AI practical teaching system based on machine vision

Abstract

The invention relates to the technical field of artificial intelligence, and discloses an AI practical training teaching system based on machine vision, which comprises a teaching interface box used for collecting a real-time current signal of an electronic practical training tool; and the current fingerprint identification engine is used for determining the accurate working state of the tool by performing dual feature matching on the macroscopic waveform of the current signal and the common-mode noise spectrum, and the cognitive logic engine is used for judging the compliance of the operation based on the state sequence. According to the method, the transient current is converted into direct interpretation of the internal electrical characteristics of the tool, the deterministic recognition capability of the working state of the tool without environmental interference is obtained, and the physical safety of the operation process and the dynamic adaptability of the teaching content are considered through monitoring of the transient current and self-adaptive learning of the operation time sequence.

Description

Technical Field

[0001] This invention relates to an AI training and teaching system based on machine vision, belonging to the field of artificial intelligence technology. Background Technology

[0002] Currently, in AI training and teaching systems, a commonly adopted technical strategy is the external observation method based on machine vision. This involves acquiring videos of students' operations through image acquisition devices, and then using computer vision algorithms to analyze and identify the students' body movements, postures, and interactions with tools, thereby judging the standardization of their operations and the correctness of their procedures.

[0003] However, this method, which relies on external optical imaging, faces insurmountable bottlenecks due to its inherent indirectness when dealing with practical training scenarios such as electronic circuit assembly and debugging, which require precise interaction of tool states. For example, the system may be able to recognize that the trainee is holding a multimeter and touching the circuit board, but it cannot definitively know whether the multimeter is at the correct resistance setting or the wrong voltage setting. Similarly, the system cannot determine whether the soldering iron has reached the required working temperature for soldering based solely on external images before the trainee operates. The critical working states of these tools are precisely the prerequisites for assessing skill mastery and ensuring operational safety. At a deeper level, the fundamental limitation of this technical approach lies in its focus on the trainee's complex physical movements rather than the core interaction between the trainee and the toolchain. In fact, this not only leads to distortion of the source of information but also results in high decoding costs and deployment barriers.

[0004] Specifically, existing technologies suffer from the following shortcomings: 1. All inferences based on external vision cannot escape the ambiguity caused by physical occlusion, changes in lighting, and confusion arising from similar actions, making the judgment of key operational nodes always risky; 2. To improve recognition accuracy, multiple cameras and complex AI models are usually required, which not only increases hardware costs but also places excessive demands on the technical capabilities of the application unit for subsequent data annotation and model training and maintenance; 3. The system cannot penetrate the tool's outer shell to directly obtain the true state information representing its internal working principle, which is a physical limitation of this technology. Therefore, how to avoid the inherent defects of external observation methods and obtain the accurate working state of students interacting with electronic training tools through a reliable and low-cost means, and build an efficient teaching logic judgment system based on this, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides an AI training and teaching system based on machine vision. Its main purpose is to solve the problems of existing technologies that rely on external visual observation, which cannot directly and definitively identify the internal working status of training tools, resulting in uncertainty, high cost, and environmental interference in teaching evaluation.

[0006] To achieve the above objectives, the present invention provides an AI training and teaching system based on machine vision, the system comprising:

[0007] The teaching interface box has at least one power output interface for powering at least one electronic training tool, and the teaching interface box has a built-in current sensor for collecting the current signal flowing through the power line when the electronic training tool is working.

[0008] The current fingerprint recognition engine is connected to a current sensor and stores a fingerprint template library. The fingerprint template library stores macroscopic current waveform features and common-mode noise spectrum features corresponding to various preset tool states. The current fingerprint recognition engine is configured to: extract macroscopic current waveform features and common-mode noise spectrum features in real time based on the collected current signal; and output an event signal indicating that the current working state of the electronic training tool is the preset tool state when the extracted macroscopic current waveform features and common-mode noise spectrum features match the corresponding features of a preset tool state stored in the fingerprint template library.

[0009] The cognitive logic engine receives event signals and is configured to determine whether the sequence of received event signals conforms to the teaching operation procedures, which contain multiple operation nodes.

[0010] Preferably, the current fingerprint recognition engine is deployed on the edge computing unit within the teaching interface box; and the macroscopic current waveform characteristics include the current mean and current variance of the current signal within a time window.

[0011] Preferably, the teaching interface box also has a built-in relay. The teaching interface box is configured to: calculate the slope of the change in the current signal in real time through the edge computing unit, and actively cut off the power supply to the electronic training tool through the relay when the slope of the change exceeds the safety threshold.

[0012] Preferably, the cognitive logic engine is also configured to: upon receiving a new training task instruction, record the actual time consumed by each operation node in at least one successful operation process to generate a baseline time sequence vector, and dynamically update the node timeout threshold used in subsequent judgments based on the baseline time sequence vector.

[0013] Preferably, the system further includes: an image acquisition device for acquiring facial images of the trainee; an eye movement feature analysis module, configured to be activated only when the cognitive logic engine determines that the sequence of event signals does not conform to the teaching operation procedure, to perform the following operations: analyze the facial image and calculate the spatial variance of the trainee's pupil center coordinates within the timing period to generate a gaze stability index; and the cognitive logic engine is also configured to determine whether to delay or suppress intervention instructions to the trainee based on the gaze stability index.

[0014] Preferably, the cognitive logic engine dynamically updates the node timeout threshold based on the benchmark time sequence vector, specifically including: determining the benchmark timeout for each operation node based on the benchmark time sequence vector. The updated node timeout threshold is calculated using the following formula. : in, The updated node timeout threshold. Based on the baseline time, It is a preset margin coefficient that is greater than zero.

[0015] Preferably, the current fingerprint recognition engine is also configured to: provide an initialization interface to guide the user to switch the electronic training tools to each preset tool state in sequence, and automatically record the corresponding macroscopic current waveform characteristics and common-mode noise spectrum characteristics in the process to generate a fingerprint template library.

[0016] Preferably, the common-mode noise spectrum characteristics are the position and amplitude of at least one characteristic frequency peak of the current signal in the frequency range of 1 kHz to 50 kHz.

[0017] Preferably, the edge computing unit obtains the slope of change by sampling the current signal twice consecutively at a sampling interval of less than or equal to 50 microseconds and calculating the difference quotient.

[0018] Preferably, the cognitive logic engine is also configured to generate a safety alarm command when it is detected that an electronic training tool is continuously in a non-power-off standby state for a duration exceeding an alarm duration threshold.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. By directly acquiring the current signal flowing through the power line when the electronic training tool is working through the built-in current sensor in the teaching interface box, the current fingerprint recognition engine performs dual comparison and authentication of the macroscopic waveform characteristics and common-mode noise spectrum characteristics of the signal. This method transforms the indirect inference of the student's external behavior into the direct reading of the tool's internal physical state. Since the current fingerprint is uniquely determined by the tool's circuit principle in a specific working mode, the recognition result not only eliminates the interference of all external environmental factors such as light and obstruction, but also solves the problem of state confusion that may occur due to the similarity of macroscopic current in different operating modes through the synergistic verification of macroscopic stability characteristics and microscopic noise characteristics. This allows the system's ability to identify the tool's state to penetrate to the level of its internal working mechanism. The certainty and reliability of its recognition are unattainable by existing external observation methods.

[0021] 2. This invention combines real-time monitoring of the slope of current signal changes with a relay disconnection device to construct a physical layer safety protection mechanism independent of upper-level cognitive logic judgment. When a student causes a reverse connection or short circuit due to misoperation, the system does not rely on recognizing the power-on state, but rather captures the transient characteristics of the surge current whose slope of change is far beyond the normal range. It then actively disconnects the power supply within milliseconds with the highest priority of hardware interruption. This method ensures that safety intervention occurs before the destructive consequences are fully formed, avoiding physical damage to the training equipment and circuit boards. It transforms the system's role from an evaluator of the operation process to one that can actively disconnect the power supply to avoid asset loss.

[0022] 3. By recording the node time of the initial successful operation under a new task through the cognitive logic engine, and using this as a benchmark to dynamically update the timeout threshold of subsequent judgments, a self-calibrating closed loop of teaching rules is established. This method utilizes the process data generated by the system's own operation to optimize its future judgment criteria, enabling the system's core logic parameters to automatically adapt to different difficulty and type of training tasks. This avoids the tedious process of manually reconfiguring parameters due to changes in teaching content, and also avoids a large number of misjudgments caused by the mismatch between static thresholds and actual operation time, making the system adaptable and easy to use. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system functional architecture of the present invention;

[0024] Figure 2 This is a comparison diagram of current signals under easily confused operating conditions of the present invention;

[0025] Figure 3 This is a flowchart illustrating the process of creating system calibration and teaching content for this invention.

[0026] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0028] This invention provides an AI training system based on machine vision. In a typical training scenario of electronic circuit assembly and debugging, it mainly includes a teaching interface box for signal acquisition and physical layer interaction, a current fingerprint recognition engine for tool status identification, and a cognitive logic engine for judging the compliance of student operation logic. The teaching interface box serves as the physical connection hub between the electronic training tools used by the students and the power supply system. The real-time current signals acquired by the teaching interface box are transmitted to the current fingerprint recognition engine. The engine outputs structured tool status event signals through a dual feature matching mechanism. Finally, the cognitive logic engine receives the event signal sequence and performs a final logical evaluation of the student's operation chain based on the preset teaching procedures.

[0029] In a specific deployment scenario, such as when trainees are performing circuit board soldering and testing tasks, existing technologies based on external cameras struggle to definitively determine whether a multimeter is at the correct resistance setting or whether a soldering iron has reached its operating temperature. To address this challenge, the teaching interface box of this invention is configured with at least one power output interface to provide power to electronic training tools such as soldering irons and multimeters on the trainees' desktops in a plug-and-play manner. This teaching interface box physically deviates from the existing teaching environment, integrating a non-contact Hall current sensor to collect the current signal flowing through each power line. Thus, the system obtains a one-dimensional current time series signal that directly reflects the internal electrical activity of the tools and is not easily affected by external light or physical obstructions, serving as the raw data source for subsequent analysis. Furthermore, to address the fingerprint aliasing bottleneck caused by similar macroscopic current values ​​in different tool states—for example, the average current consumption of a multimeter in diode continuity testing mode may be the same as in low-current standby mode—the system's current fingerprint recognition... The engine is configured to execute a dual-mode authentication logic. On one hand, it extracts the macroscopic current waveform features of the acquired current signal within a preset time window. Specifically, this could be the average current reflecting the overall power consumption level and the current variance reflecting its volatility. On the other hand, it performs a Fast Fourier Transform (FFT) analysis on the same current signal to obtain common-mode noise spectrum features that characterize the internal chip clock or switching power supply operating characteristics. Specifically, these spectrum features can be defined as the position and amplitude of several key characteristic frequency peaks appearing in the frequency range of 1 kHz to 50 kHz. Correspondingly, the engine's built-in fingerprint template library also pre-stores the macroscopic current waveform features and common-mode noise spectrum features corresponding to each preset tool state. Only when both sets of features extracted in real time match the template of a preset tool state in the template library will the engine output a definite event signal with a timestamp. This mechanism achieves the analysis and identification of the tool's operating state through the collaborative verification of the tool's steady-state power consumption and internal high-frequency operating mechanism.

[0030] The determination of this dual-feature matching is achieved by calculating a comprehensive matching confidence level. To achieve this, specifically, firstly, the macroscopic current mean, variance, and frequency and amplitude of multiple common-mode noise spectral characteristic peaks are extracted from the real-time acquired current signal. Each feature value is normalized using Z-score based on the mean and standard deviation of the features corresponding to the state in the fingerprint template library. Then, this confidence score... Determined by a normalized inverse weighted Euclidean distance function: ,in, For real-time normalized feature values, The corresponding feature center value in the template library, and This is a feature discrimination weight, which is obtained during the initialization phase by performing an ANOVA test on a training sample set of various easily confused working states, and normalizing the F-statistic values ​​of each feature term. It is used to quantify the contribution of this feature in distinguishing different states. Ultimately, the weight is determined if and only if the calculated... The engine only outputs a deterministic tool status event signal when the value exceeds a preset matching acceptance threshold, such as 0.92. To ensure that the fingerprint template library can be easily built and updated, the current fingerprint recognition engine is also configured to provide a standardized initialization interface. This initialization process is designed as a deterministic operating procedure: First, the interface guides the user to connect a specific electronic training tool to the teaching interface box; then, the interface prompts the user to manually switch the tool to all its preset working states, such as voltage range, resistance range, on / off buzzer range, etc.; in each state, after user confirmation, the system automatically collects a current signal during stable operation and automatically extracts its macroscopic current waveform characteristics and common-mode noise spectrum characteristics. Then, these characteristics are bound to the corresponding tool status labels, ultimately generating or updating the fingerprint template library. This process does not require any complex data annotation by the user, ensuring the ease of use and data consistency of the template library construction.

[0031] After obtaining a defined sequence of tool state events, the cognitive logic engine performs logic gating judgments based on a teaching operation procedure containing multiple operation nodes to assess whether the trainee's operation conforms to the specifications. For example, a standard soldering process may require node A (measuring components using the resistance setting of a multimeter) to occur before node B (soldering with a soldering iron). After receiving the event signal indicating that node A is complete, the cognitive logic engine opens a time window to wait for the corresponding event signal at node B to appear, thereby verifying the operation timing logic. In addition, the cognitive logic engine is also configured to detect when an electronic training tool is continuously in a non-power-off standby state. Furthermore, when the duration exceeds the preset alarm duration threshold, a corresponding safety alarm command is generated. However, when faced with training tasks of varying difficulty, the fixed node timeout threshold will generate a large number of false alarms or missed alarms due to its inability to adapt to changes in task complexity. To address this adaptability bottleneck, the cognitive logic engine is further configured to execute a self-calibrating closed-loop mechanism that follows a set of teaching rules. Upon receiving a new training task command, the engine records the actual time consumed by each operation node in the first or several successfully completed operations of the task, and generates a baseline time consumption for each node that reflects the characteristics of the new task by, for example, taking the median or truncated mean. Subsequently, the engine will dynamically update the node timeout thresholds used in subsequent judgments based on a deterministic mathematical procedure. The procedure is specifically given by the following formula: ,in, A margin coefficient greater than zero, such as 0.25, is used to provide a reasonable tolerance; thus, the system's judgment criteria can automatically and dynamically adapt to the inherent time rhythm of different teaching content; the margin coefficient in this formula... It is not a fixed value, but rather determined by a baseline margin. Determined together with a dynamic adjustment term, its complete form is: Among them, the benchmark margin For example, 0.15 is the global minimum fault tolerance time determined by the system during the offline calibration phase after statistical analysis of operational data covering various types and difficulty benchmark teaching tasks; This is the coefficient of variation of historical operation times for a specific operation node. This coefficient is defined as the ratio of the standard deviation of all recorded successful operation times for that node to its arithmetic mean, used to dimensionlessly characterize the dispersion of operation times; and This is a sensitivity constant, such as 0.5, used to adjust the response strength of the margin coefficient to the dispersion of operation time. Thus, a node with uniform internal operation steps and small differences in time consumption among trainees will obtain a smaller margin. This allows for more compact timing control; conversely, a node with complex operations and that allows for different exploration paths will automatically receive a larger value. The value is adapted to its inherent uncertainty.

[0032] In a more specific teaching scenario, a student's pause in operation does not always indicate a lack of ability; it could also be due to deep thinking and problem-solving. To enable the system to distinguish between these two states, this invention also includes an image acquisition device and an eye-tracking feature analysis module. This module is designed to follow the principle of on-demand service, meaning it is only activated when the cognitive logic engine determines that the student's operation has encountered a non-compliant situation such as a node timeout. Once activated, the module acquires the student's facial image through the image acquisition device and extracts the pupil center coordinates using a lightweight facial landmark detection algorithm. Then, it quantifies the student's cognitive load through a deterministic calculation process: specifically, the module calculates the spatial variance of the pupil center coordinates within a short timing period to generate a gaze stability index. A lower variance value indicates that the student is in a state of deep thinking with a high gaze. Upon receiving this judgment, the cognitive logic engine decides to delay or suppress intervention commands to the student, thus achieving a non-intrusive teaching interaction. Finally, to address potential equipment damage caused by student misoperation... To mitigate risks such as reversed power supply polarity or accidental short circuits, this invention integrates a physical layer safety protection mechanism independent of upper-layer logic within the teaching interface box. The core of this mechanism lies in real-time monitoring of the slope of the current signal change. This is achieved by using an edge computing unit to sample the current signal twice consecutively at intervals no greater than 50 microseconds and calculating the difference to approximate the slope. Given that destructive operations such as short circuits can trigger a transient inrush current with a slope far exceeding the normal power-on value, the system presets a safety threshold. This threshold is determined by empirically measuring the maximum current slope of all supporting tools during normal power-on and multiplying it by a safety factor, such as 2.0. Once the slope exceeds this safety threshold, a hardware interrupt program with the highest processing priority is immediately triggered. This interrupt program bypasses all upper-layer application logic and directly drives the relays built into the teaching interface box, proactively cutting off power to the corresponding electronic training tools within milliseconds. This proactive avoidance measure, based on physical layer transient characteristics and prior to logical judgment, effectively removes the risk source before physical damage occurs.

[0033] Example 1: In a training scenario for troubleshooting and repairing high-value, high-density circuit boards, the technical solution of the present invention operates as follows: This scenario requires trainees to perform a set of operating procedures, including electrical measurement and power-on debugging, on a circuit board equipped with multiple sensitive microprocessors. In the initial stage of this scenario, trainees are required to first use the diode continuity setting of a multimeter to perform a short-circuit test to ground on a critical power interface on the circuit board. When the trainee switches the multimeter to this setting, a specific technical condition immediately appears: the macroscopic average current at this setting is different from the current at another unrelated setting. Standby current is almost indistinguishable numerically. If only macroscopic current waveform characteristics are relied upon for judgment, ambiguity in state identification will arise. At this point, the parallel authentication mechanism of the current fingerprint recognition engine in this solution is executed. While performing macroscopic feature comparison, it analyzes the common-mode noise spectrum characteristics of the current signal. Since the buzzer drive circuit built into the diode on / off position will generate a characteristic frequency peak in a specific frequency band, while the other state does not have this characteristic, the engine determines the current working state of the tool as diode on / off by double matching verification of macroscopic features and spectrum characteristics, and outputs the corresponding event signal to the cognitive logic engine.

[0034] After confirming that the student has completed the short-circuit test, the teaching procedure moves to the next stage, which requires the student to apply a rated operating voltage to the circuit board. After connecting the adjustable power supply, the student pauses due to a brief lapse in memory of the operating procedure, staring at the circuit board and the instruction manual. At this point, the cognitive logic engine determines that a timeout has occurred because it has not received a power-on event signal for an extended period. This timeout event then triggers the eye-tracking feature analysis module. This module acquires the student's facial image through an image acquisition device and calculates that the spatial variance of the pupil center coordinates within the timing period is extremely small, thus generating a high gaze stability index. Upon receiving this index, the cognitive logic engine classifies the nature of this timeout event as deep thinking and decides to suppress intervention commands to the student, providing a undisturbed environment for the student to independently troubleshoot the problem.

[0035] After careful consideration and setting the voltage, the student accidentally reversed the positive and negative terminals of the power supply when finally powering it on. The instant the student pressed the power switch, a destructive reverse bias surge current began to form in the circuit. The edge computing unit deployed in the teaching interface box, with a high-priority independent thread, was continuously calculating the slope of the current signal at intervals of no more than 50 microseconds. Within the first few milliseconds of the surge current forming, its slope had already far exceeded the safety threshold calibrated based on normal power-on conditions. This over-limit event immediately triggered a hardware interrupt. The interrupt service routine bypassed all cognitive logic judgments and directly drove the relay to cut off the power supply.

[0036] Ultimately, before the sensitive chip on the circuit board suffered physical damage due to reverse breakdown, the power supply was cut off. The system then issued a safety alarm to the student regarding the detection of an abnormal current surge and prompted them to check the correctness of the power connection. The student's misoperation was stopped in time, and the training assets were protected. In this single operation process, the entire teaching system addressed the ambiguity of state recognition, the adaptability of teaching interaction, and the safety issues of physical operation through the parallel operation of its built-in state depth recognition, cognitive load perception, and transient safety protection mechanisms.

[0037] Example 2: To objectively quantify the recognition performance of the current fingerprint recognition engine in this invention under conditions with similar macroscopic electrical characteristics but vastly different functional states, the following experiment was conducted. The purpose of this experiment was to quantitatively compare and verify the improvement in the system's accuracy in recognizing the tool's working state after introducing common-mode noise spectrum characteristics as a second verification dimension. The experimental platform was built in a standardized electronics laboratory, the core of which was a system prototype containing a teaching interface box and a built-in edge computing unit. A Fluke 15B+ digital multimeter was used as the electronic training tool for this experiment because its macroscopic average operating current is highly similar in value between the diode on / off state and the 20mA current range standby state, constituting the basis for this experiment. Typical fingerprint aliasing conditions to be verified; to obtain baseline data, a Keysight PA2201A power analyzer was connected in series in the power supply circuit to synchronously record the actual macroscopic current value; during the experiment, the sampling frequency of the current signal used for spectrum analysis in the edge computing unit was set to 100kHz. This parameter setting aims to balance the data acquisition fidelity and the computational load of the edge computing unit. The basis for this value is that, in order to capture the common-mode noise spectrum characteristics in the range of 1kHz to 50kHz without distortion, according to the Nyquist sampling theorem, the sampling frequency must be at least twice the highest target frequency. The setting of 100kHz leaves a margin for signal processing under this premise, and is also within the hardware capability range of the selected ESP32 microcontroller.

[0038] The experimental procedure first executes the fingerprint template library generation procedure guided by the initialization interface, recording all working states of the Fluke 15B+ multimeter, especially the macroscopic current waveform characteristics and common-mode noise spectrum characteristics in the diode on / off state and the 20mA current range standby state, into the fingerprint template library. Subsequently, the experiment enters the testing phase, alternating the multimeter to the above two states 100 times each, forming a total of 200 independent test samples. In each test, the system prototype records the recognition results under two algorithm logics: one is single-mode recognition that relies solely on macroscopic current waveform characteristics for comparison, and the other is dual-mode recognition that compares macroscopic current waveform characteristics and common-mode noise spectrum characteristics simultaneously. Some representative data are shown in Table 1.

[0039] Table 1: Performance comparison of single-mode recognition and dual-mode recognition under state confusion conditions.

[0040]

[0041] The experimental data shows that the mean and variance of the macroscopic current in the two target states are extremely close in value, and their differences are completely within the range of normal measurement errors and individual device differences. This results in a recognition success rate of less than 5% under the single-mode recognition logic, with the system returning an uncertain state in most cases. In contrast, under the dual-mode recognition logic, when the system further detects the diode on / off state, a 2.1kHz characteristic frequency peak generated by the internal buzzer drive circuit is stably present in its current signal, while the 20mA current range state does not have this spectral characteristic. Based on this deterministic difference in another dimension, the dual-mode recognition logic gave a judgment consistent with the actual tool state in all 200 tests.

[0042] Example 3: This example combines Figures 1 to 3 A description of an AI training and teaching system based on machine vision, such as... Figure 1 As shown in the diagram, this figure clearly illustrates the internal data flow and control logic of the system. Electronic training tools, such as soldering irons or multimeters, are connected to the system through a teaching interface box. This interface box, while supplying power, simultaneously acquires a one-dimensional current time-series signal during operation. This signal is sent to the physical layer safety protection mechanism for real-time monitoring of the current slope. Once an inrush current caused by a short circuit or other conditions is detected, this mechanism triggers a relay via a hardware interrupt, achieving the protection function of actively cutting off the power supply. On the other hand, this current time-series signal is also sent to a current fingerprint recognition engine. This engine analyzes the macroscopic current waveform characteristics of the signal, such as mean and variance, and common-mode noise. The sound spectrum characteristics, such as characteristic frequency peaks, are used for dual feature matching to determine the precise working status of the tool. The structured tool status event signal is then output to the cognitive logic engine. The cognitive logic engine judges the compliance of the event sequence according to the preset teaching procedures. The judgment result is finally generated into teaching intervention or evaluation instructions and output to the student terminal. In addition, when the cognitive logic engine detects an operation timeout, it will trigger the eye movement feature analysis module. This module judges whether the student is in deep thinking by calculating the gaze stability index. If the judgment is yes, it returns an instruction to the cognitive logic engine to suppress or delay the intervention, thus forming a complete, closed-loop intelligent teaching and safety protection system.

[0043] like Figure 2As shown in the figure, the graph uses time as the horizontal axis (milliseconds) and current as the vertical axis (milliamperes). It plots the real-time current waveforms of a digital multimeter in two states: the solid line in the diode continuity setting and the dashed line in the 20mA standby current range. It can be seen from the figure that the two curves highly overlap in both value and shape, making it difficult to effectively distinguish them based solely on macroscopic waveform characteristics. This illustrates the necessity and technical value of introducing the common-mode noise spectrum as a second identification dimension in this invention.

[0044] like Figure 3 As shown, this process ensures that the calibration of core parameters and the creation of teaching content within the system can be completed in a standardized and adaptive manner. The first stage is offline calibration and initialization, which includes three parallel initialization tasks: calibrating the physical layer safety threshold by collecting the maximum current slope when various tools are started normally and multiplying it by a safety factor; configuring alarm duration thresholds based on the standby characteristics of different tool types, such as soldering irons and multimeters; and guiding users to perform visual focus and non-focus tasks to collect pupil spatial variance to calibrate the decision boundary for deep thinking. After completing the offline calibration, the system enters the second stage, namely online creation and adaptation. In this stage, the system first guides teachers to connect multiple tool samples of the same model to create an aggregated fingerprint template that covers individual differences. Then, the teacher demonstrates the operation according to the new procedure, and the system automatically records the event signal sequence to generate the teaching operation procedure. After that, the system enters the timing benchmark learning stage in the initial stage of online operation, recording the actual time consumed by the first batch of students who successfully completed the operation at each node, and finally dynamically updating the node timeout threshold used for subsequent judgments based on the learned actual time consumption, thereby realizing the self-calibration of the teaching rules.

[0045] Example 4: To ensure the objectivity and reproducibility of the judgment criteria of each core algorithm module in the teaching system of this invention when deployed in teaching scenarios with different combinations of electronic training tools and environmental conditions, a standardized offline calibration and parameter initialization procedure must be executed before the system is officially put into use. The starting point of this procedure is the structured definition and filling of the fingerprint template library. Each record in the library corresponds to a unique tool status and includes the following data fields: tool model identifier, tool status name, macroscopic current mean, macroscopic current variance, and a data structure for storing common-mode noise spectrum characteristics. Specifically, this structure contains several characteristic frequency peaks. The list of position and amplitude pairs; during the filling of the library, the physical layer security threshold is calibrated simultaneously. Specifically, for each electronic training tool included in the teaching scenario, no less than 20 independent power-on start-up operations are performed on it through the teaching interface box. During this period, the edge computing unit continuously calculates the slope of the current signal change at a sampling interval of no more than 50 microseconds and records the slope peak value in each start-up process. Finally, the security threshold for actively cutting off the power supply is set to 200% of the value corresponding to the 95th percentile of all recorded slope peak samples, thereby reserving an anomaly judgment margin above the transient current fluctuation range of normal start-up.

[0046] Furthermore, the alarm duration thresholds involved in the cognitive logic engine are set in relation to the typical usage patterns of the tools. The configuration procedure is as follows: the alarm duration threshold for tools like soldering irons, which should not be left on for extended periods, is set to 15 minutes; the alarm duration threshold for tools like multimeters, which have low power consumption and are frequently in continuous operation, is set to 60 minutes. These parameters are loaded into the system as part of the device configuration file. Simultaneously, to calibrate the judgment threshold for the gaze stability index in the eye-tracking feature analysis module, the system provides a two-stage calibration process. In the first stage, the operator is required to perform a task requiring high visual focus: continuously tracking a slowly moving, small target on the screen for 20 seconds. During this period, the eye-tracking feature analysis module records the spatial variance sequence of the pupil center coordinates and calculates its average as the baseline variance under focused conditions. In the second phase, operators were instructed to perform a visually non-focused task within the same timeframe, namely, naturally scanning the entire room. The system then calculated the baseline variance in the non-focused state in the same manner. Finally, the final decision threshold used to distinguish between the two states. The calculation is performed using the following formula: ,in, The sensitivity coefficient is a value between 0 and 1, and in this embodiment it is 0.3. This procedure collects the baseline of eye movement characteristics of individuals in two states and establishes an interpolated discrimination boundary based on this. By executing the above integrated calibration and initialization procedure, the key judgment parameters in the system are transformed from preset values ​​to deterministic values ​​based on specific application scenarios and measured data. This lays the foundation for the stability, reliability and accuracy of the subsequent operation of the system before it is delivered for teaching use.

[0047] Example 5: In a scenario where a teaching unit introduces a batch of new electronic training tools and needs to create a completely new set of teaching operation procedures, the system provides a standardized teaching content creation procedure. The operator switches the system to creation mode. In this mode, the teacher will be guided to connect multiple tool samples of the same batch under the new model to the teaching interface box one by one. The system automatically executes the initialization process, collects and records the current fingerprint of each sample in each preset tool state, and calculates an aggregated fingerprint template containing the mean and standard deviation range based on the feature data of all samples. This aggregated template is stored in the fingerprint template library to cover the individual electrical characteristic differences caused by manufacturing tolerances.

[0048] After completing the input of the aggregated fingerprint templates for all new tools, the teacher operates a demonstration circuit board in sequence according to the logical order of the new teaching operation procedure. During this process, the system automatically records the event signal sequence output by the current fingerprint recognition engine and matched by the aggregated template, and solidifies this sequence into a teaching operation procedure containing multiple operation nodes. At the same time, the initial reference timing vector bound to this procedure is automatically generated, and the reference time of all its nodes is assigned a unified default value. After the teaching content package is finally saved and sent to the student's terminal, its built-in reference timing vector is automatically updated by the system's adaptive learning mechanism with actual time data when a student successfully completes the operation process for the first time, realizing the dynamic adaptation of teaching logic and timeout judgment criteria.

[0049] Example 6: In the initial online operation of a newly deployed teaching operation procedure in the teaching system, the cognitive logic engine first enters a timing benchmark learning phase consisting of three successful operation processes. During this phase, the system's timeout judgment function for operation nodes is temporarily suppressed. Its function is limited to recording the actual time consumption of each operation node in these three processes, and using this as a sample to calculate the initial benchmark timing vector, providing an initial basis for subsequent routine teaching evaluation. To improve the robustness of the operation procedure judgment and to cope with the brief tool state recognition jumps caused by student operation jitter or instantaneous power grid noise, the cognitive logic engine incorporates a state confirmation filtering logic when receiving event signals from the current fingerprint recognition engine. A tool state transition is only recognized as a valid operation node switch by the cognitive logic engine if it is continuously reported by the current fingerprint recognition engine N times, where N is an integer greater than 1, which is set to 3 in this embodiment, or if it is stably maintained for a duration T, which is set to 100 milliseconds in this embodiment. Isolated event signals that do not meet this timing stability condition are judged as interference and discarded.

[0050] Furthermore, to ensure that the timeout thresholds for nodes in the teaching operation procedures can adapt to the inherent difficulty of different tasks, a margin coefficient is used. The value is configured to be dynamically adjusted based on statistical analysis of historical operation data; after accumulating a sufficient number of operation completion samples, the system will calculate the time spent by all trainees at a specific operation node and compare it with their baseline time. If the standard deviation is low, it indicates that the operation difficulty of that node is uniform and clear, and the system will automatically lower it. The value is used to tighten the timeout judgment; conversely, if the standard deviation is high, it indicates that the node operation is complex or there are multiple reasonable operation paths, and the system will adjust the value accordingly. The value of the margin coefficient is adjusted to relax the timeout limit. The dynamic adjustment mechanism of this margin coefficient enables the system's teaching evaluation criteria to adapt to the inherent attributes of different training tasks.

[0051] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI training and teaching system based on machine vision, characterized in that, The system includes: The teaching interface box has at least one power output interface for powering at least one electronic training tool, and the teaching interface box has a built-in current sensor for collecting the current signal flowing through the power line when the electronic training tool is working. The current fingerprint recognition engine is connected to a current sensor and stores a fingerprint template library. The fingerprint template library stores macroscopic current waveform features and common-mode noise spectrum features corresponding to various preset tool states. The current fingerprint recognition engine is configured to: extract macroscopic current waveform features and common-mode noise spectrum features in real time based on the collected current signal; and output an event signal indicating that the current working state of the electronic training tool is the preset tool state when the extracted macroscopic current waveform features and common-mode noise spectrum features match the corresponding features of a preset tool state stored in the fingerprint template library. The cognitive logic engine receives event signals and is configured to determine whether the sequence of received event signals conforms to the teaching operation procedures, which contain multiple operation nodes.

2. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The current fingerprint recognition engine is deployed on the edge computing unit within the teaching interface box; and the macroscopic current waveform characteristics include the current mean and current variance of the current signal within a time window.

3. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The teaching interface box also has a built-in relay. The teaching interface box is configured to: calculate the slope of the current signal change in real time through the edge computing unit, and actively cut off the power supply to the electronic training tool through the relay when the slope of the change exceeds the safety threshold.

4. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The cognitive logic engine is also configured to: upon receiving a new training task instruction, record the actual time consumed by each operation node in at least one successful operation process to generate a baseline time sequence vector, and dynamically update the node timeout threshold used in subsequent judgments based on the baseline time sequence vector.

5. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The system also includes: an image acquisition device for acquiring facial images of trainees; and an eye movement feature analysis module, which is configured to be activated only when the cognitive logic engine determines that the sequence of event signals does not conform to the teaching operation procedures, to perform the following operations: analyze the facial images and calculate the spatial variance of the trainee's pupil center coordinates within the timing period to generate a gaze stability index; and the cognitive logic engine is also configured to determine whether to delay or suppress intervention instructions to trainees based on the gaze stability index.

6. The AI ​​training and teaching system based on machine vision according to claim 4, characterized in that, The cognitive logic engine dynamically updates node timeout thresholds based on a baseline time sequence vector. Specifically, this includes determining a baseline timeout for each operation node based on the baseline time sequence vector. The updated node timeout threshold is calculated using the following formula. : in, The updated node timeout threshold. Based on the baseline time, It is a preset margin coefficient that is greater than zero.

7. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The current fingerprint recognition engine is also configured to provide an initialization interface to guide users to switch the electronic training tools to various preset tool states in sequence, and automatically record the corresponding macroscopic current waveform characteristics and common-mode noise spectrum characteristics during the process to generate a fingerprint template library.

8. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The common-mode noise spectrum characteristics are the location and amplitude of at least one characteristic frequency peak of the current signal in the frequency range of 1 kHz to 50 kHz.

9. The AI ​​training and teaching system based on machine vision according to claim 3, characterized in that, The edge computing unit obtains the slope of change by sampling the current signal twice consecutively at a sampling interval of less than or equal to 50 microseconds and calculating the difference quotient.

10. The AI ​​training and teaching system based on machine vision according to claim 1, characterized in that, The cognitive logic engine is also configured to generate a safety alarm command when it detects that an electronic training tool is continuously in a standby state without being powered off and the duration exceeds an alarm duration threshold.

Images