Transformer temperature controller automatic calibration method based on image recognition

Abstract

The invention provides a transformer temperature controller automatic calibration method based on image recognition, and relates to the technical field of metering calibration of computer vision, and the method comprises the steps: injecting a pseudo-random current sequence into a measured temperature controller to recognize thermal response characteristic parameters; extracting a visual approximation vector of an instrument pointer in real time through machine vision; a visual coupling type thermal impedance feedforward observer is used for fusion calculation of the state residual error of the effective temperature and the visual reading of the thermal bulb, and a double-source cooperative calibration control instruction is generated; and in response to the instruction, adjusting the temperature change rate of the external temperature field and the internal fine-tuning heat flow in parallel, and driving the pointer to approach the verification point according to a preset track. Through double-source cooperative control and dynamic thermal impedance modeling, the contradiction between heat conduction lag and visual feedback delay in traditional calibration is effectively solved, and high-precision and high-efficiency automatic calibration is realized.

Description

Technical Field

[0001] This invention relates to the field of computer vision metrology and calibration technology, specifically to an automatic calibration method for transformer temperature controllers based on image recognition. Background Technology

[0002] In the field of image analysis and intelligent chip interaction technology, and specifically in precision metrology and calibration scenarios involving power system transformer winding temperature controllers (BWR series), the deep penetration of computer vision technology has led to a significant trend in the industry's digital transformation: replacing manual readings with non-contact optical sensing. This field is dedicated to evolving static image recognition and analysis into dynamic physical state feedback control, aiming to achieve real-time closed-loop monitoring of the measured physical quantity through high-frequency visual data streams.

[0003] Existing automated testing technologies still face objective limitations when handling calibration tasks for physical objects with high thermal inertia. Traditional calibration methods primarily rely on an external constant-temperature bath as a single heat source, which, limited by the large specific heat capacity of the oil bath medium, results in significant heat conduction lag. Specifically, the effective temperature inside the sensing bulb often precedes the mechanical indication of the dial pointer. This "spatiotemporal asynchrony" makes it difficult to accurately capture the microsecond-level contact action threshold under rapid heating conditions. Although existing technology (publication number CN119152174A) improves the accuracy of static readings through image distortion correction and data dial model establishment, it is essentially still a "passive observation" method and fails to solve the inherent lag problem of visual readings relative to the actual thermal state of the sensing bulb. Furthermore, although technologies using internal current for thermal simulation exist, they are currently mostly used as independent static temperature rise testing methods. They lack a control mechanism for deep spatiotemporal fusion of the high-frequency response characteristics of internal current excitation and machine vision feedback data, making it difficult to balance calibration efficiency and steady-state accuracy during dynamic temperature changes. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic calibration method for transformer temperature controllers based on image recognition, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An automatic calibration method for transformer temperature controllers based on image recognition, comprising the following steps: S1: Drive the internal current excitation circuit to generate a pseudo-random current sequence and obtain thermal response characteristic parameters that characterize the dynamic characteristics of the thermal response of the temperature controller under test under the pseudo-random current excitation. S2: Obtain the visual approximation vector representing the geometric difference between the current position indicated by the instrument pointer of the temperature controller under test and the preset target position; S3: Using a pre-configured visually coupled thermal impedance feedforward observer, the thermal response characteristic parameters and the visual approximation vector are fused and calculated to solve for the state residual between the real-time effective temperature of the temperature sensor of the temperature controller under test and the visual reading. Based on this, a dual-source collaborative calibration control command is generated to characterize the real-time thermal hysteresis state and critical approximation degree of the temperature sensor of the temperature controller under test; and S4: In response to the dual-source collaborative calibration control command, the temperature change rate of the external temperature field generator and the fine-tuning heat flow injected into the internal current excitation circuit are adjusted in parallel to drive the instrument pointer to approach and trigger the calibration contact along a preset dynamic trajectory.

[0006] Compared with the prior art, the beneficial effects of the present invention are: By introducing a visually coupled thermal impedance feedforward observer in step S3, combined with the dual-source coordinated adjustment in step S4, the physical limits of single-heat-source control are overcome. The microsecond-level response characteristics (second component) of the internal current excitation circuit are used to compensate for the large thermal inertia hysteresis of the external temperature field (first component). Through nonlinear weighted dual-source energy distribution, the effective temperature of the sensing bulb can be highly synchronized with the visual reading. This effectively solves the pointer indication error caused by thermal conduction hysteresis in traditional methods, achieving "zero overshoot" approximation under rapid temperature change conditions.

[0007] By injecting a pseudo-random current sequence in step S1 to identify thermal response characteristic parameters, and by updating the dynamic thermal inertia compensation in step S3, it can adaptively follow the individual differences and aging characteristics of the temperature controller under test. Compared with the existing technology CN119152174A, which only performs geometric correction on static images, this technology establishes a multi-dimensional transfer function model of "current-thermal-visual," which can calculate the state residual between the virtual effective temperature of the temperature sensor and the visual reading in real time, thus maintaining high calibration accuracy even under non-steady-state conditions.

[0008] Step S4 integrates a graceful degradation strategy based on statistical variance calculation. By establishing a sliding window to monitor the confidence level of the visual proximity vector in real time, when encountering interference from oil mist, sudden changes in illumination, or occlusion that cause visual data distortion, the high-frequency fine-tuning signal can be automatically cut off and switched to the standard open-loop mode; ensuring operational safety in harsh industrial environments and avoiding control divergence or equipment damage caused by visual feedback noise. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the technical route for an automatic calibration method for transformer temperature controllers based on image recognition. Figure 2 This is a roadmap for the dual-source collaborative calibration control command generation technology of the present invention; Figure 3A numerical verification diagram of visual feedback confidence and dual-source control response; Figure 4 A roadmap for dual-source collaborative calibration control command response technology; Figure 5 Wiring diagram for thermal simulation test principle of temperature controller. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0011] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0012] Example 1: Please see Figures 1 to 5 The present invention provides a technical solution: An automatic calibration method for transformer temperature controllers based on image recognition, applied to an automated testing environment including an external temperature field generator and an internal current excitation circuit, includes the following steps: S1: Drive the internal current excitation circuit to generate a pseudo-random current sequence and obtain thermal response characteristic parameters that characterize the dynamic characteristics of the thermal response of the temperature controller under test under pseudo-random current excitation. S2: Obtain the visual approximation vector representing the geometric difference between the current position of the instrument pointer of the temperature controller under test and the preset target position; S3: Utilizing a pre-configured visually coupled thermal impedance feedforward observer, the thermal response characteristic parameters and visual approximation vector are fused and calculated to determine the state residual between the real-time effective temperature of the temperature sensor of the tested temperature controller and the visual reading. Based on this, a dual-source collaborative calibration control command is generated, characterizing the real-time thermal hysteresis state and critical approximation degree of the temperature sensor of the tested temperature controller; and S4: In response to the dual-source collaborative calibration control command, it adjusts the temperature change rate of the external temperature field generator and the fine-tuning heat flow injected into the internal current excitation circuit in parallel to drive the instrument pointer to approach and trigger the calibration contact along a preset dynamic trajectory.

[0013] Further explanation: Step S1 includes: controlling the current source to inject a pseudo-random current sequence with preset spectral characteristics into the thermal simulation circuit of the temperature controller under test; The instantaneous displacement data of the instrument pointer in response to the pseudo-random current sequence is captured at a preset frame rate using an image acquisition device; and the following parameter identification and calculation steps are performed: An observation vector is constructed, which contains the instantaneous displacement data of the previous moment and the pseudo-random current sequence of the current moment. An autocorrelation matrix and a cross-correlation vector are constructed based on the observation vector. Matrix operations are performed on the autocorrelation matrix and the cross-correlation vector to obtain thermal response characteristic parameters, including the thermal inertia time constant and the steady-state gain.

[0014] Further explanation: Step S2 includes: Obtain the target coordinate vector of the preset target position in the image coordinate system; Real-time coordinate vector of the instrument pointer tip in the image coordinate system is extracted using machine vision algorithms; and Calculate the Euclidean distance and direction vector between the real-time coordinate vector and the target coordinate vector, and combine the Euclidean distance and direction vector into a visual approximation vector.

[0015] Further explanation: The visual-coupled thermal impedance feedforward observer in step S3 performs the following fusion calculations: Based on thermal response characteristic parameters, a thermal impedance transfer function model describing the thermal conduction characteristics of the temperature controller under test is constructed. Using the thermal impedance transfer function model, the virtual effective temperature of the sensing bulb is estimated based on the current external temperature field and the injected current. Calculate the state residual between the virtual effective temperature and the visual temperature derived from the visual approximation vector; and Based on the state residual and the magnitude of the visual approximation vector, a dual-source collaborative calibration control command is generated, which includes a first component for adjusting the external temperature field and a second component for adjusting the internal fine-tuning heat flow.

[0016] Further explanation: After calculating the state residual between the virtual effective temperature and the visual temperature derived from the visual approximation vector, a dynamic thermal inertia compensation update step is also included: Obtain the thermal hysteresis decay coefficient and preset learning rate parameters from the previous control cycle; obtain the sign characteristics of the virtual effective temperature estimated in the previous time step; Perform parameter correction gradient calculation: Multiply the state residual, the sign feature of the virtual effective temperature, and the learning rate parameter to obtain the correction increment; Execution coefficient update operation: The correction increment is added to the thermal hysteresis decay coefficient of the previous control cycle to obtain the dynamic thermal hysteresis decay coefficient at the current moment, so that the estimation process of virtual effective temperature can adaptively follow the changes in the heat conduction behavior of the measured temperature controller.

[0017] Further explanation: Based on the state residual and the magnitude of the visual approximation vector, a dual-source collaborative calibration control command is generated, specifically including the execution of a nonlinear weighted dual-source energy allocation step: Define a continuous normalized weighting function based on the magnitude of the visual approximation vector; Using a normalized weighting function, we calculate the macroscopic weighting coefficients that characterize the priority of external temperature field regulation, and the microscopic weighting coefficients that characterize the priority of internal fine-tuning heat flow regulation. Calculate the first and second ideal adjustment quantities based on the state residuals, respectively; The first ideal adjustment quantity and the macroscopic weight coefficient are weighted to generate the first component, and the second ideal adjustment quantity and the microscopic weight coefficient are weighted to generate the second component. This achieves shock-free coordinated control, which is mainly based on external temperature field adjustment in the large deviation stage and smoothly transitions to mainly based on internal fine-tuning heat flow in the critical approach stage.

[0018] The following is a detailed implementation description of the above content: In the traditional calibration process of transformer temperature controllers, due to the complex oil / gas / metal medium heat conduction process inside the temperature sensing bulb, its response to the external temperature field exhibits a significant "thermal hysteresis" characteristic. Existing static comparison methods often ignore this dynamic characteristic, resulting in the instrument pointer reading always lagging behind the actual temperature field during rapid temperature rise calibration, causing "overshoot" or "inaccuracy"; moreover, relying solely on the temperature adjustment of the external constant temperature bath results in large thermal inertia, making it impossible to achieve micron-level pointer fine-tuning. This embodiment proposes an automatic calibration method based on a "visually coupled thermal impedance feedforward observer". This method uses pseudo-random excitation to extract the instrument's thermal inertia time constant and steady-state gain; using the Joule heat generated by the internal current loop as a "fast variable", combined with the "slow variable" characterized by the external temperature field, it achieves zero-hysteresis accurate capture of the pointer position; specifically: Before performing the specific calibration steps, an automated testing environment based on an integrated transformer temperature controller testing station was constructed. The hardware configuration and parameter settings for this environment are as follows: An external temperature field generator is configured: specifically, an intelligent precision thermostatic bath is used as the background temperature field source. The temperature control range of this device is set to cover (-30~160)℃, and its temperature field uniformity is ensured to be better than 0.02℃, and its stability is better than 0.01℃. In this method, the external temperature field generator is configured to respond to the first component of the subsequently generated control command, and is responsible for providing the macroscopic reference temperature environment.

[0019] See Figure 5An internal current excitation circuit is constructed, consisting of a high-accuracy programmable AC power supply (as a power frequency current generator), a converter, and a heating wire connected in series inside the temperature sensing bulb of the temperature controller under test. The current output range of the programmable AC power supply is set to (0.000~5.000) A, with an accuracy class of not less than 0.1. In this method, the internal current excitation circuit serves a dual function: it outputs a pseudo-random sequence during the system identification phase and outputs a high-frequency micro-current heat flow during the fine-tuning phase.

[0020] Image acquisition equipment deployment: An image recognition camera is used and mounted on a fixed bracket on a dedicated workbench, with its optical axis facing the dial of the temperature controller under test. The camera is configured to acquire dial images at a preset high frame rate of 200fps to extract real-time pointer readings and movement trajectories as visual feedback input. A multi-channel process signal scanning device is used, connected to the contact output terminal and remote signal terminal of the temperature controller under test, to monitor the contact continuity and remote current / resistance signals in real time.

[0021] Further, the first stage is executed, specifically the dynamic identification of thermal response characteristic parameters: In this stage, active dynamic modeling is performed using the internal current excitation loop. The excitation sequence configuration loading step is then executed: Based on a pre-built spreadsheet file of an external structured data carrier, the data loading process is initiated. Key parameters used to define the excitation signal characteristics are read and parsed. Specifically, the "sequence amplitude parameter" is read, denoted as... ; The determination logic comprehensively considers the lower limit of the signal-to-noise ratio of the image recognition system and the upper limit of the thermal safety of the temperature sensing bulb of the temperature controller under test, specifically including the following constraints: First, a lower limit is set by the signal-to-noise ratio constraint: when the injected current is less than 200mA, the tiny pointer displacement (<0.5 degrees) caused by the generated Joule heat will be submerged in the pixel noise of image recognition (caused by illumination fluctuations), causing the system recognition algorithm to diverge.

[0022] Secondly, an upper limit is set based on thermal safety constraints: According to the physical structure of the temperature sensor in the transformer temperature controller, a long-term injection of current exceeding 1000mA could lead to localized overheating and vaporization of the internal expansion medium, disrupting the instrument's linearity. Based on these dual boundary constraints, 500mA is selected as the optimal equilibrium point in this embodiment. This value is sufficient to generate a recognizable displacement of 10 to 20 pixels, while keeping the temperature rise within a safe range.

[0023] Simultaneously, the "bandwidth parameter" and "sequence duration parameter" are read to ensure that the excitation signal contains frequency components to excite the dynamic modes of the system, while avoiding overheating and damage to the sensing bulb due to excessive energy. The pseudo-random excitation injection step is then performed: based on the analyzed "bandwidth parameter" and "sequence duration parameter," a high-accuracy programmable AC power supply is controlled to inject a pseudo-random current sequence with preset spectral characteristics into the thermal simulation circuit of the temperature controller under test. This pseudo-random current sequence is coupled to the heating wire inside the sensing bulb of the temperature controller under test via a converter, generating a rapidly changing Joule thermal excitation, which is denoted as... Simultaneously, an image recognition camera mounted on a dedicated workbench synchronously acquires an image stream from the dial of the temperature controller under test at a preset high frame rate of 200 frames per second (fps). For each frame, an optical flow algorithm is used to extract the instantaneous displacement data sequence of the instrument pointer in real time, referred to as the angle sequence. This data serves as transient response data to thermal excitation and is characterized as instantaneous displacement data.

[0024] The system identification and calculation steps include: defining the "steady-state gain" (symbol: ...). ), characterizing the efficiency of converting electric current energy into mechanical displacement; defining the "thermal inertia time constant" (symbol: The value represents the speed of the system response. Based on a first-order transfer function model with pure time delay, the following is set... Joule thermal excitation of the input sequence is performed using the least squares method. and the angle sequence of the output sequence Perform time-domain fitting calculations. The pure time lag parameter, denoted as L, refers to the time delay from the moment the input excitation signal undergoes a step change until the system output response begins to change significantly (exceeding the noise floor). This parameter characterizes the transmission delay of heat from the heating wire to the temperature-sensing medium and then driving the bellows. Example values: Within the first 2.5 seconds after current injection, the instrument pointer has not yet produced a visible displacement due to the physical path delay of heat conduction. The existence of this parameter suggests that the control algorithm must consider this dead zone time when performing predictive control to avoid overshoot. The complex frequency domain variable is denoted as... : is the Laplace transform operator, used to transform time-domain differential equations into complex frequency-domain algebraic equations to facilitate transfer function analysis.

[0025] Perform the following least-squares recursive identification steps: Obtain the input dataset consisting of "pseudo-random current sequences". and the output dataset consisting of "instantaneous displacement data" Based on the discretized first-order hysteresis model structure, an observation vector is constructed. This observation vector contains two elements at each time step: the negative of the output angle value at the previous sampling time step, and the input current value at the current sampling time step.

[0026] Using the input and output datasets, construct the "autocorrelation matrix". Specifically, multiply the observation vector with its own transpose, and sum the results over all sampling times.

[0027] Construct a "cross-correlation vector". Specifically, multiply the observed vector at each time step with the actual output angle value at that time step, and sum the results of the calculation over all sampling times.

[0028] Perform a matrix inversion operation on the autocorrelation matrix to obtain the inverse matrix. Then, perform matrix multiplication on this inverse matrix and the cross-correlation vector to obtain a parameter vector containing two elements. Extract the discrete pole values ​​from the first element of the parameter vector and convert them into thermal inertia time constants using the natural logarithm based on the sampling period. The discrete gain value is extracted from the second element of the parameter vector, and the steady-state gain is calculated by combining it with the discrete pole values. .

[0029] It should be noted that the following data stream preprocessing steps must be performed before the identification and calculation: Considering the thermal inertia characteristics of the temperature controller under test, and to avoid numerical instability caused by discrete poles approaching the unit circle, this embodiment defines an identification sampling period. In this preferred embodiment, the following is set: This indicates a downsampling factor of 20. From the instantaneous displacement data sequence In the process, a sample is extracted every 20 data points to construct a sparse displacement sequence; for the pseudo-random current sequence The same extraction operation is performed to construct a sparse excitation sequence. Using this sparse displacement sequence and the sparse excitation sequence as input, subsequent least squares operations are performed. This ensures that the discrete poles calculated subsequently are distributed within the numerically stable effective interval.

[0030] The second phase involves the real-time construction of the visual approximation vector: In this phase, the temperature approximation process in physical space is mapped to a geometric approximation process in image space. The target reference locking step is performed as follows: Before calibration begins, a template matching operation is performed on the acquired dial image to identify the feature regions of the preset calibration point scale lines and extract their geometric centers in the image coordinate system, generating a target coordinate vector composed of horizontal and vertical coordinates. .

[0031] Real-time pointer tracking steps: During the temperature bath calibration process, dial images are continuously acquired. For each image frame, color thresholding and contour extraction operations are performed to locate the pixel position of the pointer tip and generate a real-time coordinate vector. .

[0032] Perform the vectorized difference calculation step: Obtain the x-coordinate component of the target coordinate vector. The horizontal component of the real-time coordinate vector Simultaneously, obtain the ordinate component of the target coordinate vector. The ordinate component of the real-time coordinate vector .right and Perform subtraction and square the difference to obtain the squared term of the lateral deviation; and Subtraction is performed, and the difference is squared to obtain the longitudinal deviation squared term. The lateral deviation squared term is added to the longitudinal deviation squared term to obtain the total deviation sum of squares. The arithmetic square root of the total deviation sum of squares is performed to obtain the non-negative Euclidean distance D. This value represents the straight-line pixel distance between the pointer tip and the target position on the image plane. Simultaneously, a "direction vector" is calculated based on the ratio of the lateral deviation to the longitudinal deviation. The Euclidean distance value and the direction vector are combined to construct the visual proximity vector.

[0033] The third phase involves visually coupled thermal impedance feedforward observation and dual-source control: a "visually coupled thermal impedance feedforward observer" is used to achieve dual-source coordinated control, specifically including: Virtual effective temperature, denoted as The physical meaning is the equivalent temperature of the core medium of the temperature sensing bulb after considering the thermal conduction hysteresis. It is an internal state variable estimated based on the model.

[0034] Thermal hysteresis decay coefficient, denoted as Physically, this is a dimensionless coefficient characterizing the system's ability to maintain its original thermal state within the current sampling period. Its value is related to the thermal inertia time constant obtained in step S1. There is a mathematical mapping relationship.

[0035] Detailed explanation of the "Dynamic Thermal Inertia Compensation Update Steps": To enable the model to adaptively follow the aging state of the sensing bulb, the following gradient descent parameter correction steps are performed: Step 1 (Baseline Acquisition): Obtain the thermal hysteresis decay coefficient of the previous control cycle. and preset learning rate parameters Learning rate parameter The sensitivity of the model to errors is defined.

[0036] The physical meaning of the state residual E is defined as the deviation between the visual reading and the model estimate, used to characterize external environmental disturbances or model parameter drift. The state residual at the current moment is obtained and denoted as E; the calculation logic of the state residual is as follows: ; and the virtual effective temperature of the previous moment. The correction increment is calculated using the gradient descent principle. This embodiment calculates the state residual E and the virtual effective temperature from the previous time step. The function value and the learning rate parameter The three components undergo continuous multiplication. This logic is derived from the objective function of minimizing the sum of squared residuals, ensuring that the adjustment direction of the thermal hysteresis decay coefficient always aims to reduce prediction error. The correction increment is then compared with the thermal hysteresis decay coefficient of the previous control cycle. Perform an addition operation and limit the result to the open interval (0,1) to obtain the updated thermal hysteresis decay coefficient at the current time. .

[0037] Define visual temperature as The physical meaning is the current displayed temperature, derived from the pointer position perceived visually and calculated using the dial mapping. This is used in calculating visual temperature. At that time, the vector-scalar mapping logic is executed: specifically, using the image scale distribution pattern obtained in S1, the Euclidean distance D between the pointer tip and the target scale line is calculated. The polarity of the temperature deviation is determined using the direction vector: the dot product of the direction vector and the preset "temperature increase direction reference vector" is calculated. If the dot product is positive, the current visual temperature is determined to be lower than the target temperature, and the Euclidean distance D is assigned a negative sign; if the dot product is negative, the current visual temperature is determined to be higher than the target temperature, and the Euclidean distance D is assigned a positive sign. Based on the Euclidean distance D with its sign attribute, the absolute visual temperature is calculated. This step ensures that the directional information in the visual approximation vector is directly used to determine the algebraic sign of the temperature deviation, thus solving the technical problem that "overheating" and "underheating" cannot be distinguished by distance magnitude alone.

[0038] The reference vector for the heating direction is denoted as... This refers to a pre-defined unit direction vector in the image coordinate system used to characterize the increasing trend of the scale value on the measured instrument's dial. Specifically, for a circular dial, this vector represents the tangential clockwise (or counterclockwise, depending on the instrument's characteristics) direction at the target scale line; for a linear dial, this vector represents the linear direction of the scale value increase. This vector serves as a geometric reference, used to determine the polarity of the pointer's current position relative to the target position (i.e., overheating or underheating state) through vector operations. Heating direction reference vector ( Example values ​​for ) are: The current calibration point is located at the "10 o'clock" position on the circular dial. At this position, the scale value increases clockwise. In the established two-dimensional image coordinate system, the tangent direction at this point points to the "upper right." After normalization, the unit vector of this tangent direction is set to... This corresponds to a 45-degree angle. When the dot product of the real-time calculated pointer deviation vector and the reference vector is positive, it indicates that the pointer is deflected to the upper right (clockwise), meaning the temperature is too low (underheating); conversely, it indicates overheating.

[0039] The preset target temperature parameter is denoted as... This refers to the theoretical physical temperature reference value that the instrument under test should indicate, determined according to the preset calibration procedure, in the current calibration process. This parameter serves as the reference offset for absolute visual temperature calculation and corresponds to the physical properties of the target scale line on the instrument panel. Preset target temperature parameter ( Example values ​​for ) This automatic calibration process performed high-temperature range verification, and the JJG verification procedure was selected. The scale line serves as the current calibration target.

[0040] Image-physical domain mapping coefficients, denoted as , refers to the conversion scaling factor used to describe the mapping relationship between spatial distance in the image coordinate system and the change in the physical temperature scale.

[0041] Its physical dimension is temperature units / pixel units, defined as ℃ / pixel. The image-physical domain mapping coefficients are pre-obtained based on the static feature recognition in step S1, and are used to map the Euclidean distance on the image plane. Converted linearly or nonlinearly to a physical temperature deviation. Image-physical domain mapping coefficients ( Example values ​​for ) Based on the visual recognition and static calibration of the dial scale in step S1, the dial markings were measured. to The arc length between the scale lines corresponds to 250 pixels in the image. This embodiment calculates the linear resolution of this region as follows: This indicates that for every 1 pixel the pointer tip deviates from the target position, it represents... Temperature error.

[0042] The dual-source collaborative calibration control command is defined, comprising two components: the "first component" adjusts the heating rate of the constant temperature bath characterized by the external temperature field generator; the "second component" adjusts the output current intensity of the fine-tuning heat flow within the internal current excitation circuit. To address the cumulative error problem inherent in relying solely on model estimation and the hysteresis problem inherent in relying solely on visual feedback, this embodiment establishes a fusion mechanism of "dynamic thermal inertia compensation and nonlinear weighted energy allocation," specifically: Implement a dynamic thermal inertia compensation update mechanism: if a fixed thermal hysteresis decay coefficient is set... It cannot adapt to changes in thermal resistance caused by the aging of the temperature sensing element. This preferred embodiment introduces dynamic update logic: in each control cycle, the visual temperature is calculated. Compared with the virtual effective temperature estimated at the previous moment The difference between the two values ​​is used to obtain the current state residual E, and then the thermal hysteresis decay coefficient correction calculation is performed.

[0043] In this embodiment, for the learning rate parameter Determination of: It is a dimensionless adjustment factor used to control the thermal hysteresis decay coefficient. The update step size is determined in response to the state residual E. Based on experimental calibration of the thermal characteristics of the oil-immersed temperature sensing bulb, this embodiment uses the learning rate parameter... The preferred value range is limited to [0.001, 0.05]. In the most preferred embodiment, the learning rate parameter... It was set to 0.01.

[0044] Define an upper limit constraint: when the learning rate parameter At that time, thermal hysteresis decay coefficient It may overreact to measurement noise, leading to an abnormal virtual effective temperature. High-frequency oscillations, which are not physical phenomena, occur.

[0045] Define a lower bound constraint: when the learning rate parameter At that time, thermal hysteresis decay coefficient The update speed is too slow and cannot effectively compensate for drift caused by the aging of the temperature sensor within a 30-minute calibration cycle.

[0046] The learning rate parameter is multiplied by the current state residual E to obtain the correction increment; this correction increment is then added to the current thermal hysteresis decay coefficient. Above; it can use visual feedback to "calibrate" its own physical parameters in real time, so that the estimated virtual effective temperature is accurate. It is getting closer and closer to the actual internal temperature.

[0047] Furthermore, the use of a threshold switching method (using a constant temperature bath for the far field and current for the near field) leads to oscillations in the switching point control. This preferred embodiment introduces continuous weighting logic: specifically defining a normalized weighting function. In this embodiment, this normalized weighting function is configured as a variant of the Sigmoid function. Its input is the Euclidean distance value D in the "visual proximity vector," and its output is the "micro-weighting coefficient," denoted as... The value range is [0,1]; the logic for determining the micro-weight coefficients is as follows: Obtain the current Euclidean distance D and the preset critical section threshold parameters. For the Euclidean distance D and The distance deviation is obtained by performing a subtraction operation. The specific steps for determining the critical zone threshold parameter are as follows: Install a standard sample meter on the transformer temperature controller testing platform and set the external temperature field generator to full-speed heating mode. Record the movement trajectory of the dial pointer using an image acquisition device, and simultaneously record the real-time oil temperature of the constant temperature bath using a high-precision reference thermometer. Calculate the maximum thermal hysteresis deviation between the pointer-indicated temperature and the real-time oil temperature under different temperature rise rates. Convert this thermal hysteresis deviation into pixel distance in the image coordinate system. Select the critical distance value at which the "thermal hysteresis deviation" begins to exceed the allowable error limit (the pixel distance corresponding to 0.1℃ in this embodiment) in all test conditions. Set the critical distance value as the critical zone threshold parameter. .

[0048] This ensures that the internal microcurrent is activated for compensation only when the thermal inertia of the external temperature field is about to cause out-of-tolerance, thus achieving an optimal balance between energy consumption and accuracy.

[0049] A preset conversion slope parameter is obtained and denoted as k1. The conversion slope parameter k1 controls the steepness of the switching process. A multiplication operation is performed on the distance deviation and the negative value of the conversion slope parameter k1 to obtain the exponential term input value. In this embodiment, the configuration of the conversion slope parameter k1 is as follows: the conversion slope parameter k1 is used to define the weighting function Sigmoid curve at the critical threshold. The steepness of the surrounding terrain, expressed in pixels (in units). In this embodiment, the transformation slope parameter k1 is set to 0.5. This value is set so that the micro-weighting coefficients... The range from 0.1 to 0.9 spans 10 pixels, corresponding to a temperature difference of approximately 0.2℃. If the conversion slope parameter k1 is too large (k1>5.0), the switching process approximates a hard switch, easily causing abrupt changes in control commands and mechanical shocks; if the conversion slope parameter k1 is too small (k1<0.1), internal current will begin to intervene even when the target is far away, causing unnecessary energy consumption and thermal interference. The input value of the exponent term is processed using the natural constant. The exponentiation is performed with 1 as the base to obtain the intermediate exponent value. The value 1 is then added to the intermediate exponent value to obtain the denominator; the reciprocal of this denominator is then performed to obtain the micro-weight coefficient. When the Euclidean distance D is far from the target, the micro-weighting coefficients... Approaching 0; when the Euclidean distance D approaches the target, the micro-weighting coefficients... The smoothness approximates 1. Simultaneously, the "macro-weight coefficient" is calculated as follows: .

[0050] It should be noted that this embodiment is based on the Euclidean distance D obtained in step S2 and the critical region threshold parameter defined in step S3. Using the Objective Threshold Derivation (OTD) methodology, the calibration process is divided into two control stages with clearly defined physical boundaries in the time domain. Define the large deviation stage: This stage corresponds to the Euclidean distance D represented by the magnitude of the visual proximity vector satisfying the following conditions. The interval within which the micro-weighting coefficients are calculated. Approaching 0, and The macro weighting coefficient ( The value approaches 1. At this point, the instrument pointer is far from the target position, and the main displacement error comes from the fact that the external temperature field has not yet reached the set value.

[0051] The corresponding control strategy is in the "macro temperature field dominant mode". The controller mainly responds to the first component of the dual-source command, driving the external temperature field generator to heat up at the maximum allowable rate to quickly compress the calibration time; at this time, the second component (internal fine-tuning heat flow) is suppressed due to its small weight, avoiding the introduction of unnecessary Joule thermal interference when moving away from the target.

[0052] Define the critical approximation stage: This stage corresponds to the Euclidean distance D represented by the magnitude of the visual approximation vector satisfying the following conditions: Within this interval, as the Euclidean distance D decreases, the output of the Sigmoid variant function undergoes a nonlinear jump, and the micro-weighting coefficients... The reading rapidly climbs from 0.1 to nearly 1.0 and becomes dominant. At this point, the instrument pointer has entered the neighborhood of the target scale line (in this embodiment, the deviation is less than 0.2℃). The residual at this point is mainly limited by the thermal inertia of the external temperature field, and simply relying on external heating can no longer achieve micron-level precise positioning. The corresponding control strategy is to smoothly transition to the "microscopic heat flow-dominated mode".

[0053] Execute external actions: The weight of the first component is reduced, and the external temperature field generator automatically reduces the heating rate or enters a temperature lock state to create a quasi-static thermal environment.

[0054] Execution of internal actions: The second component (internal fine-tuning heat flow) is fully activated. Utilizing the millisecond-level response of Joule heat as a "fast variable," it directly drives the expansion of the medium inside the temperature sensing bulb, pushing the pointer to precisely eliminate the final static error along a preset dynamic trajectory until the verification contact is triggered.

[0055] 1) Further, based on the above definition and mechanism, the following complete control loop is executed: Obtain the current external thermostat temperature sampling value. and internal injection current value Obtain the estimated virtual effective temperature from the previous time step. Using the steady-state gain obtained in step S1 And the corrected thermal hysteresis decay coefficient, then perform discrete difference evolution calculation: virtual effective temperature at the previous moment Multiplying the inertia term by the thermal hysteresis decay coefficient yields the inertia term; the injected current... With steady-state gain Perform a multiplication operation and compare the result with the temperature sample value of the external constant temperature bath. Perform addition to obtain the current total heat input; then perform addition on the current total heat input and the coefficient. Perform a multiplication operation to obtain a new input term; then perform an addition operation between the inertia term and the new input term to obtain the virtual effective temperature at the current moment. .

[0056] Obtain the visual temperature calculated from the current frame image. .right and Perform the subtraction operation to obtain the state residual E. This state residual is used to correct the thermal hysteresis decay coefficient in the next cycle.

[0057] Obtain the Euclidean distance D of the current visual proximity vector. Calculate the micro-weight coefficients based on the Euclidean distance D. and macro weighting coefficient .

[0058] Perform dual-source instruction synthesis: Generate the first component: Set the preset maximum heating rate constant, represented by the "first ideal adjustment amount"; combine this maximum heating rate constant with the macroscopic weighting coefficient. Perform multiplication. The farther Euclidean distance is from D, the faster the external temperature rises; the closer Euclidean distance is to D, the faster the external temperature rises, or even stops.

[0059] Before generating the second component, perform safety boundary constraint operations on the calculated original adjustment value: define hardware constraint boundaries: obtain the maximum output fine-tuning current of the current source. If the original adjustment amount Greater than the maximum output fine-tuning current Then the second component will be forcibly set to If the original adjustment is less than 0, the second component is forcibly set to 0. When the above limiting action is triggered, the accumulation and update of the integral term in the PID algorithm is paused to prevent system response overshoot caused by integrator oversaturation.

[0060] The maximum heating rate constant is denoted as , is a fixed parameter characterizing the upper limit of the physical heating capacity of an external temperature field generator under the premise of full-power operation and ensuring temperature field uniformity. This parameter is specifically calculated through the following heat capacity limit calibration steps: Obtain the rated electrical power of the heating unit of the external temperature field generator. (Unit: Watts), and the total mass of the medium oil in the constant temperature bath. (Unit: kg) and specific heat capacity of the medium oil (Unit: Joules / kg·°C). Obtain the preset overall system thermal efficiency coefficient. In this embodiment, 0.85 is used to characterize the effective energy percentage after heat loss. Using the principles of thermodynamic formulas, the rated electric power... With the overall thermal efficiency coefficient of the system Perform multiplication to obtain the effective heating power; calculate the total mass of the medium oil. With specific heat capacity Perform a multiplication operation to obtain the total heat capacity of the system. Then, perform a division operation between the effective heating power and the total heat capacity to obtain the theoretical maximum temperature rise rate. To ensure control stability, 80% of this theoretical value is selected as the maximum temperature rise rate constant in this embodiment. .

[0061] The second ideal adjustment is denoted as This is the target current intensity calculated using a discrete PID algorithm based on the current control deviation, used to drive the internal fine-tuning heat flux to eliminate residual errors. To achieve a smooth transition from rapid heating in the large deviation phase to millisecond-level fine-tuning in the critical approach phase, this step executes the following linearized PID calculation process based on the position tracking error: using the Euclidean distance of the visual approximation vector obtained in step S2... Including direction information, calculate the position tracking error of the pointer relative to the target position at the current moment. Specifically, the polarity of the deviation is determined by the direction vector defined in S3 (positive for overheating and negative for underheating), and the signed Euclidean distance is converted into an equivalent temperature deviation.

[0062] Perform PID discrete calculation: Obtain the preset proportional gain coefficient. For position tracking error and Perform the multiplication operation to obtain the proportional control component. This component is responsible for a rapid response to the current deviation. Obtain the preset integral gain coefficient. The current position tracking error Accumulate it into the historical error and variables, and then... Perform multiplication to obtain the integral control component. This component is used to eliminate the steady-state static error of the system. The preset differential gain coefficient is obtained. Calculate the current error Error compared to the previous time step The difference, and with Performing multiplication yields the differential control component. This component is used to suppress overshoot. The proportional control component, integral control component, and derivative control component are added together to obtain the total control quantity characterizing the heat power demand. For the total control quantity Perform the arithmetic square root operation to obtain the second ideal adjustment value. The micro-weighting coefficients obtained in step S3 are used. The second ideal adjustment amount Perform multiplication operations ( This ultimately generates a fine-tuning current command value to drive the internal current excitation loop. During the large deviation phase ( Automatically suppresses integral accumulation to prevent overcharging; during the critical approximation stage ( It seamlessly takes over control and uses internal microcurrents for millisecond-level thermal fine-tuning to ensure that the pointer triggers the verification contact with zero steady-state error.

[0063] Further explanation: Step S4 includes: In response to the first component of the dual-source collaborative calibration control command, a temperature rate adjustment signal is generated to drive the external temperature field temperature control device. In response to the second component of the dual-source cooperative calibration control command, a fine-tuned current modulation signal is generated to drive the internal current source; and When the confidence level of the visual approximation vector is lower than a preset threshold, a degradation strategy is automatically executed. The degradation strategy includes cutting off the fine-tuning current modulation signal and switching the temperature rate adjustment signal to a preset standard open-loop control mode.

[0064] Further explanation: The judgment process when the confidence of the visual proximity vector is lower than the preset threshold includes: establishing a sliding window with a preset time length; extracting the Euclidean distance values ​​of the visual proximity vectors at all historical moments within the sliding window; performing statistical variance calculation on the Euclidean distance value sequence and normalizing the reciprocal of the calculation result into a visual confidence index; and comparing the visual confidence index with the preset downgrade trigger threshold.

[0065] Further explanation: The process of generating a fine-tuning current modulation signal to drive the internal current source includes: obtaining a preset current modulation period parameter; obtaining the second component in the dual-source collaborative calibration control command, the second component representing the desired instantaneous injected power; calculating the target duty cycle based on the ratio of the second component to the preset maximum output fine-tuning current; and generating a high-frequency on / off control sequence using the target duty cycle within the current modulation period as the fine-tuning current modulation signal.

[0066] The following are specific implementation instructions for the above content: Visual confidence index, denoted as This represents the stability and reliability of the output data of the current visual feedback system under harsh operating conditions (in this embodiment, harsh operating conditions include high-temperature oil and gas disturbances and liquid surface fluctuations). The visual confidence index is a dimensionless normalized value, ranging from [0,1]. It is calculated by performing statistical dispersion analysis on the historical sequence of the "visual proximity vector". The closer the value is to 1, the more stable the visual tracking; the closer the value is to 0, the more noise or loss of the visual signal.

[0067] The downgrade trigger threshold is denoted as Define the critical judgment boundary for automatic switching from "high-precision closed-loop collaborative mode" to "standard open-loop control mode". The determination logic is as follows: Configure a standard test bench including an external temperature field generator, an internal current excitation circuit, and an image acquisition device. Introduce an adjustable concentration of "artificial oil mist generator" to simulate visual interference environments of different degrees. In an oil mist-free environment, perform a full-process calibration and record the "steady-state control error" when the instrument pointer reaches the preset target position, denoted as... Keeping the target temperature constant, the oil mist concentration is gradually increased in 5% increments according to a preset gradient. At each concentration gradient, two real-time data sequences are continuously recorded: the calculated visual confidence index... The real-time control error of the pointer under the current operating condition is denoted as... When real-time control error is detected When the error exhibits exponential divergence and exceeds the preset "maximum permissible error of open-loop control" (in this embodiment, this is derived from the statistical upper limit of 0.5℃ of historical open-loop data), that moment is marked as the "control failure critical point." The visual confidence index value corresponding to the "control failure critical point" is extracted, and a safety margin of 10% is added to this extracted value to ultimately determine the degradation trigger threshold. In this embodiment, the specific value of the downgrade trigger threshold is 0.3.

[0068] The external temperature field temperature change response coefficient is denoted as This characterizes the efficiency ratio of the external temperature field generator in converting the input control command value into the actual physical temperature change rate. Based on the pre-calibration of the constant temperature bath's hardware power and the specific heat capacity characteristics of the medium, the specific determination logic includes: setting the external temperature field generator to manual control mode. When in thermal equilibrium, a step heating command with an amplitude of [value missing] is applied. A step heating command (50% full power) is issued. The response curve of the medium temperature changing over time is recorded using a reference thermometer at a preset sampling frequency of 1 Hz. From the response curve In the analysis, a linear temperature increase segment is selected, from the point where the temperature begins to rise significantly to near saturation. A linear regression operation is performed on this segment to analytically obtain the time derivative of the temperature change, which is represented as the heating rate. Regarding the heating rate With step heating command amplitude Performing a division operation yields the external temperature field response coefficient. .

[0069] The current modulation period parameter is denoted as Define the basic time granularity for the internal fine-tuning current to execute pulse width modulation (PWM). This value is matched based on the "electromagnetic interference suppression theory." Since the temperature sensor and internal circuitry of the temperature controller under test are in a complex power frequency (50Hz) electromagnetic field environment, if the switching frequency of the PWM modulation is not synchronized with the power frequency, it will generate difficult-to-filter differential frequency interference at the signal acquisition end. This embodiment sets... Equal to the reciprocal of the power frequency period. In this embodiment, it is... It is used to ensure that each current switching operation is synchronized with the zero-crossing point or a specific phase of the power frequency voltage, thereby minimizing the electromagnetic coupling interference of high-frequency switching operations on the detection of weak contact signals at the physical level, and improving the signal-to-noise ratio without the need for additional complex filtering circuits.

[0070] Current modulation period parameters In this embodiment, the specific value is 20 milliseconds, which corresponds to one complete cycle of a 50Hz power frequency.

[0071] Furthermore, to address the issues of "inability to balance speed and accuracy" and "susceptibility to failure of visual feedback due to environmental interference" near the critical point, this embodiment constructs a "robust dual-source collaborative execution mechanism based on visual entropy flow monitoring".

[0072] 2.1) Implementation of Visual Entropy Flow Monitoring and Confidence Calculation Mechanism: Simply determining whether a target is detected ignores the "jumping" characteristic of the detection coordinates. This mechanism introduces temporal statistical analysis: maintaining a first-in-first-out (FIFO) data queue with a sliding window length of N. (Setting...) (Corresponding to an observation duration of 50ms, based on a frame rate of 200fps). This value is set to resolve the conflict between "transient noise suppression" and "real-time state tracking". If N is too small ( The calculated variance statistic will not fully reflect the random fluctuation characteristics of the data stream, leading to false triggering of the degradation strategy by single-frame image recognition jumps, causing control oscillations. If N is too large ( The update of the visual confidence index will have a significant time lag (>250ms), causing a delay in determining "reliability" and maintaining closed-loop control even when visual feedback has failed due to continuous oil mist, increasing the risk of overheating. The specific deterministic calculation logic is as follows: Obtain the sampling frame rate of the image acquisition device With the maximum allowable decision delay time ;right and Perform multiplication to obtain the theoretical upper limit of the total number of samples; select the upper limit of this theoretical upper limit. to Integers within the interval are used as the sliding window N to ensure sufficient safety margin. In each control cycle, the Euclidean distance in the "visual proximity vector" calculated in step S2 is added to this queue. The arithmetic mean of all data in this queue is calculated. The difference between each data point in the queue and the arithmetic mean is calculated, and the difference is squared to obtain a sequence of squared differences. The sequence of squared differences is summed, and the result is divided by the sliding window N to obtain the variance value. The preset stability constant 1.0 is added to the variance value to obtain the denominator; the reciprocal of the denominator is then used to obtain the current visual confidence index. This mechanism uses the reciprocal of the variance as the confidence level. When oil mist causes the visual recognition point to change drastically near the target, the variance increases and the confidence level drops rapidly, thus enabling the system to keenly perceive environmental noise.

[0073] 2.2) Execution of Dynamic Pulse Width Modulation Mapping Mechanism: For the second component (internal fine-tuning heat flow) in the dual-source command, conventional linear regulation introduces quantization errors due to power supply resolution limitations. This mechanism employs a time-division strategy: It obtains the second component (characterizing the ideal current intensity) in the dual-source collaborative calibration control command generated in step S3. It obtains the predefined hardware maximum output fine-tuning current parameter. It performs a division operation between the second component and the maximum output fine-tuning current parameter, limiting the result to the [0,1] interval to obtain the target duty cycle. It then compares the target duty cycle with the current modulation period parameter. The high-level holding time is obtained by performing a multiplication operation. By adjusting the high-level time, a continuously adjustable precision thermal power is equivalent to that of a temperature sensor with high thermal inertia, achieving microwatt-level heat flux control resolution.

[0074] 3.0) Based on the above definitions and mechanisms, the specific execution flow of step S4 is as follows: Receive the dual-source collaborative calibration control command generated in step S3. Separate the dual-source collaborative calibration control command into a first component data packet for controlling the external temperature field and a second component data packet for controlling the internal fine-tuning loop.

[0075] Using the aforementioned "visual entropy flow monitoring and confidence calculation mechanism", based on the most recent The visual proximity vector data of the frame is used to calculate the current visual confidence index in real time. The calculated visual confidence index Compared with the preset degradation trigger threshold Perform numerical comparison operations.

[0076] Set Branch 1 (High-Precision Closed-Loop Collaborative Mode): If the visual confidence index... Greater than or equal to the downgrade trigger threshold The visual feedback was deemed reliable.

[0077] Perform external temperature field conditioning: Combine the first component with the external temperature field temperature change response coefficient. Perform a multiplication operation to generate a temperature rate control signal, which drives an external temperature field generator to approach the target at that rate.

[0078] Internal microfluidic fine-tuning is performed: Utilizing the aforementioned "dynamic pulse width modulation mapping mechanism," a fine-tuning current modulation signal is generated based on the second component, driving the internal programmable AC power supply to output a pulse current with a corresponding duty cycle. This process utilizes the Joule heating of the heating wire inside the temperature sensing bulb to perform millisecond-level micro-superposition of the temperature sensing bulb's temperature, compensating for the hysteresis of the external temperature field.

[0079] Set branch two (standard open-loop control mode): If the visual confidence index Less than the downgrade trigger threshold The visual feedback is determined to be severely interfered with, including but not limited to being obscured by high-temperature oil mist.

[0080] Cut off the fine-tuning circuit: Force the data of the second component to be set to zero, stop the output of the internal current excitation circuit, and prevent incorrect heating caused by erroneous visual feedback.

[0081] Switch to open-loop scan: Ignore the first component based on the model generated in step S3, directly read the "standard linear heating rate parameter" of 0.5℃ / min from the preset configuration file, generate a standard open-loop variable temperature rate adjustment signal, and drive the external temperature field generator to pass through the calibration point at a constant speed.

[0082] In response to the generated drive signal, the external temperature field generator and the internal current excitation circuit work together to act on the temperature controller under test. As the effective temperature rises, the pointer of the drive instrument moves along a preset trajectory. The contact status signal of the temperature controller under test is monitored in real time. Once a change in the contact signal is detected (from open to closed, or vice versa), the current external standard temperature value and visual reading are immediately recorded, completing the calibration task for that calibration point.

[0083] Through the above implementation methods, a "confidence-based adaptive execution architecture" was achieved in step S4. Under normal operating conditions, precise capture with "zero thermal hysteresis" was achieved through fine-tuning of heat flow via PWM modulation; under visual failure conditions, a degradation strategy based on variance analysis ensured that the system would not diverge or shut down due to sensor noise, thereby guaranteeing high availability and intrinsic safety in industrial field applications.

[0084] The following are specific implementation instructions for the above content: Visual Confidence Index The domain is the interval [0,1]; when Time: Characterizes the "low-entropy steady state" of the visual feedback system. It indicates a small variance in the Euclidean distance between feature points in the image acquisition sequence, and a high linear correlation between visual readings and actual physical locations. In this state, the prerequisites for performing internal current fine-tuning are met, allowing dual-source collaborative control to intervene with maximum gain to achieve rapid approximation with zero thermal hysteresis.

[0085] when Time: Characterizing the "high-entropy chaotic state" of the visual feedback system. Physically, this corresponds to the random jumps in feature point recognition coordinates caused by strong oil mist interference, violent fluctuations in the liquid surface, or sudden changes in illumination. In this state, any closed-loop feedback based on visual errors will lead to positive feedback oscillations or erroneous energy injection. This trend forces the removal of fine-tuning components, degenerating into intrinsically safe open-loop control.

[0086] Define the variance of the Euclidean distance within the sliding window as: : and They exhibit a non-linear, negatively correlated relationship. The specific functional form is a variant of the inverse proportional function. In thermostat calibration scenarios, the pointer's thermal action is mechanically conducted, possessing significant thermal inertia and mechanical damping; its physical position cannot undergo drastic changes within milliseconds. High-frequency, large-amplitude position changes originate from measurement noise rather than physical motion. Using an inverse proportional function can... The variance domain is mapped to the confidence domain of (0,1), exhibiting high sensitivity when the variance is small and saturation characteristics when the variance is large. This nonlinear mapping directly supports the "degradation" technical effect of this embodiment. When environmental interference increases slightly, the visual confidence index only decreases slightly, maintaining collaborative control but reducing the weight; once the interference exceeds the critical point, the visual confidence index quickly falls below the corresponding threshold, immediately cutting off the fine-tuning loop. This ensures that in the complex electromagnetic and optical environment of industrial sites, the measured instrument will not experience uncontrolled heating due to transient noise from the sensor.

[0087] 4) Furthermore, this embodiment is configured in a high-fidelity digital twin industrial calibration scenario to simulate and verify the response characteristics of the "visually coupled thermal impedance feedforward observer" under dynamic interference environment.

[0088] The test object model is set as follows: an oil-filled pressure thermostat with a range of 0-200℃, and the standard calibration target temperature is set at 150℃. Environmental interference model: an "optical transmittance attenuation factor" is introduced as a dynamic variable to simulate the process of the oil fume concentration above the calibration constant temperature bath gradually increasing from 0% (clean) to heavy obstruction, in order to test the robustness of the system.

[0089] Downgrade trigger threshold Set to 0.3. Maximum output fine-tuning current. : Set to 100mA. External temperature field temperature change response coefficient. Set to 0.05. Visual sampling frame rate: 200fps, sliding window. .

[0090] During operation, the data flow follows the following logical path: A vision sensor captures dashboard images in real time, extracts feature point coordinates, and calculates the variance of the Euclidean distance between feature points within a sliding window. Based on this, the visual confidence index is calculated. → and Compare and determine the control strategy branch (high-precision closed-loop cooperative mode or standard open-loop control mode) → Based on the decision result, output the external temperature field temperature change rate. and internal fine-tuning current duty cycle .

[0091] The table below shows the calculated key state parameters and control command responses under different levels of environmental disturbance.

[0092] Table 1: Examples of system response calculations under different environmental disturbances This table aims to verify the nonlinear adaptive capability of the S4 algorithm in the face of environmental noise. The data intuitively presents the entire process of smoothly transitioning from "dual-source collaborative control" to "single-source open-loop control", highlighting the safe response logic of the control command when the visual confidence index approaches and falls below the corresponding threshold.

[0093] The cooperative energy efficiency ratio is used to quantify the contribution weight of the "internal fine-tuning heat flow" to the overall temperature control process in the method of this invention. The cooperative energy efficiency ratio is denoted as... This parameter is obtained by calculating the dimensionless proportion of the internal fine-tuning heat flux in the total effective temperature control energy. The calculation logic is as follows: in: The duty cycle of the internal fine-tuning current (range 0.0~1.0). The external temperature field temperature change rate (unit: °C / min); and These are pre-calibrated energy normalization weighting coefficients. In the experimental configuration of this embodiment, based on the thermal capacity characteristics of the temperature sensing bulb, the following settings are used: and This reflects the fact that, during the dynamic approximation phase, the internal Joule heating and external convective heating have control effectiveness of the same order of magnitude. When When the value approaches 0.5, it indicates that the method mechanism of this embodiment is in a state of deep synergy, driven by both internal and external heat sources, resulting in the fastest response speed; when When the value approaches 0, it indicates that the method mechanism of this embodiment relies entirely on external thermal inertia and is in a safe, baseline state.

[0094] Based on the above definitions, the data in Table 1 are analyzed as follows to verify the beneficial effects of this scheme: In Phase 1 (clean environment), with a visual confidence index of 0.952, a high duty cycle fine-tuning current of 85.0% was calculated, while maintaining a relatively fast external heating rate of 1.2℃ / min. The calculated synergistic energy efficiency ratio... This indicates that, under ideal operating conditions, the internal fine-tuning heat flow contributes 41% of the effective driving energy. This data directly proves the effectiveness of the "dual-source synergy" mechanism: by introducing an internal heat source, the thermal hysteresis problem of traditional external oil bath heating can be significantly overcome, achieving rapid and accurate driving of instrument pointers.

[0095] When environmental disturbances cause the visual confidence index to drop from 0.357 to 0.300 (approaching the degradation trigger threshold), the fine-tuning current duty cycle is compressed from 25.0% to 5.0%, corresponding to a cooperative energy efficiency ratio. The gain smoothly decreases from 0.24 to 0.08. This non-linear dynamic weight adjustment characteristic verifies the robustness of this scheme in the early stages of signal quality degradation. Compared with the hard threshold switching commonly found in existing technologies, by automatically reducing the fine-tuning gain in the critical region, the control signal oscillation caused by sensor noise is effectively suppressed, avoiding actuator malfunctions.

[0096] When the interference from cooking fumes intensifies, causing the visual confidence index to drop to 0.167 (below the degradation trigger threshold of 0.3), the example data shows that the fine-tuning current duty cycle is forcibly locked at 0.0%, and the synergistic energy efficiency ratio... The temperature is reset to zero; simultaneously, the external temperature change rate switches to the standard 0.5℃ / min. This data response confirms the execution logic of the invention's "degradation strategy": in extreme operating conditions where visual feedback is unreliable, the internal energy injection can be physically cut off, and the system automatically reverts to a preset safe open-loop scanning mode. This mechanism eliminates the risk of overheating of the sensing bulb or damage to the instrument due to visual recognition errors (including coordinate drift caused by dense smoke), ensuring high reliability in industrial field applications.

[0097] The core of this solution lies in determining the operating mode based on the visual confidence index represented by the perceived resource status. The following practical application intervals are defined. The boundaries of each interval are determined by data statistical regularity (analysis of variance) and preset safety control constraints.

[0098] Interval 1: High-fidelity dual-source collaborative interval. Its boundary is determined by the statistical upper limit of visual tracking error. Within this interval, the Euclidean distance variance within the historical sliding window is within an acceptable noise range. According to control theory analysis, the signal-to-noise ratio (SNR) of the visual feedback signal at this time is sufficient to support the stability requirements of the closed-loop control system and will not cause noise-driven divergent oscillations.

[0099] The corresponding operation is to execute full-function dual-source collaborative control. It outputs an external temperature field rate adjustment signal and an internal fine-tuning current modulation signal (PWM) in parallel. The weights of the two are dynamically calculated based on the state residuals, and the Joule heating effect of the internal current is used to compensate for the thermal hysteresis of the external temperature field, achieving millisecond-level temperature tracking and calibration.

[0100] Interval Two: Downgraded Guarantee Interval Its boundary is determined by the experimental calibration value of the control failure critical point. Within this interval, the variance of the visual data exceeds the statistically permissible "steady-state range," indicating that environmental disturbances have caused the visual measurement results to decouple from the actual physical location. If fine-tuning current is introduced at this point, it will violate the "observability" principle of the control system, posing an extremely high risk of overheating or misjudgment.

[0101] The corresponding operation is to automatically trigger the fuse mechanism. This forcibly cuts off the internal fine-tuning current modulation signal (setting the duty cycle to zero) to eliminate erroneous energy injection sources. Simultaneously, it ignores feedback commands based on the visual model and switches the external temperature field control to the preset "standard open-loop control mode" (including fixed linear heating), ensuring the calibration process completes as a safety precaution.

[0102] Figure 1 It clearly demonstrates the mapping relationship between the core execution logic of the automatic calibration method and the physical architecture. Figure 1 The transformer temperature controller (the object of execution) shown on the left is driven by a dual source of external temperature field and internal current loop, and works in conjunction with machine vision equipment to perform closed-loop control.

[0103] Step 1: Current injection and parameter identification, corresponding to step S1, which is to drive the internal current excitation circuit to generate a pseudo-random current sequence and obtain thermal response characteristic parameters to establish the dynamic thermal fingerprint of the system. Step 2: Visual vector extraction, corresponding to step S2, which is to obtain the visual proximity vector representing the difference between the current position of the instrument pointer and the target position through image recognition technology, so as to realize non-contact status reading; Step 3: Multi-source fusion calculation, corresponding to step S3, that is, using a visual coupled thermal impedance feedforward observer, the thermal parameters obtained in S1 and the visual vector obtained in S2 are fused to calculate the state residual between the effective temperature of the temperature sensing bulb and the visual reading, and generate a dual-source collaborative calibration command. Step 4: Cooperative control and calibration, corresponding to step S4, that is, in response to the cooperative command, the temperature change rate of the external temperature field and the internal fine-tuning heat flow are adjusted in parallel, and the pointer is driven to approach and accurately trigger the calibration contact with a dynamic trajectory with zero hysteresis.

[0104] Figure 3 This is a schematic diagram illustrating the results of a numerical verification experiment conducted on "Dual-source collaborative calibration control command generation and execution (step S4)" in one embodiment of the present invention. This numerical verification aims to test the theoretical response characteristics of the present invention under preset standardized boundary conditions (representing a test vector where "environmental interference intensity" increases linearly from 0% to 90%). Figure 3 In the graph, the horizontal axis represents the "environmental interference intensity (%)", the left vertical axis represents the dimensionless "visual confidence index" and the normalized "fine-tuning current duty cycle", and the right vertical axis represents the "temperature change rate (°C / min)". Specifically, the blue solid line represents the calculated visual confidence index, the red solid line represents the duty cycle command of the fine-tuning current, the green stepped line represents the temperature change rate command of the external temperature field, and the vertical dashed line represents the preset degradation trigger threshold.

[0105] like Figure 3As shown, the numerical calculation results clearly reveal the deterministic logic switching process executed by the control system as environmental disturbances intensify. The "confidence-based degradation strategy" enables intrinsically safe control. Specifically, in the "cooperative control zone" (to the left of the threshold), the calculated fine-tuning current duty cycle (red) dynamically adjusts as the confidence level (blue) decreases, while the temperature change rate (green) maintains a high dynamic response. In stark contrast, once the confidence level curve falls below the preset threshold and enters the "degradation zone" (to the right of the threshold), the fine-tuning current duty cycle immediately returns to zero (cutting off the internal heat source), and the temperature change rate switches to the preset standard open-loop value (0.5℃ / min). This significant signal step change directly confirms that, with the introduction of "visual entropy flow monitoring," logic circuit breaking can be automatically executed at the critical point of visual feedback failure, effectively preventing erroneous energy injection.

[0106] Figure 5 It is a technical schematic diagram showing the equipment connection method and working principle of the temperature controller (thermometer) for thermal simulation testing.

[0107] On the far left is a power frequency current generator. The output circuit is connected to an ammeter (A) to mark the current. The range is 500mA to 5000mA. The circuit is then connected to a converter. The secondary output current of the converter is labeled as follows: The current ranges from 700mA to 1300mA, and another ammeter (A) is connected in series. The oil flows into a constant-temperature oil bath, indicated by a dashed box. The temperature of the constant-temperature oil bath is set to... The oil bath contains a heating wire (connected to the current circuit) and a sensor. The sensor extends out of the oil bath via a capillary tube. The capillary tube connects to the rightmost thermometer dial, which has scales and a pointer to display the temperature reading.

[0108] Furthermore, this verification is based on a general numerical computing environment. This environment was chosen to verify the logical closed-loop performance and numerical stability of step S4, "Visually Coupled Thermal Impedance Feedforward Observation and Dual-Source Cooperative Control," within a deterministic computing framework.

[0109] To quantify the evaluation, the actual temperature controller calibration process is abstracted into a discrete-time state-space model. The system state is represented by a state vector. Representation. This variable is strictly defined in memory as a time-series data stream, where... It is a floating-point scalar based on statistical variance normalization. The state update follows the control logic described above: ,in This is a piecewise nonlinear mapping function. When... When, perform dual-source weighted calculation; when At that time, forced execution The truncation logic. To comprehensively test the effectiveness of the solution, a standardized input set covering typical application scenarios was constructed: Baseline / Steady-State Test Vector (corresponding to a cleanroom scenario): Configure "Environmental Interference Intensity" to 0%. This vector aims to verify the algorithm's maximum collaborative gain response capability under ideal visual feedback conditions.

[0110] Boundary / stress test vector (for heavily occluded scenarios): Configure "Environmental Interference Intensity" as a linearly increasing sequence (0%). (90%), simulating the entire process of oil fume concentration crossing a critical value. This configuration constitutes the logical boundary condition for triggering the "degradation" mechanism in the verification experiment.

[0111] This verification uses the "cooperative energy efficiency ratio" "and fine-tuning current cutoff response time" are used as core evaluation indicators. Their definition logic follows: The calculation is the proportion of internal heat flux power in the total effective temperature control power; the cutoff response time is defined as... Pass through Time to Logical delay at the zero-point time. Non-zero values ​​in the high confidence interval directly characterize the system's dynamic compensation capability for thermal hysteresis; while the zero-delay characteristic of the cutoff response time theoretically directly characterizes the ability to actively avoid the risk of visual failure. Although the above verification data is based on specific... and While the parameters are set, the conclusions are universally applicable. The core advantage of this solution stems from the mathematical mapping relationship between "confidence assessment based on statistical variance" and "segmented control strategy." The effectiveness of this mechanism does not depend on specific non-critical parameters (including specific heating power or oil tank volume). In any application scenario that meets the requirements of "visual feedback and dual-source actuators," applying this technical solution can achieve robustness and safety improvements consistent with the above numerical verification results.

[0112] During contact action error measurement, the contact status is monitored in real time using a multi-channel process signal scanning device. Unlike traditional methods that rely on a constant temperature bath with a slow linear heating rate of (0.8~1.0)℃ / min, the "dual-source collaborative control" in step S4 above achieves an extremely low-speed approximation equivalent to <0.01℃ / min near the action point. This allows for the simultaneous reading of the standard thermometer value as the switching value at the moment of contact action, eliminating dynamic errors caused by thermal inertia and improving the accuracy of switching difference measurement.

[0113] Furthermore, by injecting the thermal response fingerprint of the pseudo-random current sequence active identification system and using the geometric approximation vector extracted by machine vision as the input of the state observer, a dynamic mapping relationship between the virtual effective temperature and the visual reading is constructed. By constructing a dual-source collaborative architecture of "external macroscopic temperature field + internal fine-tuning heat flow", the detection blind zone caused by traditional thermal inertia is fundamentally overcome, realizing zero-hysteresis tracking and high-fidelity state perception of the transformer temperature controller during dynamic calibration, thereby improving the intelligence level and operational efficiency of metrological calibration.

[0114] In another embodiment, to meet the national standard acceptance requirements in different scenarios, in addition to the core dynamic calibration mode mentioned above, it is also configured to perform the following standardized testing mode: static thermal simulation verification mode. In this mode, a "thermal simulation characteristic test" conforming to the standards of "JB / T6302-2016 'Transformer Oil Level Temperature Controller' and JB / T8450-2016 'Transformer Winding Temperature Controller'" is performed. A high-accuracy programmable AC power supply is controlled to output a constant heating current (740mA or 1040mA) and maintain the constant temperature bath temperature at 80℃±1℃. After the system stabilizes for a preset time of 45 minutes, the processor automatically reads the following parameters and calculates the indicated temperature rise deviation based on the following simultaneous formulas: ; ; ;in, To account for the indicated temperature rise error, The pointer value read by the camera. This is the reading from a standard platinum resistance thermometer. The standard temperature rise value obtained from the table; It is the temperature rise error indicated by the remote signal device of the temperature controller being tested, in degrees Celsius (°C). This is the correction value for a standard thermometer, expressed in degrees Celsius (°C). It is the temperature rise error indicated by the remote signal device of the temperature controller being tested, in degrees Celsius (°C). It is the temperature value converted from the output of the remote signal device of the temperature controller being measured, and the unit is degrees Celsius (°C).

[0115] The same hardware is used to perform rapid measurement of the influence of ambient temperature. The processor controls the intelligent precision thermostatic bath to automatically switch between a first temperature point (100℃±0.5℃) and a second temperature point (60℃±0.5℃). After the image recognition camera determines that the reading has stabilized, it automatically records the data and calculates the influence of ambient temperature according to the following formula. : ;in, It is the influence of the ambient temperature on the measured temperature controller, expressed as a percentage per degree Celsius (% / ℃). It is the indication error of the capillary tube of the temperature controller under high temperature conditions, and the unit is degrees Celsius (°C). It is the indication error of the capillary tube of the temperature controller under normal temperature conditions, in degrees Celsius (°C). It is the upper limit of the temperature that the temperature controller being tested can measure, in degrees Celsius (°C). It is the lower limit of the temperature measurement temperature of the temperature controller being measured, in degrees Celsius (°C). This is the high temperature value of the capillary tube, in degrees Celsius (°C). This is the capillary temperature at room temperature, expressed in degrees Celsius (°C).

[0116] The computational logic involved in this application can be constructed using algorithms such as regression analysis in machine learning, establishing a mathematical model by analyzing the inherent trends and interrelationships of the collected parameters. This process can be implemented using specialized computational tools (such as Python's Scikit-learn library or the R language environment). Throughout all calculations, to eliminate the influence of different physical dimensions and ensure that data is compared and analyzed on the same scale, the input parameters in each formula are dimensionless. The dimensionless techniques used include, but are not limited to, max-min normalization or Z-score standardization.

[0117] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this in-memory data structure.

[0118] It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. This application also provides a computer-readable storage medium having computer program instructions stored thereon.

Claims

1. An automatic calibration method for transformer temperature controllers based on image recognition, applied to an automated detection environment including an external temperature field generating device and an internal current excitation circuit, characterized in that... The specific steps include: S1: Drive the internal current excitation circuit to generate a pseudo-random current sequence and obtain thermal response characteristic parameters that characterize the dynamic characteristics of the thermal response of the temperature controller under test under the pseudo-random current excitation. S2: Obtain the visual approximation vector representing the geometric difference between the current position indicated by the instrument pointer of the temperature controller under test and the preset target position; S3: Using a pre-configured visually coupled thermal impedance feedforward observer, the thermal response characteristic parameters and the visual approximation vector are fused and calculated to solve for the state residual between the real-time effective temperature of the temperature sensor of the temperature controller under test and the visual reading. Based on this, a dual-source collaborative calibration control command is generated to characterize the real-time thermal hysteresis state and critical approximation degree of the temperature sensor of the temperature controller under test; and S4: In response to the dual-source collaborative calibration control command, the temperature change rate of the external temperature field generator and the fine-tuning heat flow injected into the internal current excitation circuit are adjusted in parallel to drive the instrument pointer to approach and trigger the calibration contact along a preset dynamic trajectory.

2. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 1, characterized in that: Step S1 includes: A pseudo-random current sequence with preset spectral characteristics is injected into the thermal simulation circuit of the temperature controller under test by a control current source. The instantaneous displacement data of the instrument pointer in response to the pseudo-random current sequence is captured at a preset frame rate using an image acquisition device; and parameter identification and calculation steps are performed. An observation vector is constructed, which includes the instantaneous displacement data of the previous time step and the pseudo-random current sequence of the current time step; an autocorrelation matrix and a cross-correlation vector are constructed based on the observation vector; matrix operations are performed on the autocorrelation matrix and the cross-correlation vector to obtain the thermal response characteristic parameters, including the thermal inertia time constant and the steady-state gain.

3. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 2, characterized in that: Step S2 includes: Obtain the target coordinate vector of the preset target position in the image coordinate system; extract the real-time coordinate vector of the instrument pointer tip in the image coordinate system; and Calculate the Euclidean distance and direction vector between the real-time coordinate vector and the target coordinate vector, and combine the Euclidean distance and direction vector into the visual approximation vector.

4. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 3, characterized in that: The visual-coupled thermal impedance feedforward observer execution in step S3 includes the following fusion calculations: Based on the thermal response characteristic parameters, a thermal impedance transfer function model describing the thermal conduction characteristics of the temperature controller under test is constructed. Using the aforementioned thermal impedance transfer function model, the virtual effective temperature of the temperature sensing bulb is estimated based on the current external temperature field and the injected current. Calculate the state residual between the virtual effective temperature and the visual temperature derived based on the visual approximation vector; Based on the state residual and the magnitude of the visual approximation vector, the dual-source collaborative calibration control command is generated, which includes a first component for adjusting the external temperature field and a second component for adjusting the internal fine-tuning heat flow.

5. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 4, characterized in that: After calculating the state residual between the virtual effective temperature and the visual temperature derived from the visual approximation vector, a dynamic thermal inertia compensation update step is also included: Obtain the thermal hysteresis decay coefficient and the preset learning rate parameter of the previous control cycle; obtain the sign feature of the virtual effective temperature estimated at the previous moment; perform parameter correction gradient operation: multiply the state residual, the sign feature of the virtual effective temperature and the learning rate parameter to obtain the correction increment; The correction increment is added to the thermal hysteresis decay coefficient of the previous control cycle to obtain the dynamic thermal hysteresis decay coefficient at the current moment, so that the estimation process of the virtual effective temperature can adaptively follow the changes in the heat conduction behavior of the temperature controller under test.

6. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 5, characterized in that: Based on the state residual and the magnitude of the visual approximation vector, the dual-source collaborative calibration control command is generated, specifically including the step of performing nonlinear weighted dual-source energy allocation: Define a continuous normalized weighting function based on the magnitude of the visual approximation vector; Using a normalized weighting function, we calculate the macroscopic weighting coefficients that characterize the priority of external temperature field regulation, and the microscopic weighting coefficients that characterize the priority of internal fine-tuning heat flow regulation. Calculate the first and second ideal adjustment quantities based on the state residuals, respectively; A weighted operation is performed on the first ideal adjustment amount and the macroscopic weight coefficient to generate the first component, and a weighted operation is simultaneously performed on the second ideal adjustment amount and the microscopic weight coefficient to generate the second component.

7. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 6, characterized in that: Step S4 includes: in response to the first component of the dual-source collaborative calibration control command, generating a temperature change rate adjustment signal for driving the external temperature field temperature control device; In response to the second component of the dual-source collaborative calibration control command, a fine-tuning current modulation signal for driving the internal current source is generated; and when the confidence level of the visual approximation vector is lower than a preset threshold, a degradation strategy is automatically executed, the degradation strategy including cutting off the fine-tuning current modulation signal and switching the temperature rate adjustment signal to a preset standard open-loop control mode.

8. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 7, characterized in that: The process for determining when the confidence level of the visual proximity vector is lower than a preset threshold includes: establishing a sliding window with a preset time length; extracting the Euclidean distance values ​​of the visual proximity vectors at all historical moments within the sliding window to form a sequence of Euclidean distance values; performing statistical variance calculation on the sequence of Euclidean distance values ​​and normalizing the reciprocal of the calculation result to a visual confidence index; and comparing the visual confidence index with a preset downgrade trigger threshold.

9. The automatic calibration method for transformer temperature controllers based on image recognition according to claim 8, characterized in that: The process of generating a fine-tuning current modulation signal for driving an internal current source includes: obtaining a preset current modulation period parameter; obtaining the second component in the dual-source collaborative calibration control command, and calculating a target duty cycle based on the ratio of the second component to the preset maximum output fine-tuning current; and generating a high-frequency on / off control sequence using the target duty cycle within the current modulation period as the fine-tuning current modulation signal.

Images