interaction events between child i's device and peer j's device within a period T is recommended to be set to one day or a period of independent activity. and These are the start and end global timestamps of the m-th low-frequency interaction event, respectively. The indicator function for determining social substantial association (taking a value of 0 or 1) is controlled and constrained by the following conditions: For low-frequency near-field bidirectional spatial sensing constraints. Physical audio track energy interleaving and overlap constraints This is a semantic round probability constraint.

[0012] To balance device power consumption, hardware cost, and data alignment quality, this invention abandons the high-cost, high-power millisecond-level high-frequency synchronization scheme, limiting the overall system clock deviation tolerance to within ±100ms. The IoT gateway, acting as the master clock source within the classroom's local area network, broadcasts time synchronization frames to all badge devices in the classroom at preset minute-level intervals (e.g., every 5 minutes). Upon receiving the synchronization frame, the master control module of each badge device corrects its local RTC clock, significantly extending battery life.

[0013] To avoid air interference with the 125kHz signal when multiple toddlers gather within a 1-meter radius and speak simultaneously, this invention introduces a random retransmission mechanism within a fixed time period. Once a badge's VAD is triggered, the low-frequency module will initiate multiple retransmissions based on a random delay algorithm within the subsequent fixed time window.

[0014] Because young children's language expression is discontinuous, fragmented, and easily interrupted, prolonged continuous recording not only generates a large amount of redundant data containing blank ambient sounds, but also places a huge computational burden on the cloud-based speech-to-text (ASR) process for long speech alignment. This invention employs a dynamic segmentation and streaming mechanism. The main control module monitors the VAD (Voice-Activated Audio Device) status in real time. When a child's speech is interrupted, the main control module starts timing. If no valid sound is detected for more than 3 consecutive seconds, the intention of the dialogue or the end of a segment of speech is determined. The main control module immediately sends an interrupt to the flash memory, encapsulates, closes, and saves the currently recorded file (injecting the current device ID and precise timestamp into the filename). Subsequently, the main control module clears the cache, recreates and mounts a new blank recording file, and the system re-enters standby VAD monitoring state. This "3-second silence equals segmentation" logic ensures that every audio file uploaded to the cloud is a highly condensed "language interaction essence," and significantly extends the battery life of the hardware device.

[0015] Due to a maximum clock hardware asynchronous drift of ±100ms between the badge devices, when the cloud platform receives audio slice files and low-frequency interaction event logs from two different IDs, directly splicing them according to the literal timestamps may result in slight overlap or gaps in the dialogue audio tracks. To eliminate this ±100ms hardware clock inconsistency, the cloud platform introduces a flexible alignment compensation algorithm. First, the cloud calculates the relative clock offset reference value based on the low-frequency broadcast signal transmission timestamp (TX_Time) and reception timestamp (RX_Time) recorded by the two devices. Then, it retrieves the corresponding multi-channel audio slices and, within a ±100ms sliding window, fine-tunes the alignment of the audio data from the two channels based on the principle of maximizing audio envelope energy mutation point matching or cross-correlation. This mechanism allows the hardware to maintain a low-power, low-precision ±100ms local area network clock, while transferring the high-precision alignment burden to the cloud, achieving perfect bidirectional dual-channel clear dialogue restoration at extremely low hardware cost.

[0016] The cloud platform constructs a class social network graph matrix G = (V, E) by analyzing low-frequency interaction events, where V is the set of nodes of the recognition device and E is the set of social edges between nodes.

[0017] Interaction weight calculation formula: In the formula: The total frequency of successful bidirectional low-frequency event handshakes between device i and device j within period T; α represents the duration of the k-th interaction; α and β are preset frequency and duration weighting coefficients.

[0018] Social marginalization screening metrics (degree centrality): When a certain early childhood milestone If a student's score falls below two standard deviations from the class average for M consecutive teaching days, the system will automatically trigger a "social isolation risk warning," prompting the teacher to provide targeted care and guidance.

[0019] The present invention provides a recognition device, system and method based on early childhood social interaction, which solves the problems mentioned in the background art, such as strong environmental noise interference and voiceprint recognition failure, poor near-field ranging accuracy and "misjudgment through walls", battery life disaster caused by high frequency and high precision synchronization, and the disconnect between recording and positioning events. This invention introduces an "energy comparison algorithm" into the main control chip: the main control chip calculates the energy ratio of the two microphones in real time. The system triggers VAD recording and low-frequency transmission only when R is greater than a preset absolute threshold θ. When there is ambient noise or other children speaking nearby (far-field sound sources), the energy and time it takes for the sound waves to reach both microphones are essentially the same (R is close to 1.0), and the system directly performs hardware-level filtering. This organic linkage between physics and algorithms perfectly solves the industry pain points of inaccurate recording and misinterpretation of voiceprints in noisy kindergarten environments. This invention employs 125kHz low-frequency near-field communication, a frequency band characterized by a strong spatial boundary descent effect (extremely rapid edge attenuation). To flexibly adjust the interaction detection distance for different classroom sizes or dynamic scenarios, a drive level control program is written into the main control module. The drive strength (i.e., current drive capability) of the low-frequency transmitting antenna is directly adjusted via software commands, allowing selection of different signal transmission power levels. By fine-tuning the antenna drive strength, the effective sensing radius of the low-frequency signal is locked within a 1-meter physical boundary. As long as the other party is within 1 meter and facing you (with the badge directly opposite), 100% near-field event capture can be achieved, ensuring the accuracy of "social interaction occurs" at the physical layer. To ensure that the system timeline does not conflict, this invention limits the total random retransmission period for low-frequency collision avoidance. This is significantly shorter than the 3-second silent slice waiting time of VAD, creating a substantial difference in magnitude. This ensures that low-frequency interaction events are completely closed within milliseconds, never interfering with the second-level audio slicing logic, thus guaranteeing robustness under high concurrency across multiple devices. This "3-second silence for slicing" logic ensures that every audio file uploaded to the cloud is a highly condensed "language interaction essence," and greatly extends the battery life of hardware devices. The cloud platform introduces an elastic alignment compensation algorithm, which allows the hardware to maintain a low-power, low-precision ±100ms local area network clock, while transferring the pressure of high-precision alignment to the cloud, achieving perfect bidirectional dual-channel clear dialogue restoration at extremely low hardware cost. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1 Two toddlers are at opposite ends of a space and each is talking to themselves (long-distance physical isolation eliminates the temporal flow). While student 01 (Device ID: Tag_St_01) is quietly reading a picture book in the reading area on the west side of the classroom, student 06 (Device ID: Tag_St_06 was in the science area on the east side of the classroom (about 8 meters away) loudly discussing a science experiment with a classmate. Both were doing this at the same time. There is continuous vocalization within each segment. The internal data flow and rigorous judgment logic of the system are as follows: Physical space layer logic determination: Since the two children are far apart in the classroom, exceeding the 1-meter local spatial sensing boundary set by this device, and the 125KHz near-field low-frequency signal has a sudden drop effect on the spatial electromagnetic boundary.

[0022] When student 01 speaks, the VAD of its badge Tag_St_01 is triggered and coded recording is started (generating Rec_St01), while the low-frequency antenna broadcasts the local ID packet; however, the low-frequency receiver of Tag_St_06, located in the eastern science area, is completely unable to capture any low-frequency electromagnetic pulses from Tag_St_01 during normal listening.

[0023] Similarly, when student 06 speaks, triggering the low-frequency broadcast of its badge, student 01's badge is also in a signal sensing blind zone, and there are no timestamp records of mutual capture in the event logs between the two devices.

[0024] Processor determination result: Space is completely isolated, cloud processor instantly determines condition one. Final integral filtering effect: Based on logical AND control constraints, no subsequent conditions need to be evaluated, and the indicator function... The cloud-based big data engine, in processing the first step (scanning the low-frequency event matrix), relies on... The feature directly suspends and blocks the social association between the two IDs, without even requiring significant subsequent computational resources for deep ASR transcription and NLP semantic comparison. Ultimately, the actual social time integral between the two individuals is 0 seconds. This embodiment demonstrates that the present invention possesses excellent adaptive pruning capability and rigorous spatiotemporal filtering when processing mixed audio, video, and location data streams uploaded concurrently from dozens of devices across the class.

[0025] Example 2 Two toddlers engage in focused, face-to-face interaction at close range (positive social interaction assessment and precise time-series integration). During free play time, Student 01 (Device ID: Tag_St_01) and Student 02 (Device ID: Tag_St_02) are sitting face-to-face in the role-playing area (dollhouse). They are approximately 0.5 meters apart and are focused on assigning roles and simulating a cooking scene. The internal data flow and rigorous judgment logic of the system are as follows: Physical space layer bidirectional mutual inductance determination: At 14:00:00.000 (based on the global reference clock), student 01 speaks, and the VAD of its badge Tag_St_01 is instantly triggered, starting the encoding and recording of the audio file Rec_St01_140000.mp3 locally; at the same time, its 125KHz low-frequency antenna instantly radiates a near-field electromagnetic broadcast packet carrying the local ID.

[0026] Because the two children were facing each other in a focused social posture, and the physical distance (0.5 meters) was within the 1-meter low-frequency sensing boundary, the receiver of student badge Tag_St_02 instantly captured the broadcast packet during normal listening and recorded the reception event locally.

[0027] At 14:00:02.000, Student 01 finished speaking, and Student 02 immediately took over the conversation. The VAD of Tag_St_02 was triggered, enabling local recording and inversely radiating a low-frequency broadcast packet; this packet was successfully captured by Student 01's badge Tag_St_01 within the flexible tolerance window of the global clock of ±100ms.

[0028] Processor determination result: Throughout the entire dialogue cycle, the two devices generated a high-frequency, highly symmetrical bidirectional near-field low-frequency handshake response pair. Cloud processor determination condition one. .

[0029] 2. Signal energy layer audio track interleaving determination: After the cloud platform accepts and uses the elastic alignment compensation algorithm to align the time axes of the two audio slices, it calls the acoustic comparison module to compare the sound pressure envelope waveforms of the two channels. and Perform sliding window analysis.

[0030] Analysis shows that: when When in the high-energy sound emission range In the low-energy listening silence zone; when When the envelope energy drops to zero, within ±100ms of the elastic sliding window, The envelope energy instantly generates an upward abrupt peak. The waveform energy across the entire timeline exhibits the perfect interwoven characteristics of interpersonal dialogue: "rising and falling, alternating in turn, with slightly overlapping boundaries."

[0031] Processor determination result: The acoustic waveform has highly substantial interactive characteristics, and is not a source of strong noise; cloud processor determination condition two. .

[0032] Semantic text layer round association determination: The ASR engine converts the aligned dual-channel audio stream into a text stream.

[0033] Student 01's transcribed text reads: "Xiaohong, how about you be the mom and I be the dad today?"; Student 02's transcribed text immediately follows: "Okay, then I'll be in charge of cooking at the dollhouse, and you go out to work and earn money."

[0034] The NLP semantic engine calculates the contextual relevance of the two texts. Semantic matrix analysis shows that the two texts are highly focused on the theme of "role-playing division of labor", and the latter has a clear semantic response to the former. The probability of their dialogue round relevance is much higher than the preset substantive social threshold (e.g., greater than or equal to 0.85).

[0035] Processor determination result: Contextual semantic logic closed loop, cloud processor determination condition three .

[0036] Final integration and representation output results: Based on the ternary AND control constraints, during the 180-second continuous face-to-face communication period from 14:00:00 to 14:03:00, the characteristic function... .

[0037] The cloud-based big data engine classifies this period as a "high-credibility, substantial peer social event," and its points system is used for cumulative accumulation. Second The platform accurately counts 180 seconds into the two people's "Total Social Time (TST)" for the day.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A recognition device for early childhood social interaction, characterized in that, include: The badge-style casing is designed to be attached to the wearer's chest. The voice acquisition module, fixed inside the badge-shaped housing, is used to acquire voice data to form an audio recording file. It includes a first microphone and a second microphone. When worn, the first microphone faces upward toward the wearer's mouth, and the second microphone faces downward toward the wearer's mouth. The low-frequency communication module, fixed inside the badge-shaped housing, includes a miniature high-permeability ferrite wire-wound antenna for transmitting and receiving near-field low-frequency broadcast signals within a set local spatial sensing range. The transmitting module supports drive level control to achieve adjustment of different sensing ranges. The main control module is electrically connected to the voice acquisition module and the low-frequency communication module respectively. The main control module has a built-in drive level control program and a voice activity detection (VAD) algorithm. The main control module triggers an effective speaking state based on the difference in acoustic physical parameters of the first microphone and the second microphone. The drive level control program adjusts the drive strength of the low-frequency communication module to control the local spatial sensing range within the set default distance. At the same time, it controls the low-frequency communication module to transmit a near-field low-frequency broadcast signal containing the device ID and routinely listens to the low-frequency broadcast signals transmitted by other nearby identification devices to generate low-frequency interactive events. The data storage module is electrically connected to the main control module and is used to locally store the recording files and the low-frequency interaction events. The wireless transmission module is electrically connected to the main control module and is used to upload locally stored data to a remote cloud platform.

2. The recognition device for early childhood social interaction according to claim 1, characterized in that: The first microphone and the second microphone are arranged vertically inside the badge-shaped housing, with a physical center-to-center distance of 2.5cm to 3.5cm.

3. The recognition device for early childhood social interaction according to claim 1, characterized in that: When detecting a valid speaking state, the main control module calculates the real-time signal energy ratio between the first microphone and the second microphone. ,in This represents the real-time signal energy value of the first microphone. The real-time signal energy value of the second microphone; the energy ratio R is determined to be the wearer's own valid speech if and only if the energy ratio R is greater than the preset absolute energy threshold θ.

4. The recognition device for early childhood social interaction according to claim 3, characterized in that: When the main control module determines that there is a valid speaking state, it controls the data storage module to automatically start encoding and recording; when it detects that there is no valid speaking sound for more than 3 consecutive seconds, it controls the data storage module to automatically close the current recording file and save it, and then recreate a new recording file locally and restart the voice detection process.

5. The recognition device for early childhood social interaction according to claim 4, characterized in that: The low-frequency communication module has a built-in anti-collision mechanism; the length of the low-frequency broadcast signal data packets it transmits is limited to a preset millisecond-level time window, and the anti-collision mechanism is configured as follows: after a valid speaking state is triggered, the low-frequency communication module transmits the low-frequency broadcast signal by randomly delaying and retransmitting it a certain number of times within a fixed time period, and the total period of the random delay and retransmission is... This is much smaller than the 3-second silent slice time of the audio file.

6. The recognition device for early childhood social interaction according to claim 1, characterized in that: The low-frequency communication module transmits and receives near-field low-frequency broadcast signals of 125KHz, and the local spatial sensing range is set to 1 meter by default.

7. A recognition system for early childhood social interactions, characterized in that, include: Multiple identification devices for early childhood social interaction as described in any one of claims 1-7 are worn by children and teachers in the class, and the device ID of each identification device is bound to the corresponding child's student ID or teacher's employee ID in the system; The IoT gateway, as the reference clock source within the local area network, is used to periodically send clock synchronization frames to each of the identification devices over a long period of time; each of the identification devices synchronizes its time with the IoT gateway through its wireless transmission module, keeping the global clock deviation of all identification devices within ±100ms during the observation period. The remote cloud platform, including a processor and memory, communicates with the wireless transmission module via a wireless network to receive audio files and low-frequency interaction events uploaded by each identification device, and performs bidirectional intersection matching of device ID and timestamp on the same global timeline to identify social interaction behaviors.

8. A method for recognizing preschool social interactions based on the system described in claim 7, characterized in that, Includes the following steps: S1: Low-power local clock alignment: After each identification device is powered on, it performs low-frequency local network clock synchronization with the IoT gateway through the wireless transmission module to maintain the global clock deviation among all devices in the class not exceeding ±100ms. S2: Dual-microphone energy comparison VAD-triggered recording: During device operation, the main control module uses the real-time signal energy ratio R of the vertically placed dual microphones to determine whether the wearer is speaking; when valid speech is detected, encoding recording is started and the recording file with device ID and transmission timestamp is stored locally, which is set as the start time of valid speech; S3: Near-field low-frequency anti-collision broadcast: At the moment the VAD triggers recording, the device broadcasts a local near-field signal to the outside through a 125KHz low-frequency communication module and a miniature ferrite antenna. The signal is randomly retransmitted within a fixed time period to avoid air collisions between signals from multiple devices. The signal contains a transmission timestamp, the device ID, and the low-frequency sensing range. S4: Near-field event bidirectional capture: A neighboring device within 1 meter sensing range of this device captures the signal through its normally monitored low-frequency receiver, and packages the device ID, receiving timestamp, transmitting device ID, transmitting timestamp, and low-frequency sensing range into a low-frequency interactive event and stores it locally; S5: Cloud-based elastic spatiotemporal alignment and multimodal big data analysis: The remote cloud platform receives data uploaded by each device, and according to the bidirectional ID and timestamp in the event information, slices and splices the recording files of two devices that have interactive conflicts or intersections on the global timeline within a tolerance window of ±100ms. Based on the audio waveform or timestamp offset, elastic alignment compensation is performed to restore the multi-channel bidirectional complete dialogue audio track. Based on the aligned audio track and events, multi-dimensional social topology and teacher-child interaction indicators are calculated. S6: Cloud-based social metrics algorithm calculation based on NLP semantics The rigorous total social time of target i within the observation period T Defined as: in: The total number of low-frequency near-field interaction events between child i's device and peer j's device within a period T. and These are the start and end global timestamps of the m-th low-frequency interaction event, respectively. The indicator function for determining social substantial association is controlled and constrained by the following conditions: For low-frequency near-field bidirectional spatial sensing constraints. Physical audio track energy interleaving and overlap constraints This is a semantic round probability constraint.

Images