Modular camera auto focus curve intelligent fitting method based on electric hong operating system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU JINYUAN TECH DEV CO LTD
- Filing Date
- 2026-04-15
- Publication Date
- 2026-07-21
Smart Images
Figure CN122437998A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent vision systems, and in particular to a method for intelligent fitting of autofocus curves of modular cameras based on the Elec-Tech operating system. Background Technology
[0002] With the widespread application of smart cameras in autofocus systems, improving focusing accuracy and environmental adaptability has become a key focus for the industry. Existing camera autofocus technologies typically rely on single sharpness evaluation algorithms and static threshold filtering methods, achieving automated focusing control by fitting the relationship curve between object distance, image plane position, and image sharpness. To further improve focusing speed and robustness, some systems have introduced mechanical displacement pulse counting, motor drive feedback, and joint modeling of environmental parameters such as temperature, humidity, and illuminance, attempting to enhance the adaptability and anti-interference capabilities of the focusing process through multimodal data assistance. With the development of distributed collaborative architectures, systems are increasingly inclined to integrate multi-source heterogeneous data through distributed soft buses and time synchronization technologies, driving the intelligentization, closed-looping, and real-time evolution of focusing algorithms at the edge.
[0003] Currently, mainstream focusing curve fitting and outlier identification methods mainly fall into three categories. First, most products employ traditional median filtering, sliding window elimination, or single statistical distributions (such as the 3σ rule) to screen for outlier data points. These methods are simple and efficient, but cannot handle complex nonlinear noise caused by mechanical tolerances, sudden environmental changes, or dynamic temperature drift, resulting in limited accuracy and robustness in outlier identification. Second, some high-end systems introduce multi-dimensional data filtering based on sensor fusion, improving focusing closed-loop performance through simple data alignment and synchronization. However, these solutions generally lack temporal consistency guarantees, and outlier identification is mainly based on static physical thresholds and preset rules, failing to dynamically perceive changes in the physical causes of noise sources under different operating conditions. The third category attempts to apply edge-side neural network algorithms for pattern recognition in focusing operations. Although this improves intelligent decision-making to some extent, limitations in computing resources and inconsistent data temporal sequences mean that the problem of misidentification / missed detection of outliers caused by fine-grained environmental disturbances and mechanical anomalies remains serious, ultimately making it difficult to guarantee the accuracy and convergence speed of the fitted curve.
[0004] The existing autofocus technology has the following main drawbacks: First, during the cross-node transmission and acquisition of multi-source heterogeneous data, factors such as clock drift and network latency cause a lack of precise nanosecond-level alignment, resulting in asynchronous environmental parameters and mechanical response trajectories. Outlier removal carries the risk of timing misjudgment, and the accuracy of focusing curve fitting is impaired.
[0005] Second, traditional outlier removal mechanisms rely heavily on static thresholds and simple statistical features, making it difficult to distinguish dynamic anomalies introduced by factors such as environmental temperature drift, mechanical structural errors, and sudden lighting. This often results in the erroneous removal of valid transitional data or the erroneous retention of complex noise points, leading to insufficient system robustness.
[0006] Third, existing pattern recognition algorithms are sensitive to nonlinear trajectory changes under complex physical conditions, but their generalization ability and interpretability are limited. They are difficult to automatically adjust the fitting strategy for specific physical disturbances and lack semantic interpretation and traceable data foundation for focus control.
[0007] Fourth, the focus curve fitting stage usually adopts global regression or static segmentation strategy, which cannot achieve dynamic interval incremental refitting and interval collaborative response. As a result, the model cannot adaptively optimize when the environment or mechanical state evolves dynamically, affecting the system's real-time response speed and closed-loop accuracy.
[0008] Therefore, the industry urgently needs a method for intelligent reconstruction of focus curves that can support nanosecond-level alignment of multimodal heterogeneous data and has the ability to detect dynamic causal chain breakpoints and semantic segmentation modeling. Summary of the Invention
[0009] This invention provides a modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system, aiming to solve the problems existing in the prior art mentioned in the background section.
[0010] The technical solution of this invention is: a modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system, comprising the following steps: S1: Based on the distributed soft bus of the Elec-Tech operating system, timestamp broadcast calibration is performed on the camera optical module, image acquisition unit and environmental sensor node to generate multimodal raw data stream; S2: Map the multimodal raw data stream to the same time axis for cross-node temporal alignment processing to generate a multidimensional feature trajectory sequence; S3: Based on the phase offset of the peak image sharpness, lens position pulse count and environmental parameter change rate in the multidimensional feature trajectory sequence, construct the causal trajectory consistency criterion and generate a set of transient points to be verified. S4: Using a lightweight neural network inference framework on the terminal side, the set of transient points to be verified is input into a pre-trained local trajectory pattern recognition model for sliding window analysis to generate dynamic decay weight labels that match typical nonlinear transition modes. S5: For the remaining data points that do not match typical nonlinear transition modes, perform similarity retrieval and residual energy comparison with the same trajectory segments of historical working conditions to generate a pure focus data point set; S6: Based on the dominant influencing factor labels of the pure focal data point set, the object distance-image plane position space is divided into several local intervals with independent physical causes, and local model parameter sets with specific fitting basis functions and learning rates for each interval are generated. S7: Based on the newly added data points, perform incremental refitting on the local model parameter group of the local interval, and trigger the collaborative convergence of adjacent intervals by broadcasting parameter disturbance information through a distributed soft bus to generate a segmented focusing curve.
[0011] Preferably, in S1, under the distributed soft bus topology of the Elec-Tech operating system, the unique hardware identifiers and network connection status of the camera optical module, image acquisition unit, temperature and humidity sensor node, and motor drive controller are loaded as initial conditions into the clock synchronization management module of the main control chip. Preferably, in S1, in a distributed sensor network clock environment in a steady-state synchronization state, the master control chip is selected as the global control source.
[0012] Preferably, S3 includes the following steps: Extreme point detection processing is performed on the multidimensional feature trajectory sequence. The sliding window differential algorithm is used to extract the local maximum moments of the image sharpness evaluation function value and the zero-crossing moments of the first derivative of the lens position pulse count, generating a transient feature time stamp set. Based on the transient feature time stamp set, the ambient illuminance change rate signal and temperature gradient signal are subjected to synchronous time interception operation. The light intensity change value and thermal drift rate value corresponding to the sharpness peak time and the mechanical steady state time are calculated to generate the instantaneous state vector group of environmental parameters. Perform temporal alignment difference operation on the transient feature time stamp set and the instantaneous state vector group of environmental parameters to calculate the time lag of image sharpness response relative to lens mechanical displacement and the phase delay angle of environmental parameter abrupt change relative to sharpness response, and generate phase offset; Based on the phase offset, the preset causal logic rule base is called to perform threshold comparison and logic gate judgment processing. The phase delay angle exceeding the dynamic tolerance window is judged as a causal chain break event, and an initial abnormal point index list is generated. Spatiotemporal neighborhood aggregation is performed on the initial list of outlier points to merge discrete data points marked as causal chain break events within continuous time slices into non-steady-state process segments and bind the corresponding phase offset feature vectors to generate a set of transient points to be verified marked with non-steady-state transition processes.
[0013] Preferably, step S4 includes the following steps: A pre-trained local trajectory pattern recognition model is loaded using a lightweight neural network inference framework on the edge. A sliding window truncation process is performed on the multi-dimensional feature trajectory sequence within a fixed time window before and after each data point in the transient point set to be verified, so as to generate a standardized time-series feature fragment containing image sharpness peak, lens position pulse count, ambient illuminance change rate and temperature gradient. Using standardized temporal feature fragments as input, multi-scale spatiotemporal features are extracted through the convolutional coding layer in the local trajectory pattern recognition model. The standardized temporal feature fragments are then mapped into high-dimensional latent space feature vectors to obtain deep pattern feature encoding. Based on deep pattern feature encoding, similarity measurement calculation with pre-set typical nonlinear transition pattern prototypes is performed in the classification decision layer of the local trajectory pattern recognition model. The deep pattern feature encoding is compared with the standard pattern prototypes stored in the pattern library using Euclidean distance to generate a probability distribution confidence matrix. Based on the maximum confidence value in the probability distribution confidence matrix and its corresponding category index, a dynamic decay weight mapping strategy is executed to convert the probability distribution confidence matrix into dynamic decay weight labels. By integrating the dynamic decay weight labels with the original transient point identifier information to be verified, an enhanced transient point dataset with pattern matching results is generated.
[0014] Preferably, step S5 includes the following steps: Based on the remaining data points of the unmatched typical nonlinear transition mode, its multidimensional time series feature vector is extracted, and the standard trajectory segment set under the same object distance interval is extracted from the historical working condition database using the distributed storage interface of the Elec-Tech operating system. A historical working condition feature index library containing timestamp alignment information and dominant influencing factor labels is constructed. Based on the standard trajectory fragment set in the historical working condition feature index library, the multidimensional time series feature vector of the remaining data points is processed by the dynamic time warping algorithm. The cumulative distance measurement value between the current test trajectory and each historical standard trajectory in three dimensions—image clarity, lens position pulse count, and environmental parameter change rate—is calculated, and a similarity retrieval result sequence representing the similarity of trajectory morphology is generated. Based on candidate noise points with similarity below a preset threshold in the similarity retrieval result sequence, and combined with the motor drive current waveform and lens mechanical vibration spectrum during the current zoom process, a local residual energy calculation model is constructed. Least square fitting residual analysis is performed to quantify the deviation energy between the actual observed value and the theoretical predicted value, and a residual energy comparison index characterizing the mechanical tolerance or temperature drift noise intensity is generated. Based on the residual energy comparison index and the background noise level benchmark, an adaptive dual-threshold decision logic processing is performed to mark data points whose residual energy is higher than the dynamic noise threshold and whose similarity retrieval result sequence shows low correlation as real outliers, while retaining data points whose residual energy is within the tolerance range as valid observations, and generating a data point classification list containing removal and retention labels. Based on the retained marked data points in the data point classification list, spatiotemporal coordinate reconstruction and confidence-weighted aggregation are performed to remove all abnormal data marked as true outliers and integrate them to form a clean set of focus data points.
[0015] Preferably, step S6 includes the following steps: Based on the dominant influencing factor label carried by each data point in the pure focus data set, cluster analysis is used to generate multiple physical cause homogeneous data subsets. Based on the distribution boundaries of the multiple homogeneous physical cause data subsets in the object distance dimension, an interval segmentation algorithm is executed to generate several independent physical cause local intervals that do not overlap and cover the entire range. For each independent physical cause local interval containing data point distribution characteristics, the radial basis function library is called to perform kernel function matching and filtering to generate a set of exclusive fitting basis functions that fit the nonlinear characteristics of the interval. Based on the complexity index of the dedicated fitting basis function set and the variance of the data point density within the local interval of the independent physical cause, an adaptive gain calculation is performed to generate a dedicated adaptive learning rate scalar for each interval. The dedicated set of fitting basis functions and the dedicated adaptive learning rate scalar are structurally encapsulated and bound to generate local model parameter sets corresponding to each independent local interval of physical cause.
[0016] Preferably, S7 includes the following steps: Obtain the newly added pure focus data point set and its corresponding dominant influence factor labels, construct a dedicated fitting basis function set based on the radial basis function kernel method, and calculate the local model parameter set including the initial weight coefficients; The local model parameter set is used to perform residual gradient descent on the newly added clean focal data point set to generate a parameter update vector containing the learning rate adjustment factor, so as to correct the local fitting bias caused by mechanical tolerance or temperature drift and output the optimized local model parameter set. Based on the optimized local model parameter set, the boundary gradient perturbation value is calculated. The parameter perturbation information packet containing the boundary gradient perturbation value is broadcast to adjacent physical cause local interval nodes through the Dianhong distributed soft bus to trigger the linkage response mechanism of the neighborhood model and generate cross-interval collaborative convergence instructions. It receives cross-interval collaborative convergence instructions and performs weighted smoothing fusion processing on the local model parameter groups of adjacent intervals to generate a global continuous piecewise focusing curve function; The semantic feature identifiers of each segment interval are extracted based on the global continuous segmented focusing curve function, and the dominant influencing factor labels are mapped to the curve function segments to generate segmented focusing curves with semantic interpretability.
[0017] Preferably, the method further includes: S8: Based on the real-time prediction and compensation of the segmented focus curve, the zoom motor group is driven to perform a precise focusing action.
[0018] Preferably, S8 includes the following steps: Obtain the target object distance interval identifier and the corresponding exclusive fitting basis function parameters in the segmented focusing curve. Use the lightweight neural network inference framework of the Elec-Tech OS to load the pre-trained motor dynamic response prediction model. Use the target object distance interval identifier and the ambient temperature parameter as input vectors to perform forward propagation calculation and generate a motor dynamic characteristic descriptor that characterizes the motor mechanical hysteresis and electrical time constant under the current working condition. Based on the motor dynamic characteristic descriptor, a discrete state-space equation containing inertial and damping terms is constructed. The inverse system solution process is performed on the theoretical image plane position sequence output by the piecewise focusing curve to generate a theoretical motor pulse count target trajectory sequence containing the predicted lead amount. The current real-time rotor position pulse count value and speed feedback signal of the zoom motor are collected. The instantaneous position deviation between the theoretical motor pulse count target trajectory sequence and the real-time rotor position pulse count value is calculated. An adaptive proportional-integral-derivative adjustment algorithm is introduced to correct the instantaneous position deviation and generate a motor drive control voltage instruction set optimized by delay compensation. The motor drive control voltage command set is sent to the motor drive controller node via the distributed soft bus of the electric motor, driving the zoom motor group to perform precise focusing action according to the theoretical motor pulse count target trajectory sequence.
[0019] The beneficial technical effects of this invention are as follows: 1) This invention constructs a low-overhead, high-precision time synchronization channel through the distributed soft bus of the Elec-Tech operating system, achieving nanosecond-level clock alignment between the camera optical module, image acquisition unit, temperature and humidity sensor, illuminance sensor, and motor drive controller; before each zoom start, it automatically triggers full-link timestamp broadcast calibration to ensure that each frame of image, each pulse count, and each set of environmental parameters are bound to a unified spatiotemporal coordinate, thereby mapping heterogeneous data streams to the same time axis and forming a strictly time-consistent multidimensional feature trajectory sequence; 2) This invention introduces a causal trajectory consistency criterion and combines it with a lightweight edge neural network to construct a two-level intelligent verification mechanism. When a focal data point exhibits asynchronous responses in dimensions such as peak sharpness, lens position, rate of change of illumination, and temperature gradient, the system marks it as a transient point to be verified instead of directly removing it. Subsequently, a pre-trained local trajectory pattern recognition model is used to perform sliding window analysis to determine whether it belongs to a known nonlinear transition mode. If a match is successful, it is retained and given dynamic decay weights to participate in fitting. For unmatched points, a second-level judgment is further performed through historical operating condition similarity retrieval and residual energy analysis. Only when both conditions are met is it removed as a true outlier. 3) This invention adopts a segmented adaptive fitting architecture based on dynamic interval division. According to the aforementioned trajectory analysis results, the system divides the object distance-image plane space into multiple local regions with independent physical causes. Each region is configured with a dedicated basis function and learning rate, and its local parameters are updated only when new data belongs to the corresponding interval. At the same time, the system broadcasts perturbation information to neighboring nodes through a soft bus to trigger collaborative convergence. The final generated focus curve is a set of semantically interpretable segmented results, with each segment accompanied by a dominant influencing factor label. This not only improves the robustness and convergence stability of the model in a variable environment, but also provides a traceable and analyzable technical path for system fault attribution, performance diagnosis, and continuous evolution. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention; Figure 2 A schematic diagram illustrating the process of generating a set of transient points to be verified for an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating dynamic decay weight labels in an embodiment of the present invention. Detailed Implementation
[0021] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0022] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0023] likeFigure 1 As shown, this embodiment provides a modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system, including the following steps: S1: Based on the distributed soft bus of the Elec-Tech operating system, timestamp broadcast calibration is performed on the camera optical module, image acquisition unit, and environmental sensor nodes to generate a multimodal raw data stream; details are as follows: S1.1: Based on the distributed soft bus topology of the Elec-Tech operating system, a global clock synchronization request command is sent to the camera optical module, image acquisition unit, temperature and humidity sensor node and motor drive controller to obtain the current local oscillator frequency deviation value and phase offset of each slave node, and generate a set of synchronization handshake messages containing node identifier and initial clock status information. Under the distributed soft bus topology of the Elec-Tech OS, the unique hardware identifiers and network connection status of the camera optical module, image acquisition unit, temperature and humidity sensor node, and motor drive controller are loaded as initial conditions into the clock synchronization management module of the main control chip. For each slave node, a global clock synchronization request command is sent through the low-latency communication channel of the soft bus, ensuring that the unique node identifier and request timestamp are included in the command broadcast process. Upon receiving the request, each slave node calls the local oscillator monitoring loop to measure its current oscillation frequency, obtains the local frequency characteristic value based on a high-precision counter, and records the phase offset from the master control reference time. The above frequency characteristic value and phase offset are encapsulated into the payload field of the synchronization handshake message, and node identification information is added so that the master control chip can distinguish the source node and perform subsequent delay compensation calculations. All handshake messages are returned to the master control chip through the distributed soft bus, forming a set of synchronization handshake messages containing multiple node identifiers, local oscillator frequency deviations, and phase offsets. By binding a unified format timestamp to each handshake message, the node status information of the previous step is transformed into a standardized data set that can be used for bidirectional timestamp exchange, realizing the preparation conditions for nanosecond-level time calibration.
[0024] S1.2: Perform bidirectional timestamp exchange algorithm processing on the synchronous handshake message set, use the main control chip as the time reference source to calculate the propagation delay compensation factor and crystal oscillator drift correction coefficient of each slave node, so as to eliminate the timing error caused by the physical link transmission delay and generate a dynamic clock calibration parameter set with nanosecond-level accuracy. S1.3: Based on the dynamic clock calibration parameter group, perform closed-loop feedback adjustment operation on the internal timing register of each node to lock the dispersed local clock sources to a unified global virtual clock domain, so as to establish a full-link microsecond-level time synchronization channel and generate a distributed sensor network clock environment in a steady-state synchronization state. S1.4: Utilizing the clock environment of the distributed sensor network in a steady-state synchronization state, a synchronization trigger signal with a globally unique sequence number is broadcast to all data acquisition nodes at the moment zoom is started, so as to force each node to capture the image sharpness evaluation function value, lens position pulse count, ambient illuminance and temperature parameters within the same absolute time slice, and generate a discrete multi-source sensing data packet carrying a unified global timestamp. In a distributed sensor network clock environment in a steady-state synchronization state, the master control chip is selected as the global control source. A synchronization trigger signal parameter package with a globally unique sequence number is constructed, which includes an absolute time slice identifier, acquisition mode settings, and node execution priority index. Based on the low-latency transmission path of the Elec-Tech distributed soft bus, the synchronization trigger signal is broadcast to all data acquisition nodes, including the camera optical module, image acquisition unit, temperature and humidity sensor, and illuminance sensor. Within the same absolute time slice when the node receives the trigger signal, the acquisition interrupt service routine inside each node is called to latch the current status register value and start the acquisition process. A precise interrupt scheduling strategy is adopted to ensure that each node completes the calculation of image sharpness evaluation function value, accumulation of lens position pulse count, instantaneous sampling of ambient illuminance, and instantaneous sampling of temperature within a microsecond-level synchronization window. The above multi-source data are then initially encapsulated in raw numerical form along with a global timestamp. The encapsulated data from each node is sent back to the master control chip in a distributed queue, where field verification and sequence number consistency verification are performed. All verified data are aggregated into a structured discrete multi-source sensing data packet. On the main control chip side, global spatiotemporal coordinate information is uniformly attached to each data packet to ensure that the batch of data has an accurate time synchronization reference during cross-node processing and subsequent feature trajectory construction. By using a unified global timestamp to eliminate the drift of the start and end times of acquisition, strict synchronization of data acquisition on the end side is achieved.
[0025] S1.5: Perform spatiotemporal coordinate binding and encapsulation processing on the discrete multi-source sensing data packets carrying a unified global timestamp, mapping optical position information, image quality indicators and environmental physical quantities to the same four-dimensional spatiotemporal reference system to form logically indivisible data atom units, generating a multimodal raw data stream with unified spatiotemporal coordinates.
[0026] S2: Map the multimodal raw data stream to the same time axis for cross-node temporal alignment processing to generate a multidimensional feature trajectory sequence; details are as follows: S2.1: Perform non-uniform sampling interval analysis on the multimodal raw data stream with unified spatiotemporal coordinates, calculate the absolute time offset of each sensor node relative to the master control clock reference based on the network transmission jitter delay parameters recorded by the Dianhong distributed soft bus, and generate a biased raw data set containing accurate timestamp correction information. For multimodal raw data streams with unified spatiotemporal coordinates, the transmission event log interface of the Elec-Tech distributed soft bus is called to obtain the network transmission jitter delay parameters generated by each node during the sending and receiving of data packets, forming a delay record table containing node identifiers and delay values.
[0027] By performing a differential operation between each delay value in the delay record table and the timestamp record of the master clock reference, a preliminary time offset dataset for each sensor node in the current sampling session is obtained.
[0028] S2.2: Based on the absolute time offset in the biased original dataset, a hybrid algorithm of linear interpolation and cubic spline interpolation is used to resample the image sharpness evaluation function value, lens position pulse count and environmental parameters. The data with heterogeneous sampling frequencies are uniformly converted to a preset microsecond-level standard time grid to generate a resampled data sequence with equal time intervals. S2.3: Perform cross-node phase synchronization verification processing on the resampled data sequence with equal time intervals, compensate for the inherent time delay difference between the optical module and the motor encoder based on the photoelectric signal propagation physical model, eliminate the time misalignment between mechanical response and image acquisition, and generate phase-aligned multi-source synchronized data frames. S2.4: Based on phase-aligned multi-source synchronous data frames, vector splicing and feature normalization techniques are used to map the values of four dimensions—peak image sharpness, lens position pulse count, ambient illuminance change rate, and temperature gradient—into high-dimensional state vectors, thereby constructing a composite feature vector set that reflects the comprehensive state of the system at a single moment. S2.5: Perform sliding window serialization assembly on the composite feature vector set in chronological order, form a continuous temporal evolution trajectory through chain connection, and finally generate a multi-dimensional feature trajectory sequence with strict temporal consistency, which serves as the direct input object for the subsequent construction of causal trajectory consistency criteria.
[0029] S3: Based on the phase offset of image sharpness peaks, lens position pulse counts, and environmental parameter change rates in multidimensional feature trajectory sequences, a causal trajectory consistency criterion is constructed, generating a set of transient points to be verified; such as... Figure 2 As shown, the details are as follows: S3.1: Perform extreme point detection processing on the multidimensional feature trajectory sequence with strict temporal consistency, and use the sliding window differential algorithm to extract the local maximum moment of the image sharpness evaluation function value and the zero-crossing moment of the first derivative of the lens position pulse count, and generate a transient feature time stamp set containing sharpness peak timestamp and mechanical steady state timestamp; S3.2: Based on the transient feature time stamp set, perform synchronous time interception operation on the ambient illuminance change rate signal and temperature gradient signal, calculate the light intensity change value and thermal drift rate value corresponding to the sharpness peak time and the mechanical steady state time, and generate an instantaneous state vector group of environmental parameters characterizing the immediate impact of environmental disturbances. S3.3: Perform time-domain aligned difference operation on the transient feature time stamp set and the instantaneous state vector group of environmental parameters to calculate the time lag of the image sharpness response relative to the lens mechanical displacement and the phase delay angle of the environmental parameter mutation relative to the sharpness response, and generate a multi-dimensional phase offset dataset that reflects the coupling response characteristics of multiple physical quantities. S3.4: Based on the multidimensional phase offset dataset, call the preset causal logic rule base to perform threshold comparison and logic gate judgment processing, determine the phase delay angle exceeding the dynamic tolerance window as a causal chain break event, and generate an initial abnormal point index list that indicates the system is in an unbalanced transition state. S3.5: Perform spatiotemporal neighborhood aggregation processing on the initial anomaly point index list, merge discrete data points marked as causal chain break events within continuous time slices into non-steady-state process segments with duration, and bind the corresponding phase offset feature vectors, finally generating a set of transient points to be verified marked with non-steady-state transition processes.
[0030] S4: Utilizing a lightweight neural network inference framework on the terminal side, the set of transient points to be verified is input into a pre-trained local trajectory pattern recognition model for sliding window analysis to generate dynamically decaying weight labels that match typical nonlinear transition modes; such as... Figure 3 As shown, the details are as follows: S4.1: Based on the distributed soft bus transmission of the Elec-Tech operating system, the set of transient points to be verified is used to load the pre-trained local trajectory pattern recognition model by calling the lightweight neural network inference framework on the end side. The sliding window truncation process is performed on the multi-dimensional feature trajectory sequence within a fixed time window before and after each data point in the set of transient points to be verified to generate a standardized time-series feature segment containing four-dimensional information including image sharpness peak, lens position pulse count, ambient illuminance change rate and temperature gradient. For example, in the modular camera mounted on a power line inspection drone, the set of transient points to be verified includes four data sources with node identifiers CAM01, MOTOR02, LUX01, and TEMP01, with a unified global timestamp of 162345678901234 nanoseconds. The edge-side lightweight neural network inference framework sets the time window length to 50 milliseconds and the sampling step size to 1 millisecond when calling the local trajectory pattern recognition model, collecting a total of 50 time-series samples within the coverage window. The original range of the image sharpness peak signal column is 0 to 500, which is normalized and mapped to 0 to 1; the range of the lens position pulse count column is 0 to 2000, which is also normalized and mapped to 0 to 1; the range of the ambient illumination change rate is... 5 to 5 lux / millisecond, normalized to 1 to 1; the temperature gradient range is 0.2 to 0.2 degrees Celsius / millisecond, normalized to 1 to 1. During the time window index alignment process, it was found that two frames of illuminance change rate data for LUX01 were missing. The missing values were filled in using cubic spline interpolation, and phase alignment compensation was performed within a range of 0.8 milliseconds using the inherent time lag difference parameters between the motor and the optical module. The final generated standardized temporal feature fragment matrix has a dimension of 50×4, with four columns corresponding to the normalized image sharpness peak, lens position pulse count, ambient illuminance change rate, and temperature gradient, respectively. The matrix is used as an input vector and fed into the local trajectory pattern recognition model to achieve accurate matching analysis of patterns such as temperature drift-induced gradual change and mechanical rebound oscillation. The output feature encoding significantly improved the pattern recognition capability and fitting accuracy of the focus curve in subsequent steps.
[0031] S4.2: Using the standardized temporal feature fragments as input vectors, multi-scale spatiotemporal feature extraction is performed through the convolutional coding layer in the local trajectory pattern recognition model to map the standardized temporal feature fragments into high-dimensional latent space feature vectors, so as to obtain deep pattern feature encodings that can characterize the essential attributes of typical nonlinear transition modes such as temperature drift-induced gradual change, mechanical rebound oscillation, or light abrupt change adaptation. S4.3: Based on the deep pattern feature encoding, perform similarity measurement calculation with the preset typical nonlinear transition pattern prototype in the classification decision layer of the local trajectory pattern recognition model, and perform Euclidean distance comparison analysis between the deep pattern feature encoding and the standard pattern prototype stored in the pattern library to generate a probability distribution confidence matrix characterizing the current transient point to be verified to belong to a specific physical cause category. S4.4: Based on the maximum confidence value in the probability distribution confidence matrix and its corresponding category index, execute the dynamic decay weight mapping strategy to convert the probability distribution confidence matrix into a dynamically decay weight label with physical interpretability, so as to quantify the credibility of the transient point to be verified belonging to the effective transition process and assign it a differentiated contribution weight in the subsequent fitting. S4.5: Integrate the dynamic decay weight labels and the original transient point identification information to be verified to generate an enhanced transient point dataset with pattern matching results. Output the enhanced transient point dataset to the subsequent clean focus data point set generation module to support the remaining data points that do not match typical patterns to enter the secondary verification process and complete the final focus curve reconstruction.
[0032] S5: For the remaining data points that do not match typical nonlinear transition modes, perform similarity retrieval and residual energy comparison with trajectory segments of the same type as historical working conditions to generate a clean set of focal data points; as detailed below: S5.1: Based on the remaining data points of the unmatched typical nonlinear transition modes generated in the previous steps, extract their multi-dimensional time-series feature vectors, and use the distributed storage interface of the Dianhong operating system to extract the set of standard trajectory segments under the same object distance interval from the historical working condition database, and construct a historical working condition feature index library containing timestamp alignment information and dominant influencing factor labels to establish a benchmark reference system for similarity comparison. S5.2: Based on the set of standard trajectory segments in the historical working condition feature index library, the dynamic time warping algorithm is applied to the multidimensional time-series feature vector of the remaining data points to calculate the cumulative distance measurement value between the current trajectory to be tested and each historical standard trajectory in three dimensions: image clarity, lens position pulse count and environmental parameter change rate, and generate a similarity retrieval result sequence that characterizes the similarity of trajectory morphology. S5.3: Based on the candidate noise points with similarity below the preset threshold in the similarity retrieval result sequence, and combined with the motor drive current waveform and lens mechanical vibration spectrum during the current zoom process, a local residual energy calculation model is constructed, least squares fitting residual analysis is performed, the deviation energy between the actual observed value and the theoretical predicted value is quantified, and a residual energy comparison index characterizing the mechanical tolerance or temperature drift noise intensity is generated. The lens mechanical vibration spectrum signal is subjected to Fast Fourier Transform (FFT) and peak frequency band truncation processing to extract the inherent vibration mode frequency and its harmonic distribution of the mechanical system in the time interval, which are used as quantitative indicators of mechanical disturbance intensity and bound to the multidimensional feature vector corresponding to the candidate noise point.
[0033] Based on the electrical characteristic curve of the motor and the spectrum of mechanical disturbance, a local residual energy calculation model is constructed. The theoretical prediction value sequence and the actual observation value sequence are input into the model, and least squares fitting residual analysis is performed to obtain the residual energy value. The formula for calculating the least squares residual is as follows: in, For the first The actual observed values of each data point For the first Predicted values for each data point For the number of data points, This represents the residual energy value.
[0034] The residual energy value and the amplitude of the vibration peak frequency band are normalized to form a quantitative residual energy comparison index that characterizes the mechanical tolerance or temperature drift noise intensity, and the index is associated with and stored with the candidate noise point identification information.
[0035] S5.4: Based on the residual energy comparison index and the background noise level benchmark, perform adaptive dual threshold decision logic processing, mark data points with residual energy higher than the dynamic noise threshold and low correlation in the similarity retrieval result sequence as real outliers, and retain data points with residual energy within the tolerance range as valid observations, and generate a data point classification list containing removal and retention labels; S5.5: Based on the retained labeled data points in the data point classification list, perform spatiotemporal coordinate reconstruction and confidence-weighted aggregation processing to remove all abnormal data marked as true outliers and integrate them to form a pure focal data point set containing only high-reliability observations, which serves as the sole input source for subsequent local interval-specific fitting basis function construction and adaptive learning rate adjustment.
[0036] S6: Based on the dominant influencing factor labels of the pure focal data point set, the object distance-image plane position space is divided into several local intervals with independent physical causes, generating local model parameter sets with specific fitting basis functions and learning rates for each interval; as detailed below: S6.1: Perform cluster analysis on the dominant influencing factor label carried by each data point in the pure focus data point set to generate multiple physical cause homogeneous data subsets, including the thermal drift dominant cluster, the mechanical hysteresis compensation cluster, and the light adaptation transition cluster. For example, in a power line inspection scenario, the clean focal data set contains 500 observation points, with the dominant influencing factor labels distributed as follows: thermal drift (200 points), mechanical hysteresis (150 points), and illumination adaptation (150 points). One-hot encoding is performed on these three types of labels to form a three-dimensional state vector. For example, the thermal drift label is encoded as follows: Mechanical hysteresis is coded as Light adaptation coding is For each data point, calculate the weighted Euclidean distance between the object distance parameter (unit: meters) and the tag state vector. The formula is: ; in, The object distance weight is set to 0.7. The tag similarity weight is set to 0.3. For the first With the Data point distance difference This represents the difference between the label state vectors.
[0037] After constructing a 500×500 distance matrix, agglomerative hierarchical clustering is performed, and the dynamic threshold is calculated using the following formula: in, The maximum distance value. The total number of data points. It is the distribution density factor (calculated from the standard deviation of the data points on the object distance axis).
[0038] After calculating the threshold, clustering and merging were stopped, and the 500 data points were finally stably divided into three clusters. In this scenario, the output thermally drift-dominant cluster, mechanical hysteresis compensation cluster, and illumination adaptation transition cluster significantly improved the matching accuracy of the fitting basis function selection in the subsequent S6.2 interval partitioning. Moreover, in the actual scenario verification, the object distance boundaries of the three clusters did not overlap, achieving clear separation of physical causes and a significant improvement in model adaptability.
[0039] S6.2: Based on the distribution boundaries of the multiple homogeneous physical cause data subsets in the object distance dimension, perform interval segmentation algorithm processing to generate several independent physical cause local intervals that do not overlap and cover the entire range. S6.3: For the distribution characteristics of data points contained in each independent physical cause local interval, the radial basis function library is called to perform kernel function matching and filtering to generate a set of exclusive fitting basis functions that are suitable for the nonlinear characteristics of the interval. S6.4: Based on the complexity index of the dedicated fitting basis function set and the variance of the data point density within the local interval of the independent physical cause, perform adaptive gain calculation to generate a dedicated adaptive learning rate scalar for each interval. S6.5: The dedicated fitting basis function set and the dedicated adaptive learning rate scalar are structurally encapsulated and bound to generate local model parameter sets corresponding to each independent physical cause local interval, thus completing the initial configuration of the focus curve segmented modeling resources.
[0040] S7: Based on the newly added data points, perform incremental refitting on the local model parameter group of the corresponding local interval, and trigger collaborative convergence of adjacent intervals by broadcasting parameter perturbation information through a distributed soft bus, generating a piecewise focusing curve; as detailed below: S7.1: Obtain the newly added pure focus data point set and its corresponding dominant influencing factor labels, construct a dedicated fitting basis function set based on the radial basis function kernel method, and calculate the local model parameter set including the initial weight coefficients to establish the benchmark fitting shape of the current object distance interval. Based on the local interval invocation radial basis function kernel method, kernel function mapping is performed on existing clean focal data points and newly added data points within the interval to form an initial kernel matrix structure with a center vector and bandwidth parameters. The response values of the radial basis functions in the object distance-image plane position space are calculated based on this initial kernel matrix structure. The kernel function response values are associated and stored with the dominant influencing factor labels carried by each data point, constructing a feature map set containing multi-kernel weight distributions.
[0041] Initial weight coefficients are calculated for the feature mapping set, and the weight coefficients are solved using the least squares method to obtain the coefficient matrix between the target output and the kernel function response. The formula is: ; in, The kernel function response matrix; This is the target image plane position value vector.
[0042] S7.2: Use the local model parameter set to perform residual gradient descent operation on the newly added clean focal data point set to generate a parameter update vector containing an adaptive learning rate adjustment factor, so as to correct the local fitting deviation caused by mechanical tolerance or temperature drift and output the optimized local model parameter set. S7.3: Calculate the boundary gradient perturbation value based on the optimized local model parameter set, and broadcast the parameter perturbation information packet containing the boundary gradient perturbation value to the adjacent physical cause local interval nodes through the Dianhong distributed soft bus, so as to trigger the linkage response mechanism of the neighborhood model and generate cross-interval collaborative convergence instructions. S7.4: Receive the cross-interval collaborative convergence instruction and perform weighted smoothing fusion processing on the local model parameter groups of adjacent intervals to generate a global continuous piecewise focusing curve function that eliminates abrupt changes in the gaps between intervals, so as to ensure the continuity of focus prediction in the switching regions of different dominant influencing factors. The input conditions for receiving cross-interval collaborative convergence instructions based on the optimized local model parameter set include the boundary gradient perturbation value sequence, the dominant influence factor labels of adjacent intervals, and the fitted basis function parameter matrix of each interval. During the smooth fusion process, a weight allocation matrix is constructed for the local model parameter sets of adjacent intervals according to the similarity of the dominant influence factors. Normalization is performed on the weight allocation matrix to maintain the consistency of physical quantity dimensions while ensuring that the sum of each weight is 1. A weighted arithmetic mean strategy is used to fuse the optimized parameter set with the parameter sets of adjacent intervals point by point within the boundary buffer, forming a smooth parameter transition sequence. For boundaries with nonlinear changes, the radial basis function interpolation method is used to perform local curvature correction on the fusion result, ensuring the continuity of the second derivative of the curve at the boundary. The fused parameter values are calculated using the following formula. : ; in, The current interval parameter value, For adjacent interval parameter values, The fusion weights are determined based on the similarity of the impact factors.
[0043] Phase consistency verification is performed on the fused global continuous segmented focus curve function to ensure that the focus prediction in different dominant influencing factor switching areas remains continuous in time and space.
[0044] S7.5: Based on the global continuous segmented focusing curve function, extract the semantic feature identifiers of each segment interval, map the dominant influencing factor labels to the curve function segments to generate a segmented focusing curve with semantic interpretability, so as to complete the adaptive evolution of the focusing curve and provide a control basis with physical meaning for precise focusing action.
[0045] S8: Based on the segmented focus curve, it predicts and compensates for the focus control command in real time, driving the zoom motor assembly to perform precise focusing actions; specifically as follows: S8.1: Obtain the target object distance interval identifier and the corresponding exclusive fitting basis function parameters in the segmented focusing curve, load the pre-trained motor dynamic response prediction model using the lightweight neural network inference framework of the Elec-Tech OS, and perform forward propagation calculation with the target object distance interval identifier and the ambient temperature parameter as input vectors to generate a motor dynamic characteristic descriptor that characterizes the motor mechanical hysteresis and electrical time constant under the current working condition. S8.2: Based on the motor dynamic characteristic descriptor, construct a discrete state-space equation containing inertia and damping terms, and perform inverse system solution processing on the theoretical image plane position sequence output by the segmented focusing curve to offset the inherent phase lag effect of the motor and generate a theoretical motor pulse count target trajectory sequence containing the predicted lead amount. S8.3: Collect the current real-time rotor position pulse count value and speed feedback signal of the zoom motor group, calculate the instantaneous position deviation between the theoretical motor pulse count target trajectory sequence and the real-time rotor position pulse count value, introduce an adaptive proportional-integral-derivative adjustment algorithm to perform closed-loop correction calculation on the instantaneous position deviation, and generate a motor drive control voltage instruction set optimized by delay compensation. S8.4: The motor drive control voltage instruction set is sent to the motor drive controller node via the distributed soft bus of the electric motor, which drives the zoom motor group to perform a precise focusing action according to the theoretical motor pulse count target trajectory sequence. After the action is completed, the new image sharpness evaluation function value and the actual lens position pulse count are captured simultaneously to generate a real-time end-side calibration feedback data pair for verifying the focusing accuracy. S8.5: Based on the residual energy distribution characteristics in the real-time calibration feedback data pair at the end side, update the weight parameters of the motor dynamic response prediction model and correct the adaptive learning rate of the dedicated fitting basis function parameters to complete the adaptive evolution iteration of the focusing curve and form a closed loop of real-time calibration at the end side with continuous self-optimization capability.
[0046] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0047] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0048] Unless otherwise defined, the technical or scientific terms used herein should be understood in their ordinary sense by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” involved in the embodiments of this invention refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0049] The above description is merely an exemplary embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system, characterized in that, Includes the following steps: S1: Based on the distributed soft bus of the Elec-Tech operating system, timestamp broadcast calibration is performed on the camera optical module, image acquisition unit and environmental sensor node to generate multimodal raw data stream; S2: Map the multimodal raw data stream to the same time axis and perform cross-node time alignment processing to generate a multidimensional feature trajectory sequence; S3: Based on the phase offset of the peak image sharpness, lens position pulse count and environmental parameter change rate in the multidimensional feature trajectory sequence, construct a causal trajectory consistency criterion and generate a set of transient points to be verified. S4: Using a lightweight neural network inference framework on the terminal side, the set of transient points to be verified is input into a pre-trained local trajectory pattern recognition model for sliding window analysis to generate dynamic decay weight labels that match typical nonlinear transition modes. S5: For the remaining data points that do not match typical nonlinear transition modes, perform similarity retrieval and residual energy comparison with the same trajectory segments of historical working conditions to generate a pure focus data point set; S6: Based on the dominant influencing factor labels of the pure focal data point set, the object distance-image plane position space is divided into several local intervals with independent physical causes, and local model parameter sets with specific fitting basis functions and learning rates for each interval are generated. S7: Based on the newly added data points, perform incremental refitting on the local model parameter group of the local interval, and trigger the collaborative convergence of adjacent intervals by broadcasting parameter disturbance information through a distributed soft bus to generate a segmented focusing curve.
2. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, S3 specifically includes the following steps: Extreme point detection processing is performed on the multidimensional feature trajectory sequence. The sliding window differential algorithm is used to extract the local maximum moments of the image sharpness evaluation function value and the zero-crossing moments of the first derivative of the lens position pulse count, generating a transient feature time stamp set. Based on the transient feature time stamp set, the ambient illuminance change rate signal and temperature gradient signal are subjected to synchronous time interception operation. The light intensity change value and thermal drift rate value corresponding to the sharpness peak time and the mechanical steady state time are calculated to generate the instantaneous state vector group of environmental parameters. Perform temporal alignment difference operation on the transient feature time stamp set and the instantaneous state vector group of environmental parameters to calculate the time lag of image sharpness response relative to lens mechanical displacement and the phase delay angle of environmental parameter abrupt change relative to sharpness response, and generate phase offset; Based on the phase offset, the preset causal logic rule base is called to perform threshold comparison and logic gate judgment processing. The phase delay angle exceeding the dynamic tolerance window is judged as a causal chain break event, and an initial abnormal point index list is generated. Spatiotemporal neighborhood aggregation is performed on the initial list of outlier points to merge discrete data points marked as causal chain break events within continuous time slices into non-steady-state process segments and bind the corresponding phase offset feature vectors to generate a set of transient points to be verified marked with non-steady-state transition processes.
3. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, S4 specifically includes the following steps: A pre-trained local trajectory pattern recognition model is loaded using a lightweight neural network inference framework on the edge. A sliding window truncation process is performed on the multi-dimensional feature trajectory sequence within a fixed time window before and after each data point in the transient point set to be verified, so as to generate a standardized time-series feature fragment containing image sharpness peak, lens position pulse count, ambient illuminance change rate and temperature gradient. Using standardized temporal feature fragments as input, multi-scale spatiotemporal features are extracted through the convolutional coding layer in the local trajectory pattern recognition model. The standardized temporal feature fragments are then mapped into high-dimensional latent space feature vectors to obtain deep pattern feature encoding. Based on deep pattern feature encoding, similarity measurement calculation with pre-set typical nonlinear transition pattern prototypes is performed in the classification decision layer of the local trajectory pattern recognition model. The deep pattern feature encoding is compared with the standard pattern prototypes stored in the pattern library using Euclidean distance to generate a probability distribution confidence matrix. Based on the maximum confidence value in the probability distribution confidence matrix and its corresponding category index, a dynamic decay weight mapping strategy is executed to convert the probability distribution confidence matrix into dynamic decay weight labels. By integrating the dynamic decay weight labels with the original transient point identifier information to be verified, an enhanced transient point dataset with pattern matching results is generated.
4. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, S5 specifically includes the following steps: Based on the remaining data points of the unmatched typical nonlinear transition mode, its multidimensional time series feature vector is extracted, and the standard trajectory segment set under the same object distance interval is extracted from the historical working condition database using the distributed storage interface of the Elec-Tech operating system. A historical working condition feature index library containing timestamp alignment information and dominant influencing factor labels is constructed. Based on the standard trajectory fragment set in the historical working condition feature index library, the multidimensional time series feature vector of the remaining data points is processed by the dynamic time warping algorithm. The cumulative distance measurement value between the current test trajectory and each historical standard trajectory in three dimensions—image clarity, lens position pulse count, and environmental parameter change rate—is calculated, and a similarity retrieval result sequence representing the similarity of trajectory morphology is generated. Based on candidate noise points with similarity below a preset threshold in the similarity retrieval result sequence, and combined with the motor drive current waveform and lens mechanical vibration spectrum during the current zoom process, a local residual energy calculation model is constructed. Least square fitting residual analysis is performed to quantify the deviation energy between the actual observed value and the theoretical predicted value, and a residual energy comparison index characterizing the mechanical tolerance or temperature drift noise intensity is generated. Based on the residual energy comparison index and the background noise level benchmark, an adaptive dual-threshold decision logic processing is performed to mark data points whose residual energy is higher than the dynamic noise threshold and whose similarity retrieval result sequence shows low correlation as real outliers, while retaining data points whose residual energy is within the tolerance range as valid observations, and generating a data point classification list containing removal and retention labels. Based on the retained marked data points in the data point classification list, spatiotemporal coordinate reconstruction and confidence-weighted aggregation are performed to remove all abnormal data marked as true outliers and integrate them to form a clean set of focus data points.
5. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, S6 specifically includes the following steps: Based on the dominant influencing factor label carried by each data point in the pure focus data set, cluster analysis is used to generate multiple physical cause homogeneous data subsets. Based on the distribution boundaries of the multiple homogeneous physical cause data subsets in the object distance dimension, an interval segmentation algorithm is executed to generate several independent physical cause local intervals that do not overlap and cover the entire range. For each independent physical cause local interval containing data point distribution characteristics, the radial basis function library is called to perform kernel function matching and filtering to generate a set of exclusive fitting basis functions that fit the nonlinear characteristics of the interval. Based on the complexity index of the dedicated fitting basis function set and the variance of the data point density within the local interval of the independent physical cause, an adaptive gain calculation is performed to generate a dedicated adaptive learning rate scalar for each interval. The dedicated set of fitting basis functions and the dedicated adaptive learning rate scalar are structurally encapsulated and bound to generate local model parameter sets corresponding to each independent local interval of physical cause.
6. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that: S7 specifically includes the following steps: Obtain the newly added pure focus data point set and its corresponding dominant influence factor labels, construct a dedicated fitting basis function set based on the radial basis function kernel method, and calculate the local model parameter set including the initial weight coefficients; The local model parameter set is used to perform residual gradient descent on the newly added clean focal data point set to generate a parameter update vector containing the learning rate adjustment factor, so as to correct the local fitting bias caused by mechanical tolerance or temperature drift and output the optimized local model parameter set. Based on the optimized local model parameter set, the boundary gradient perturbation value is calculated. The parameter perturbation information packet containing the boundary gradient perturbation value is broadcast to adjacent physical cause local interval nodes through the Dianhong distributed soft bus to trigger the linkage response mechanism of the neighborhood model and generate cross-interval collaborative convergence instructions. It receives cross-interval collaborative convergence instructions and performs weighted smoothing fusion processing on the local model parameter groups of adjacent intervals to generate a global continuous piecewise focusing curve function; The semantic feature identifiers of each segment interval are extracted based on the global continuous segmented focusing curve function, and the dominant influencing factor labels are mapped to the curve function segments to generate segmented focusing curves with semantic interpretability.
7. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, The method further includes: S8: Based on the real-time prediction and compensation of the segmented focus curve, the zoom motor group is driven to perform a precise focusing action.
8. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 7, characterized in that, S8 specifically includes the following steps: Obtain the target object distance interval identifier and the corresponding exclusive fitting basis function parameters in the segmented focusing curve. Use the lightweight neural network inference framework of the Elec-Tech OS to load the pre-trained motor dynamic response prediction model. Use the target object distance interval identifier and the ambient temperature parameter as input vectors to perform forward propagation calculation and generate a motor dynamic characteristic descriptor that characterizes the motor mechanical hysteresis and electrical time constant under the current working condition. Based on the motor dynamic characteristic descriptor, a discrete state-space equation containing inertial and damping terms is constructed. The inverse system solution process is performed on the theoretical image plane position sequence output by the piecewise focusing curve to generate a theoretical motor pulse count target trajectory sequence containing the predicted lead amount. The current real-time rotor position pulse count value and speed feedback signal of the zoom motor are collected. The instantaneous position deviation between the theoretical motor pulse count target trajectory sequence and the real-time rotor position pulse count value is calculated. An adaptive proportional-integral-derivative adjustment algorithm is introduced to correct the instantaneous position deviation and generate a motor drive control voltage instruction set optimized by delay compensation. The motor drive control voltage command set is sent to the motor drive controller node via the distributed soft bus of the electric motor, driving the zoom motor group to perform precise focusing action according to the theoretical motor pulse count target trajectory sequence.
9. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, In S1, under the distributed soft bus topology of the Elec-Tech OS, the unique hardware identifiers and network connection status of the camera optical module, image acquisition unit, temperature and humidity sensor node, and motor drive controller are loaded as initial conditions into the clock synchronization management module of the main control chip.
10. The modular camera autofocus curve intelligent fitting method based on the Elec-Tech operating system according to claim 1, characterized in that, In S1, in a distributed sensor network clock environment that is in a steady-state synchronization state, the master control chip is selected as the global control source.