Radio equipment full-automatic calibration system based on machine vision

Abstract

The invention discloses a radio equipment full-automatic calibration system based on machine vision, which relates to the technical field of industrial automation and machine vision, and comprises a collaborative perception module, a consensus diagnosis module, a visual guidance module and a calibration analysis module, registration and consistency verification of multi-modal sensing data are achieved through space-time mapping of a visual coordinate system and a radio frequency signal field, an environment multi-physical-field consensus graph is dynamically constructed by fusing data of multiple devices, physical effect image sequences such as excited near-field thermal radiation, fluorescence and polarization state disturbance are collected, and the consensus graph is obtained. Spatial feature comparison is carried out on an acquired image sequence and an ideal state standard map, the actual deviation value and the accurate compensation value of each adjustable parameter of the radio frequency front end are directly calculated, and after the accurate calibration compensation value is received, a plurality of peripheral health devices are dynamically scheduled to form a coordinated regulation array, and the regulation accuracy is improved. A virtual ideal calibration field is synthesized at a spatial specific position, and an optimal reference signal environment is provided for equipment to be calibrated.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation and machine vision technology, and in particular to a fully automated calibration system for wireless equipment based on machine vision. Background Technology

[0002] In the logistics system of smart factories, the stable operation of automated guided vehicles (AGVs) depends heavily on accurate positioning and reliable wireless communication. However, the increasingly dense industrial Internet devices have led to an exceptionally complex electromagnetic environment in the workshop, posing a severe challenge to the AGV system. Existing technical solutions typically employ an isolated and passive approach when addressing related issues, which has fundamental flaws: The calibration of existing AGVs and their communication modules is an offline, periodic preventative maintenance that requires removing the AGVs from the production line and performing the calibration in a dedicated shielded room using standard instruments. When an AGV loses connection due to wireless interference, the industry currently generally adopts a reactive approach: maintenance personnel go to the site with a handheld spectrum analyzer to scan and manually locate the approximate direction of the interference. This results in low efficiency in interference troubleshooting and cannot completely resolve persistent interference problems caused by equipment failure, misconfiguration, or unauthorized access. In existing industrial vision applications, machine vision is mostly used for code reading, size measurement, or simple defect detection. Its perception dimension is limited to optical images and does not realize visual features such as equipment model, antenna shape, physical location, and radio frequency features such as signal fingerprint, modulation characteristics, and propagation path. In the deep integration of the data layer and semantic layer, vision, radio frequency, and navigation information are still isolated data silos. They have failed to form a unified cognitive model to understand the global state of the physical-electromagnetic coupling space. They have limitations and lack an intelligent management that can uniformly understand physical space coordinates, electromagnetic field strength distribution, and equipment visual entities, which reduces the efficiency of intelligent management of factory equipment. To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0003] The purpose of this invention is to construct a unified environmental perception field through the spatiotemporal alignment and fusion of visual, radio frequency, and pose data, providing a reliable data foundation for calibration; to generate a consensus truth about the environment using observation data from multiple devices; to quickly locate hidden faults in single devices through comparison, achieving intelligent diagnosis; and to infer internal parameter deviations by analyzing near-field thermal, optical, and polarization multi-physical effect images excited by the radio frequency front end, without physical contact or disassembly. Furthermore, it involves scheduling surrounding healthy devices to form an array, synthesizing a virtual ideal field in the air at the target location, and achieving contactless, dynamic calibration of faulty devices. From data acquisition, anomaly diagnosis, parameter inversion to field-coordinated calibration, everything is executed automatically, significantly improving calibration efficiency and system availability.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a fully automatic calibration system for wireless equipment based on machine vision, comprising a collaborative perception module, a consensus diagnosis module, a visual guidance module, a calibration analysis module, and a calibration execution module; The collaborative sensing module is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field in the operating space, collect and register the visual image sequence, radio frequency spectrum scanning data and pose information of the device, generate sensing data, calculate the spatiotemporal consistency index, and trigger the system reference recalibration when the index is lower than the threshold. The consensus diagnostic module acquires and fuses the perception data of multiple devices on the same time axis, dynamically generates an environmental multiphysics consensus map, and diagnoses the implicit parameter drift of a specific radio frequency link based on the distance between the real-time perception feature vector acquired by a single device and the environmental multiphysics consensus map, and performs cross-validation with the vision of adjacent devices, generating a structured device health assessment report, which is then converted into targeted calibration instructions. The vision guidance module is used to control the target device to emit specific test signals according to the target calibration instructions, and to acquire the image sequence of physical effects excited by the near-field electromagnetic effect of the device's radio frequency front end, including thermal radiation distribution, characteristic fluorescence and polarization state perturbation. The calibration analysis module is used to compare the spatial features of the acquired physical effect image sequence with the pre-stored standard effect spectrum of ideal transmission state, and calculate the actual deviation and calibration compensation value of each adjustable parameter of the RF front end of the device to be calibrated through a pre-trained visual effect-RF parameter inverse mapping model. The calibration execution module receives calibration compensation values, schedules multiple health devices to form a collaborative control array, synthesizes a virtual ideal calibration field at the location of the device to be calibrated, and obtains the optimal solution for performance indicators.

[0005] Furthermore, it is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field within the operating space, and to acquire and register the visual image sequence, radio frequency spectrum scan data, and pose information of the device. The specific acquisition process is as follows: Based on visual sensing devices, radio frequency scanning devices, and high-precision position sensors, visual image sequences, radio frequency spectrum data, and pose information are obtained. The acquired visual image sequence is processed, and pre-deployed coded markers are used to construct a visual coordinate system in space. The pose trajectory in the visual coordinate system is mapped to the visual coordinate system based on the fixed spatial transformation relationship between the radio frequency scanning device and the visual sensor. The radio frequency spectrum data corresponding to each timestamp is spatially bound to the precise position of the device in the visual coordinate system to construct a continuous radio frequency signal field model. The fused perception data is output, including the visual coordinate system model, the device pose trajectory, and the radio frequency feature vector corresponding to each spatial coordinate point, forming a queryable spatiotemporal correlation database.

[0006] Furthermore, a spatiotemporal consistency index is calculated. When the index falls below a threshold, a system baseline recalibration is triggered. The specific calculation process is as follows: S100. Based on the generated perception data, periodically calculate the spatiotemporal consistency index to obtain the position vector measured by the real-time vision device and the coordinate vector measured by the position sensor. Based on the coordinate vector and the position vector, obtain the original measurement value, calculate the consistency between vision and position, and perform residual calculation based on the position trajectory of the two to obtain the first index value. S101. Based on the established radio frequency signal field model, and based on the actual coordinate position and the coordinate position obtained from the radio frequency signal field model, perform a consistency analysis between the visual and radio frequency spaces to obtain the second index value. S102. The first indicator value and the second indicator value are merged to generate a comprehensive spatiotemporal consistency score. A standardized consistency judgment value is set, and a judgment is made at the end of each calculation cycle. When the comprehensive score is lower than the threshold, it is determined that there is a significant spatiotemporal mismatch in the system, and the recalibration process is triggered.

[0007] Furthermore, it is used to fuse sensing data from multiple devices on the same timeline to dynamically generate a multiphysics consensus map of the environment. The specific generation process is as follows: S200: Acquire perception data from multiple devices, and transform the local visual coordinates of each device into the visual coordinate system through a pre-calibrated relative pose transformation matrix between devices, forming a unified spatiotemporal grid. S201. Based on the spatiotemporal grid, acquire observation data of devices adjacent to each spatial location, classify them according to the physical field, including radio frequency field, visual field and sensor coordinate field, acquire the observation values ​​provided by the radio frequency field and visual field of the device in each grid and the observation uncertainty calibrated by the device, and perform pose estimation uncertainty correlation on the observation values ​​of the sensor field. S202. In the established unified spatiotemporal grid, for each physical field in each spatial location, the observations from different devices are fused with their overall uncertainty. The inverse variance weighted algorithm is used to calculate the optimal consensus estimate and its confidence level of the physical field at this location. Based on the spatiotemporal grid space, the final environmental multiphysics consensus map is generated for each spatial location.

[0008] Furthermore, the system diagnoses the latent parameter drift of specific RF links, generates a structured device health assessment report, and translates it into targeted calibration instructions. The specific diagnostic process is as follows: S300. Extract real-time sensing feature vectors from the real-time sensing data obtained from a single device, compare the consistency of the single device's real-time sensing feature vectors with the environmental multiphysics consensus map, and perform initial anomaly screening. If the consistency is lower than the standard judgment value, it is preliminarily determined that the device has a sensing anomaly, triggering the deep diagnosis process. S301. Schedule nearby devices with normal visual and radio frequency sensing functions, perform a triple comparison of visual observations of the same spatial area where the devices are located and their own radio frequency data, and generate analysis results. S302. Based on the analysis results, automatically generate a structured equipment health assessment report and automatically convert it into executable targeted calibration instructions.

[0009] Furthermore, it is used to control the target device to emit specific test signals according to the targeted calibration instructions, and to obtain a sequence of physical effect images of the device's radio frequency front end excited by near-field electromagnetic effects. The specific acquisition process is as follows: The system obtains a target calibration command to enable the target device to perform a diagnostic process. It also configures the radio frequency front-end of the target device through the device's control interface and transmits a specific test signal. Simultaneously with the transmission of the characteristic test signal, it activates all imaging devices to acquire data through a synchronous pulse signal, thereby obtaining thermal radiation distribution image sequences, characteristic fluorescence image sequences, and polarization state perturbation image sequences. The system then extracts dynamic features that are strongly correlated with the working state of the radio frequency front-end. Based on the thermal radiation distribution image sequence, the maximum temperature rise and thermal diffusion rate of the key area are calculated to evaluate power consumption and heat dissipation characteristics. The average fluorescence intensity is calculated from the characteristic fluorescence image sequence. Based on the polarization state perturbation image sequence, the amplitude and principal direction of the polarization state perturbation are extracted to characterize the vector properties of the electromagnetic field. The dynamic features are then fused to generate a structured physical effect feature vector.

[0010] Furthermore, the pre-trained visual effects-RF parameter inverse mapping model is constructed using the following steps: Select the tunable frequency parameter values ​​of the same batch of target RF front-end devices when transmitting specific test signals, construct them as parameter vectors, collect the physical effect image sequence of the target RF front-end devices in the transmission state, form the original image dataset, perform data annotation and pairing, and generate sample datasets; Feature vectors are extracted and preprocessed from the samples in the sample dataset to obtain structured feature vectors, which are used as input features. Based on a deep neural network with a fully connected layer architecture, a loss function for the visual effect-RF parameter inverse mapping model is defined. The model is trained by minimizing the loss function through the backpropagation algorithm to obtain a pre-trained visual effect-RF parameter inverse mapping model. The output is an estimated vector of RF parameter deviation and its confidence vector.

[0011] Furthermore, using a pre-trained visual effect-RF parameter inverse mapping model, the actual deviation and calibration compensation value of each adjustable parameter of the RF front-end of the device to be calibrated are calculated. The specific calculation process is as follows: The physical effect image sequence and the pre-stored standard effect spectrum of ideal emission state are compared to calculate the difference features and combine the features to form a difference feature vector; The obtained differential feature vector is input into the pre-trained visual effect-RF parameter inverse mapping model, and the output is the RF parameter deviation estimation vector and its confidence vector. The model is validated based on material constraints. If the estimated value violates the strong constraints, the constraint-based optimization correction is triggered to obtain the corrected deviation. The parameter deviation is then converted into a specific, executable calibration compensation value.

[0012] Furthermore, to receive calibration compensation values, multiple health devices are scheduled to form a collaborative control array to synthesize a virtual ideal calibration field at the location of the device to be calibrated, thereby obtaining the optimal solution for performance indicators. The specific process is as follows: The received calibration compensation value is analyzed, and an optimization problem is constructed with the goal of synthesizing an electromagnetic field that conforms to the ideal characteristics at the location of the device to be calibrated. The optimal transmission parameter configuration of each device is obtained by taking the transmission amplitude and phase of each healthy device as variables, the matching degree between the synthesized field and the ideal field as the optimization objective, and the device performance and safety specifications as constraints. The transmission parameter configuration that achieves the highest performance score and the corresponding measured performance index are output as the optimal solution for this collaborative calibration, and the array is controlled to transmit according to this configuration, thereby actually generating a virtual ideal calibration field at the target location.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This machine vision-based fully automated calibration system for wireless equipment utilizes the spatiotemporal mapping between the visual coordinate system and the radio frequency signal field to achieve registration and consistency verification of multimodal sensing data. When a reference deviation is detected, recalibration is automatically triggered. By fusing data from multiple devices to dynamically construct a multiphysics consensus map of the environment, and based on the deviation between the sensing characteristics of a single device and the consensus map, combined with visual cross-verification from neighboring devices, it can accurately diagnose the implicit parameter drift of a single radio frequency link. It transforms ambiguous device anomalies into structured health reports and targeted calibration commands, controls the target device to transmit test signals, and collects the excited near-field thermal data. Image sequences of physical effects such as radiation, fluorescence, and polarization state perturbation are transformed into visually interpretable high-dimensional image features. Through a pre-trained visual effect-RF parameter inverse mapping model, the acquired image sequences are spatially compared with the ideal state standard spectrum. The actual deviation and accurate compensation value of each adjustable parameter of the RF front end are directly calculated, realizing calibration analysis from indirect signal inference to direct physical effect inversion. After receiving the accurate calibration compensation value, multiple peripheral health devices are dynamically scheduled to form a collaborative control array. By synthesizing a virtual ideal calibration field at a specific location in space, the optimal reference signal environment is provided for the device to be calibrated. Attached Figure Description

[0014] Figure 1 A schematic diagram of the overall system steps of the present invention is shown. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: like Figure 1 As shown, the fully automated calibration system for wireless equipment based on machine vision includes a collaborative perception module, a consensus diagnosis module, a vision guidance module, a calibration analysis module, and a calibration execution module. The collaborative sensing module is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field in the operating space, collect and register the visual image sequence, radio frequency spectrum scanning data and pose information of the device, generate sensing data, calculate the spatiotemporal consistency index, and trigger the system reference recalibration when the index is lower than the threshold. The consensus diagnostic module acquires and fuses the perception data of multiple devices on the same time axis, dynamically generates an environmental multiphysics consensus map, and diagnoses the implicit parameter drift of a specific radio frequency link based on the distance between the real-time perception feature vector acquired by a single device and the environmental multiphysics consensus map, and performs cross-validation with the vision of adjacent devices, generating a structured device health assessment report, which is then converted into targeted calibration instructions. The vision guidance module is used to control the target device to emit specific test signals according to the target calibration instructions, and to acquire the image sequence of physical effects excited by the near-field electromagnetic effect of the device's radio frequency front end, including thermal radiation distribution, characteristic fluorescence and polarization state perturbation. The calibration analysis module is used to compare the spatial features of the acquired physical effect image sequence with the pre-stored standard effect spectrum of ideal transmission state, and calculate the actual deviation and calibration compensation value of each adjustable parameter of the RF front end of the device to be calibrated through a pre-trained visual effect-RF parameter inverse mapping model. The calibration execution module receives calibration compensation values, schedules multiple health devices to form a collaborative control array, synthesizes a virtual ideal calibration field at the location of the device to be calibrated, and obtains the optimal solution for performance indicators.

[0017] Specifically, it is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field within the operating space, and to acquire and register the visual image sequence, radio frequency spectrum scan data, and pose information of the device. The specific acquisition process is as follows: Based on visual sensing devices, radio frequency scanning devices, and high-precision position sensors, visual image sequences, radio frequency spectrum data, and pose information are obtained. The acquired visual image sequence is processed, and pre-deployed coded markers are used to construct a visual coordinate system in space. The pose trajectory in the visual coordinate system is mapped to the visual coordinate system based on the fixed spatial transformation relationship between the radio frequency scanning device and the visual sensor. The radio frequency spectrum data corresponding to each timestamp is spatially bound to the precise position of the device in the visual coordinate system to construct a continuous radio frequency signal field model. The fused perception data is output, including the visual coordinate system model, the device pose trajectory, and the radio frequency feature vector corresponding to each spatial coordinate point, forming a queryable spatiotemporal correlation database.

[0018] In this solution, the collaborative perception module first achieves synchronous data acquisition of the target device through a multi-view vision system, a broadband radio frequency scanning array, and a high-precision positioning network deployed on the site. The system processes the acquired visual image sequence, identifies coded markers pre-placed in the environment, and constructs a three-dimensional visual coordinate system V with the site center as the origin through multi-view geometric calculation. Simultaneously, the high-precision positioning system (such as UWB) provides the device's pose in the global coordinate system in real time. Through a pre-calibrated coordinate transformation matrix Transform it to a 3D vision coordinate system to obtain the precise pose trajectory of the device in V. ; The key lies in the spatial mapping of radio frequency (RF) data. RF scanning equipment and vision sensors are connected via a rigid structure or undergo precise calibration, and their fixed spatial transformation matrix is... For each RF sampling time t, the system determines the device's current pose. With attitude rotation matrix The beam pointing vector measured by the radio frequency scanner Mapped to the three-dimensional spatial position in the visual coordinate system : ; This step assigns precise spatial coordinates to the originally independent radio frequency sampling points; Subsequently, the system performs spatiotemporal binding and field modeling, associating the radio frequency spectrum data corresponding to each timestamp t with its corresponding spatial coordinates. Binding is performed by constructing a continuous radio frequency signal field model through three-dimensional interpolation of time-series and spatially sampled point cloud data. Finally, the module outputs the fused perception data packet, whose data structure includes: parameters of the visual coordinate system V, device pose trajectory sequence, and radio frequency feature vector field based on spatial grid index. This data packet constitutes a database that can be queried in a spatiotemporal manner.

[0019] Specifically, a spatiotemporal consistency index is calculated. When the index falls below a threshold, a system baseline recalibration is triggered. The specific calculation process is as follows: S100. Based on the generated perception data, periodically calculate the spatiotemporal consistency index to obtain the position vector measured by the real-time vision device and the coordinate vector measured by the position sensor. Based on the coordinate vector and the position vector, obtain the original measurement value, calculate the consistency between vision and position, and perform residual calculation based on the position trajectory of the two to obtain the first index value. S101. Based on the established radio frequency signal field model, and based on the actual coordinate position and the coordinate position obtained from the radio frequency signal field model, perform a consistency analysis between the visual and radio frequency spaces to obtain the second index value. S102. The first indicator value and the second indicator value are merged to generate a comprehensive spatiotemporal consistency score. A standardized consistency judgment value is set, and a judgment is made at the end of each calculation cycle. When the comprehensive score is lower than the threshold, it is determined that there is a significant spatiotemporal mismatch in the system, and the recalibration process is triggered.

[0020] In this implementation scheme, to ensure the long-term reliable fusion of multi-source heterogeneous sensor data, the system periodically performs self-checks, determines its own status by calculating spatiotemporal consistency indicators, extracts data within a time window, and obtains the device location sequence calculated from visual markers. and the coordinate sequence directly output by the positioning sensor Calculate the residual between the two trajectories to obtain the first index value.

[0021] ; For the actual visual coordinates of the device at time t The predicted radio frequency feature vector is obtained by querying the radio frequency signal field model. Simultaneously, the actual radio frequency feature vector scanned at that moment is obtained. Calculate the cosine similarity between the two, and take the average within the window as the second index value. : ; The two indicators are combined to generate a comprehensive spatiotemporal consistency score. : ; like , If the consistency judgment value is set based on the average of the consistency judgment values ​​of historical periods, it is determined that there is a significant spatiotemporal mismatch in the multi-sensor data. The system automatically triggers the benchmark recalibration process and re-executes the sensor extrinsic parameter calibration and coordinate system one.

[0022] Specifically, it is used to fuse sensing data from multiple devices on the same timeline to dynamically generate a multiphysics consensus map of the environment. The specific generation process is as follows: S200: Acquire perception data from multiple devices, and transform the local visual coordinates of each device into the visual coordinate system through a pre-calibrated relative pose transformation matrix between devices, forming a unified spatiotemporal grid. S201. Based on the spatiotemporal grid, acquire observation data of devices adjacent to each spatial location, classify them according to the physical field, including radio frequency field, visual field and sensor coordinate field, acquire the observation values ​​provided by the radio frequency field and visual field of the device in each grid and the observation uncertainty calibrated by the device, and perform pose estimation uncertainty correlation on the observation values ​​of the sensor field. S202. In the established unified spatiotemporal grid, for each physical field in each spatial location, the observations from different devices are fused with their overall uncertainty. The inverse variance weighted algorithm is used to calculate the optimal consensus estimate and its confidence level of the physical field at this location. Based on the spatiotemporal grid space, the final environmental multiphysics consensus map is generated for each spatial location.

[0023] In this scheme, sensing data from M devices is acquired, and a pre-calibrated relative pose transformation matrix between the devices is used. The local observation data of all devices are converted into a unified reference visual coordinate system. Then, the entire space is divided into a regular three-dimensional grid to form a unified spatiotemporal analysis framework.

[0024] In each grid cell Internally, observation data from different devices are categorized by physical field type: radio frequency field observations. Visual field observations Sensor coordinate field observations Meanwhile, the uncertainty information of each observation is correlated: the uncertainty of radio frequency and visual observations stems from the calibration error and noise level of the equipment itself; the uncertainty of sensor coordinate field observations is related to the accuracy of the positioning system and the pose estimation algorithm.

[0025] Inverse variance weighted fusion generates consensus graph: for each grid For each physical field within the grid cell, an inverse variance weighted algorithm is used for optimal fusion. Taking the radio frequency field as an example, its optimal consensus estimate within that grid cell is... and the uncertainty variance after fusion The calculation is as follows: ; ; in, For the k-th three-dimensional spatial grid cell, For the j-th device in the grid Radio frequency field observations within, For the j-th device in the grid Visual field observations within, Let j be the pose estimate of the j-th device. By traversing all grid cells and physical fields, a multiphysics consensus graph of the environment is finally generated, which contains the optimal estimate and its confidence level. This graph represents the most reliable physical facts in the current environment.

[0026] Specifically, the process involves diagnosing latent parameter drift in a specific RF link, generating a structured device health assessment report, and then converting it into targeted calibration instructions. The specific diagnostic process is as follows: S300. Extract real-time sensing feature vectors from the real-time sensing data obtained from a single device, compare the consistency of the single device's real-time sensing feature vectors with the environmental multiphysics consensus map, and perform initial anomaly screening. If the consistency is lower than the standard judgment value, it is preliminarily determined that the device has a sensing anomaly, triggering the deep diagnosis process. S301. Schedule nearby devices with normal visual and radio frequency sensing functions, perform a triple comparison of visual observations of the same spatial area where the devices are located and their own radio frequency data, and generate analysis results. S302. Based on the analysis results, automatically generate a structured equipment health assessment report and automatically convert it into executable targeted calibration instructions.

[0027] In this solution, feature vectors are extracted from the real-time sensing data of the device to be diagnosed. ,calculate Consensus Map The Mahalanobis distance at the corresponding spatial location takes into account both the magnitude and uncertainty of the difference. Exceeding the preset threshold Therefore, the equipment was initially determined. Detecting an anomaly triggers in-depth diagnosis. Cross-validation and root cause analysis, system scheduling Several health devices located nearby These health devices are Independent visual observation and radio frequency scanning are performed on the same spatial region, and the system performs triple comparison: Image comparison: Analyze the SIFT feature matching rate of images of the same scene captured by multiple devices; RF data comparison: calculation With each Dynamic time-warped distance of the spectral morphology measured at the same location; Coordinate position comparison: verification Self-report location and reason The difference in position estimated by triangulation; By comprehensively analyzing inconsistencies, the root cause of the anomaly can be located to a specific radio frequency link; Report and instruction generation: Based on the location results, a structured equipment health assessment report is automatically generated, including: abnormal equipment ID, abnormal link identifier, suspected parameter drift type, drift severity level and confidence level. The system automatically converts this report into machine-readable targeted calibration instructions.

[0028] Specifically, it is used to control the target device to emit a specific test signal according to the target calibration command, and to obtain a sequence of physical effect images of the device's radio frequency front end excited by near-field electromagnetic effects. The specific acquisition process is as follows: The system obtains a target calibration command to enable the target device to perform a diagnostic process. It also configures the radio frequency front-end of the target device through the device's control interface and transmits a specific test signal. Simultaneously with the transmission of the characteristic test signal, it activates all imaging devices to acquire data through a synchronous pulse signal, thereby obtaining thermal radiation distribution image sequences, characteristic fluorescence image sequences, and polarization state perturbation image sequences. The system then extracts dynamic features that are strongly correlated with the working state of the radio frequency front-end. Based on the thermal radiation distribution image sequence, the maximum temperature rise and thermal diffusion rate of the key area are calculated to evaluate power consumption and heat dissipation characteristics. The average fluorescence intensity is calculated from the characteristic fluorescence image sequence. Based on the polarization state perturbation image sequence, the amplitude and principal direction of the polarization state perturbation are extracted to characterize the vector properties of the electromagnetic field. The dynamic features are then fused to generate a structured physical effect feature vector. In this scheme, the maximum value of the temperature distribution matrix is ​​obtained from the thermal radiation distribution image sequence. The temperature rise is calculated for each time t, and the maximum value in the temperature rise matrix is ​​identified and recorded as the local maximum temperature rise at time t. The maximum value among all local maximum temperature rises is taken as the final result. By fitting the temperature rise change curve, the time constant for heat accumulation and dissipation, and the thermal diffusion rate are obtained. In the area covered by the electromagnetically sensitive material, the fluorescence region is extracted and the preset static background light intensity is subtracted. The average intensity within the region is taken as the arithmetic mean over the entire excitation cycle as the final average fluorescence intensity. Based on all average intensities, the instantaneous gradient at each internal time point t is calculated. The maximum value of the gradient sequence, the average gradient of the rising edge, and the time required to reach steady state are extracted as dynamic response features. Based on the polarization state perturbation image sequence, four Stokes parametric images are directly output by the polarization camera. Four Stokes parameters ; Linear polarization degree DoLP: ; Polarization angle θ: ; Selecting a reference frame without RF excitation Calculate its average degree of polarization ; The perturbation amplitude image is defined as the change in polarization degree relative to the reference frame at each time step: ; Based on the region with the largest disturbance amplitude, calculate the peak value of all pixels within that region at the time of the disturbance. polarization angle The dominant statistical direction, as the main direction angle ; Among them, the Stokes parameter is used to completely describe the polarization state of a beam of light. The values ​​represent the total light intensity, the difference between horizontal and vertical linear polarization components, the difference between +45° and -45° linear polarization components, and the difference between right-handed and left-handed circular polarization components. i usually represents the row index of the image, and j usually represents the column index of the image.

[0029] Specifically, the pre-trained visual effect-RF parameter inverse mapping model is constructed using the following steps: Select the tunable frequency parameter values ​​of the same batch of target RF front-end devices when transmitting specific test signals, construct them as parameter vectors, collect the physical effect image sequence of the target RF front-end devices in the transmission state, form the original image dataset, perform data annotation and pairing, and generate sample datasets; Feature vectors are extracted and preprocessed from the samples in the sample dataset to obtain structured feature vectors, which are used as input features. Based on a deep neural network with a fully connected layer architecture, a loss function for the visual effect-RF parameter inverse mapping model is defined. The model is trained by minimizing the loss function through the backpropagation algorithm to obtain a pre-trained visual effect-RF parameter inverse mapping model. The output is an estimated vector of RF parameter deviation and its confidence vector.

[0030] Specifically, using a pre-trained visual effect-RF parameter inverse mapping model, the actual deviation and calibration compensation value of each adjustable parameter of the RF front-end of the device to be calibrated are calculated. The specific calculation process is as follows: The physical effect image sequence and the pre-stored standard effect spectrum of ideal emission state are compared to calculate the difference features and combine the features to form a difference feature vector; The obtained differential feature vector is input into the pre-trained visual effect-RF parameter inverse mapping model, and the output is the RF parameter deviation estimation vector and its confidence vector. The model is verified based on material constraints. If the estimated value violates the strong constraints, the constraint-based optimization correction is triggered to obtain the corrected deviation. The parameter deviation is then converted into a specific, executable calibration compensation value. In this scheme, during the online calibration phase, feature vectors are first extracted from the sequence of physical effect images acquired by the device to be calibrated. Simultaneously, the eigenvectors corresponding to the standard effect spectrum of this equipment model under ideal launch conditions are retrieved from the database. Calculate the difference feature vectors between the two. ; Will In the pre-trained inverse mapping model, after forward propagation, two vectors are output: the radio frequency parameter deviation estimation vector Δθ^. The confidence vector represents the reliability of each bias estimate; Subsequently, verification and correction based on physical constraints are performed, traversing the hardware constraints of the system's built-in RF front-end. For each estimated value in the equation, if it causes the adjusted parameters to violate the strong constraints, then a constraint optimization correction is triggered. The correction problem can be formalized as:

[0031] Where W is a weight matrix with confidence levels as diagonal elements, and lb and ub are the lower and upper bounds of the parameters, respectively. Solving for W yields the physically feasible corrected bias. ; Finally, Convert the values ​​into specific calibration compensation values ​​that the device can execute, such as converting the frequency deviation into the frequency control value written to the phase-locked loop; or converting the gain deviation into the control voltage or digital codeword of the variable gain amplifier. In this scheme, the acquired thermal radiation, characteristic fluorescence, and polarization state perturbation image sequences are spatially registered with the pre-stored ideal emission state standard effect spectrum of the corresponding mode at the subpixel level, for each.

[0032] Specifically, it is used to receive calibration compensation values, schedule multiple health devices to form a collaborative control array, synthesize a virtual ideal calibration field at the location of the device to be calibrated, and obtain the optimal solution for performance indicators. The specific process is as follows: The received calibration compensation value is analyzed, and an optimization problem is constructed with the goal of synthesizing an electromagnetic field that conforms to the ideal characteristics at the location of the device to be calibrated. The optimal transmission parameter configuration of each device is obtained by taking the transmission amplitude and phase of each healthy device as variables, the matching degree between the synthesized field and the ideal field as the optimization objective, and the device performance and safety specifications as constraints. The transmission parameter configuration that achieves the highest performance score and the corresponding measured performance index are output as the optimal solution for this collaborative calibration, and the array is controlled to transmit according to this configuration, thereby actually generating a virtual ideal calibration field at the target location.

[0033] In this scheme, the calibration execution module receives the final compensation value from the calibration analysis module. Instead of directly adjusting the device to be calibrated, it uses multiple surrounding healthy devices to construct a collaborative control array to synthesize an electromagnetic field environment that meets ideal characteristics at the location of the device to be calibrated, thereby performing air calibration. First, analyze the calibration compensation value and calculate the current position of the device to be calibrated. The ideal calibration field required to achieve the desired performance indicators. Based on the characteristics of the network, K geographically suitable and functionally healthy devices are then scheduled from the network to form a coordinated control array. Using the amplitude and initial phase of the transmitted signals from each healthy device as optimization variables, a field synthesis optimization problem is constructed: Next, K geographically appropriate and functionally healthy devices are selected from the network to form a coordinated control array, with the transmission signal amplitude of each healthy device being controlled. and initial phase To optimize the variables, a field synthesis optimization problem is constructed:

[0034]

[0035]

[0036]

[0037] in, It is an array synthesis field. It is the power regularization coefficient used to control the total transmit power, and SAR is the specific absorption rate, which must meet safety specifications. After the array is configured to operate, a virtual ideal calibration field is actually generated. The device to be calibrated operates in this field, and its key performance indicators, such as the received signal-to-noise ratio (SNR) and error vector magnitude (EVM), are monitored in real time. The system iteratively fine-tunes the configuration until the performance indicators measured by the device to be calibrated reach the preset optimal solution range. Finally, the array configuration at this point and the corresponding measured optimal performance indicators are output as the result of this collaborative calibration, completing the calibration.

[0038] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0039] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. In the two embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways; for example, the device embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or modules may be electrical, mechanical or other forms. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fully automated calibration system for wireless equipment based on machine vision, characterized in that, It includes a collaborative perception module, a consensus diagnosis module, a visual guidance module, a calibration analysis module, and a calibration execution module; The collaborative sensing module is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field in the operating space, collect and register the visual image sequence, radio frequency spectrum scanning data and pose information of the device, generate sensing data, calculate the spatiotemporal consistency index, and trigger the system reference recalibration when the index is lower than the threshold. The consensus diagnostic module acquires and fuses the perception data of multiple devices on the same time axis, dynamically generates an environmental multiphysics consensus map, and diagnoses the implicit parameter drift of a specific radio frequency link based on the distance between the real-time perception feature vector acquired by a single device and the environmental multiphysics consensus map, and performs cross-validation with the vision of adjacent devices, generating a structured device health assessment report, which is then converted into targeted calibration instructions. The vision guidance module is used to control the target device to emit specific test signals according to the target calibration instructions, and to acquire the image sequence of physical effects excited by the near-field electromagnetic effect of the target device's radio frequency front end, including thermal radiation distribution, characteristic fluorescence and polarization state perturbation; The calibration analysis module is used to compare the spatial features of the acquired physical effect image sequence with the pre-stored standard effect spectrum of ideal transmission state, and calculate the actual deviation and calibration compensation value of each adjustable parameter of the RF front end of the device to be calibrated through a pre-trained visual effect-RF parameter inverse mapping model. The calibration execution module is used to receive calibration compensation values, schedule multiple health devices to form a collaborative control array, synthesize a virtual ideal calibration field at the location of the device to be calibrated, and obtain the optimal solution for performance indicators.

2. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, This is used to establish the mapping relationship between the visual coordinate system and the radio frequency signal field within the operating space, and to acquire and register the device's visual image sequences, radio frequency spectrum scan data, and pose information. The specific acquisition process is as follows: Based on visual sensing devices, radio frequency scanning devices, and high-precision position sensors, visual image sequences, radio frequency spectrum data, and pose information are obtained. The acquired visual image sequence is processed, and pre-deployed coded markers are used to construct a visual coordinate system in space. The pose trajectory in the visual coordinate system is mapped to the visual coordinate system based on the fixed spatial transformation relationship between the radio frequency scanning device and the visual sensor. The radio frequency spectrum data corresponding to each timestamp is spatially bound to the precise position of the device in the visual coordinate system to construct a continuous radio frequency signal field model. The fused perception data is output, including the visual coordinate system model, the device pose trajectory, and the radio frequency feature vector corresponding to each spatial coordinate point, forming a queryable spatiotemporal correlation database.

3. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, The spatiotemporal consistency index is calculated, and when the index falls below a threshold, the system baseline recalibration is triggered. The specific calculation process is as follows: S100. Based on the generated perception data, the spatiotemporal consistency index is calculated periodically to obtain the position vector measured by the real-time vision device and the coordinate vector measured by the position sensor. Based on the coordinate vector and the position vector, the original measurement value is obtained, and the consistency between vision and position is calculated. The residual is calculated based on the position trajectory of the two to obtain the first index value. S101. Based on the established radio frequency signal field model, and based on the actual coordinate position and the coordinate position obtained from the radio frequency signal field model, perform a consistency analysis between the visual and radio frequency spaces to obtain the second index value. S102. The first indicator value and the second indicator value are merged to generate a comprehensive spatiotemporal consistency score. A standardized consistency judgment value is set, and a judgment is made at the end of each calculation cycle. When the comprehensive score is lower than the threshold, it is determined that there is a significant spatiotemporal mismatch in the system, and the recalibration process is triggered.

4. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, This is used to fuse sensing data from multiple devices on the same timeline and dynamically generate a multiphysics consensus map of the environment. The specific generation process is as follows: S200: Acquire perception data from multiple devices, and transform the local visual coordinates of each device into the visual coordinate system through a pre-calibrated relative pose transformation matrix between devices, forming a unified spatiotemporal grid. S201. Based on the spatiotemporal grid, acquire observation data of devices adjacent to each spatial location, classify them according to the physical field, including radio frequency field, visual field and sensor coordinate field, acquire the observation values ​​provided by the radio frequency field and visual field of the device in each grid and the observation uncertainty calibrated by the device, and perform pose estimation uncertainty correlation on the observation values ​​of the sensor field. S202. In the established unified spatiotemporal grid, for each physical field in each spatial location, the observations from different devices are fused with their overall uncertainty. The inverse variance weighted algorithm is used to calculate the optimal consensus estimate and its confidence level of the physical field at this location. Based on the spatiotemporal grid space, the final environmental multiphysics consensus map is generated for each spatial location.

5. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, The system diagnoses latent parameter drift in specific RF links, generates structured device health assessment reports, and translates these into targeted calibration instructions. The specific diagnostic process is as follows: S300. Extract real-time sensing feature vectors from the real-time sensing data obtained from a single device, compare the consistency of the single device's real-time sensing feature vectors with the environmental multiphysics consensus map, and perform initial anomaly screening. If the consistency is lower than the standard judgment value, it is preliminarily determined that the device has a sensing anomaly, triggering the deep diagnosis process. S301. Schedule nearby devices with normal visual and radio frequency sensing functions, perform a triple comparison of visual observations of the same spatial area where the devices are located and their own radio frequency data, and generate analysis results. S302. Based on the analysis results, automatically generate a structured equipment health assessment report and automatically convert it into executable targeted calibration instructions.

6. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, This is used to control the target device to emit specific test signals according to the targeted calibration instructions, and to obtain a sequence of physical effect images of the device's radio frequency front end excited by near-field electromagnetic effects. The specific acquisition process is as follows: The system obtains a target calibration command to enable the target device to perform a diagnostic process. It also configures the radio frequency front-end of the target device through the device's control interface and transmits a specific test signal. Simultaneously with the transmission of the characteristic test signal, it activates all imaging devices to acquire data through a synchronous pulse signal, thereby obtaining thermal radiation distribution image sequences, characteristic fluorescence image sequences, and polarization state perturbation image sequences. The system then extracts dynamic features that are strongly correlated with the working state of the radio frequency front-end. Based on the thermal radiation distribution image sequence, the maximum temperature rise and thermal diffusion rate of the key area are calculated to evaluate power consumption and heat dissipation characteristics. The average fluorescence intensity is calculated from the characteristic fluorescence image sequence. Based on the polarization state perturbation image sequence, the amplitude and principal direction of the polarization state perturbation are extracted to characterize the vector properties of the electromagnetic field. The dynamic features are then fused to generate a structured physical effect feature vector.

7. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, The pre-trained visual effects-RF parameter inverse mapping model is constructed using the following steps: Select the tunable frequency parameter values ​​of the same batch of target RF front-end devices when transmitting specific test signals, construct them as parameter vectors, collect the physical effect image sequence of the target RF front-end devices in the transmission state, form the original image dataset, perform data annotation and pairing, and generate sample datasets; Feature vectors are extracted and preprocessed from the samples in the sample dataset to obtain structured feature vectors, which are used as input features. Based on a deep neural network with a fully connected layer architecture, a loss function for the visual effect-RF parameter inverse mapping model is defined. The model is trained by minimizing the loss function through the backpropagation algorithm to obtain a pre-trained visual effect-RF parameter inverse mapping model. The output is an estimated vector of RF parameter deviation and its confidence vector.

8. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, Using a pre-trained visual effect-RF parameter inverse mapping model, the actual deviation and calibration compensation value of each adjustable parameter of the RF front-end of the device to be calibrated are calculated. The specific calculation process is as follows: The physical effect image sequence and the pre-stored standard effect spectrum of ideal emission state are compared to calculate the difference features and combine the features to form a difference feature vector; The obtained differential feature vector is input into the pre-trained visual effect-RF parameter inverse mapping model, and the output is the RF parameter deviation estimation vector and its confidence vector. The model is validated based on material constraints. If the estimated value violates the strong constraints, the constraint optimization-based correction is triggered to obtain the corrected deviation. The parameter deviation is then converted into a specific, executable calibration compensation value.

9. The fully automated calibration system for wireless equipment based on machine vision according to claim 1, characterized in that, This is used to receive calibration compensation values, schedule multiple health devices to form a collaborative control array, synthesize a virtual ideal calibration field at the location of the device to be calibrated, and obtain the optimal solution for performance indicators. The specific process is as follows: The received calibration compensation values ​​are analyzed, and an optimization problem is constructed with the goal of synthesizing an electromagnetic field that conforms to ideal characteristics at the location of the device to be calibrated. The optimal transmission parameter configuration for each device is obtained by taking the transmission amplitude and phase of each healthy device as variables, the matching degree between the synthesized field and the ideal field as the optimization objective, and the device performance and safety specifications as constraints. The transmission parameter configuration that achieves the highest performance score and the corresponding measured performance index are output as the optimal solution for this collaborative calibration, and the array is controlled to transmit according to this configuration, thereby actually generating a virtual ideal calibration field at the target location.

Images