Surface temperature measurement correction method based on thermal simulation and BP neural network

Abstract

The application discloses a surface temperature measurement correction method based on thermal simulation and BP neural network, belongs to the technical field of contact type temperature measurement and temperature measurement error compensation, and aims at solving the problem of systematic deviation of a surface temperature measurement caused by coupling of conduction, convection, radiation, contact state and sensor thermal response in a complex field. The application trains a multi-output BP neural network by using transient conjugate heat transfer simulation data, traces test calibration error components and working condition weights, synthesizes total compensation and outputs a dominant error type and a reliable coverage interval, and according to an online calibration criterion, implements amplitude limiting update or parameter freezing, so that the technical effect of improving surface temperature measurement accuracy, repeatability and compensation stability in a cross-temperature zone and a cross-environment is achieved.

Description

Technical Field

[0001] This invention relates to the field of contact temperature measurement and temperature measurement error compensation, and in particular to a surface temperature measurement correction method based on thermal simulation and BP neural network. Background Technology

[0002] Contact surface temperature sensors are widely used in the operation monitoring of cryogenic storage and transportation equipment, petrochemical plants, pressure vessels, and high-temperature pipelines. In actual measurements, the contact heat transfer between the sensor and the measured object, the convective heat dissipation from the sensor package to the surrounding fluid, the radiative heat transfer with the surrounding surface, and the thermal inertia of the sensing element all affect the original temperature reading. Moreover, these effects vary with ambient temperature, fluid state, contact pressure, material thermal properties, and structural dimensions.

[0003] Existing surface temperature measurement methods mostly rely on static calibration coefficients in the laboratory or establish empirical correction relationships for a single temperature zone and a single installation condition. These methods struggle to cover the combined variations in temperature difference direction, convection intensity, radiation boundary, and contact condition within the range of -60 to 600°C. When training a single total error model based solely on limited measured samples, various heat transfer errors lack identifiable labels, making it difficult for the model to output the dominant error sources. Furthermore, using only heat transfer simulation results may result in systematic residuals between the simulation domain and the measured domain due to material parameters, contact conditions, and deviations from on-site boundaries.

[0004] Furthermore, continuous online adjustment of all parameters of a neural network is susceptible to compensation drift due to erroneous reference values, out-of-domain operating conditions, and long-term cumulative updates. Existing technologies typically do not simultaneously define the traceability of the reference temperature, the width of the reliable coverage interval, the component consistency residual, the update amplitude, and the offline parameter backoff conditions, making it difficult to provide a reliable boundary and maintain online compensation stability while outputting the corrected temperature.

[0005] Therefore, there is a need for a surface temperature measurement correction method based on thermal simulation and BP neural network that can overcome the shortcomings of the existing technology. Summary of the Invention

[0006] One objective of this invention is to propose a surface temperature measurement correction method based on thermal simulation and a BP neural network. Addressing the problems in existing technologies, such as the difficulty in handling the coupling of conduction, convection, radiation, contact, and dynamic hysteresis in static calibration, the lack of a unified calibration chain between simulation error components and the total measured error, and the susceptibility to compensation drift during online learning, this invention proposes a technical solution that collects surface temperature measurement operating condition characteristics, uses transient conjugate heat transfer simulation data to form component labels and trains a multi-output BP neural network, uses traceable reference surface temperature test residuals to calibrate output parameters and operating condition weights, synthesizes the total compensation amount, dominant error type, and reliable coverage interval, and performs limited updates or freezes on parameters based on online calibration criteria. This invention improves the consistency of temperature measurement correction across temperature zones and environments, identifies dominant errors, and suppresses online compensation drift.

[0007] This invention provides a surface temperature measurement correction method based on thermal simulation and BP neural network, including: S1. Collect the raw temperature readings and their rate of change over time from the contact surface temperature sensor, along with the ambient temperature, local convection state, surrounding radiation temperature, contact state, and structural parameters of the sensor and the measured object, to form a working condition feature vector. S2. Input the operating condition feature vector into a multi-output BP neural network trained based on transient conjugate heat transfer simulation data and calibrated with traceable reference surface temperature test data, and output contact and heat conduction, convection, radiation and dynamic hysteresis error components and corresponding non-negative operating condition weights. S3. Synthesize each error component according to the non-negative working condition weight to obtain the total compensation amount and the contribution amount of each error component. Add the total compensation amount to the original temperature indication to obtain the corrected surface temperature. S4. Determine the dominant error type by sorting the absolute values ​​of the contribution, and determine the reliable coverage range of the corrected surface temperature based on the measured residual distribution of the working condition matching. S5. When a traceable reference surface temperature is obtained, an online calibration criterion is set based on the traceable reference surface temperature, the corrected surface temperature, the reliable coverage range, and the total compensation amount; if the temperature is not obtained, the criterion is set to not meet the preset update conditions. S6. According to the criterion, if the preset update conditions are met, the output parameters of the amplitude-limited update error component or the non-negative operating condition weight are updated; if not, the parameters are frozen.

[0008] Optionally, S1 includes: obtaining the original temperature reading by the sensor acquisition circuit and determining the rate of change of time based on the difference between the readings at adjacent sampling times and the sampling time interval; The ambient temperature is obtained by an ambient temperature sensor; the local convection state, which characterizes at least one of fluid temperature, flow velocity, and convective heat transfer coefficient, is obtained by a flow velocity sensor, a fluid temperature sensor, or heat transfer calculation based on flow boundary conditions; and the ambient radiation temperature is obtained by a radiation temperature sensor or by calculation of ambient surface temperature and emissivity. The contact state is characterized by at least one of contact pressure, contact area and contact thermal resistance, and the sensor package size, temperature sensing element position, material thermal conductivity and heat capacity, as well as the thickness, thermal conductivity and heat capacity of the measured body are used as the material and structural parameters. The parameters are normalized according to the value range of each parameter in the offline training samples to form the working condition feature vector.

[0009] Optionally, S2 includes: the transient conjugate heat transfer simulation data is obtained by: establishing a finite element transient conjugate heat transfer model that includes the measured body, contact layer, sensor package, temperature sensing element, surrounding fluid and radiation boundary, and combining and setting the ambient temperature, fluid state, radiation boundary, contact state and material and structural parameters within the measured surface temperature range of -60 to 600℃. The undisturbed surface temperature without the sensor and the temperature sensing element temperature after the sensor is installed are calculated separately, and the difference between the undisturbed surface temperature and the temperature sensing element temperature is used as the total error label. Extract the heat flow conducted at the contact boundary, the heat flow convected on the sensor surface, the heat flow radiated on the sensor surface, and the transient heat storage of the temperature sensing element. Calculate and normalize the equivalent temperature difference corresponding to each heat flow according to the preset boundary heat flow separation order to form component labels for the contact and heat conduction error components, convection error components, radiation error components, and dynamic hysteresis error components. Make the composite value of each component label correspond to the total error label. Furthermore, the multi-output BP neural network includes a shared hidden layer that receives the working condition feature vector, four error component output terminals that output four error components respectively, and a working condition weight module that outputs the non-negative working condition weight. The operating condition weighting module generates at least four non-negative numbers based on the difference between the original temperature reading and the ambient temperature, the time change rate, the convective heat transfer state quantity obtained by normalizing the local convection state, the radiative heat transfer state quantity obtained by normalizing the ambient radiation temperature, and the contact heat transfer state quantity obtained by normalizing the contact state, and normalizes the sum of the four non-negative numbers to one. The training loss includes the component loss between each error component and its corresponding component label, the consistency loss between the error synthesized by each error component according to the non-negative operating condition weight and the total error label, and the sign constraint loss for violating the corresponding boundary heat flow direction. Furthermore, the traceable reference surface temperature test data is divided into multiple operating condition units according to the temperature zone, the direction of the temperature difference between the original temperature reading and the ambient temperature, the convective heat transfer coefficient range, and the contact state range. Within each operating unit, the difference between the traceable reference surface temperature and the synchronously acquired original temperature reading is taken as the total measured error. The total compensation amount is obtained by summing each error component according to the non-negative operating weight. The difference between the total measured error and the total compensation amount is taken as the paired residual. While keeping the shared hidden layer unchanged, offline calibration parameters corresponding to each operating condition unit are obtained by considering at least one of the bias, scale, and non-negative operating condition weights at the output of each error component of the paired residual calibration.

[0010] Optionally, S3 includes: applying the bias and scale corresponding to the current operating condition unit to each error component, multiplying the calibrated error component by the corresponding non-negative operating condition weight to obtain the contribution of the error component, and summing the contribution of the four error components to obtain the total compensation amount. When the operating condition feature vector is located in the boundary region of multiple adjacent operating condition units, the offline calibration parameters of each adjacent operating condition unit are interpolated based on the normalized distance from the operating condition feature vector to the center of each adjacent operating condition unit, and the total compensation amount is calculated using the interpolated parameters.

[0011] Optionally, S4 includes: arranging each contribution in descending order of absolute value, and determining the error component corresponding to the contribution at the top of the list as the dominant error type; The residual samples that were not used to obtain the offline calibration parameters are reserved from the traceable reference surface temperature test data of each working unit. The empirical quantile of the absolute value of the residual sample is determined according to the preset target coverage probability. The reliable coverage interval is obtained by subtracting and adding the empirical quantile from the corrected surface temperature. When the current operating condition falls within the applicable range of two or more adjacent operating condition units, the empirical quantiles of the adjacent operating condition units are arranged in descending order of value, and the reliable coverage interval is expanded using the empirical quantile at the top of the list.

[0012] Optionally, S5 includes: verifying the calibration certificate identifier, validity period, and measurement uncertainty of the traceable reference surface temperature to obtain the reference traceability status; The difference between the upper and lower limits of the reliable coverage interval is calculated to obtain the width of the coverage interval; Calculate the difference between the traceable reference surface temperature and the original temperature reading, and subtract the total compensation amount to obtain the component consistency residual; The online calibration criterion is composed of the reference traceability status, the coverage interval width, and the component consistency residual.

[0013] Optionally, S6 includes setting the following as the preset update conditions: the reference traceability status is valid, the coverage interval width is not greater than the width threshold determined by the offline verification data, the absolute value of the component consistency residual is not greater than the consistency threshold determined by the offline verification data, and the operating condition feature vector falls within the applicable range of only one offline calibration parameter. When the preset update conditions are met, the output bias, scale or non-negative operating condition weights of the error components associated with the applicable range are updated with the goal of reducing the component consistency residual, and the update amount of each update is limited to the corresponding parameter update range determined by offline verification data. The update stops when the preset update conditions are not met, and the offline calibration parameter that has been called before the current sampling time and whose applicable range covers the current working condition continues to be called, while maintaining the output of the corrected surface temperature and the reliable coverage range. Furthermore, a version identifier, applicable range of operating condition feature vector, and verification time are set for each offline calibration parameter version; When the operating condition feature vector exceeds the applicable range of the current version, the preset number of samplings determined by offline verification data is not satisfied with the preset update condition, or the online calibration criterion changes from satisfying the preset update condition to not satisfying the preset update condition after calling the online update parameters, the error component output parameters or non-negative operating condition weights after online update are cancelled, and the process is rolled back and the offline calibration parameters that have passed offline verification the most recently and whose applicable range covers the current operating condition are called according to the version identifier. If there are no offline calibration parameters that cover the current operating condition, freeze the parameters, first output a status flag indicating that the current operating condition exceeds the trusted coverage range, and stop outputting the interval calculated in step S4 as the trusted coverage range.

[0014] The beneficial effects of this invention are: 1. The contact and conduction, convection, radiation and dynamic hysteresis error components are labeled by the transient conjugate heat transfer model and synthesized by a multi-output BP neural network combined with non-negative operating condition weights, so that the total compensation can be adjusted with the heat transfer state. At the same time, the dominant error type can be determined by the component contribution.

[0015] 2. By calibrating the output bias, scale, or operating condition weight with the paired residual between the traceable measured total error and the total compensation within the operating condition unit, the system difference between the simulation boundary and the field conditions can be corrected, and a reliable coverage range for the corrected surface temperature can be given by leaving the empirical quantile of the residual.

[0016] 3. By controlling online updates jointly with the reference traceability status, coverage interval width, component consistency residuals, and the applicable range of offline parameters, the update range is limited, and parameters are frozen, canceled, or rolled back when conditions are not met or operating conditions exceed limits. This can reduce compensation drift caused by erroneous references, out-of-domain operating conditions, and long-term online updates. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the surface temperature measurement correction method based on thermal simulation and BP neural network of the present invention.

[0018] Figure 2 This is a flowchart of the online calibration and parameter management steps S6 of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A surface temperature measurement correction method based on thermal simulation and BP neural network includes: S1. Collect the raw temperature readings and their rate of change over time from the contact surface temperature sensor, along with the ambient temperature, local convection state, surrounding radiation temperature, contact state, and structural parameters of the sensor and the measured object, to form a working condition feature vector. S2. Input the operating condition feature vector into a multi-output BP neural network trained based on transient conjugate heat transfer simulation data and calibrated with traceable reference surface temperature test data, and output contact and heat conduction, convection, radiation and dynamic hysteresis error components and corresponding non-negative operating condition weights. S3. Synthesize each error component according to the non-negative working condition weight to obtain the total compensation amount and the contribution amount of each error component. Add the total compensation amount to the original temperature indication to obtain the corrected surface temperature. S4. Determine the dominant error type by sorting the absolute values ​​of the contribution, and determine the reliable coverage range of the corrected surface temperature based on the measured residual distribution of the working condition matching. S5. When a traceable reference surface temperature is obtained, an online calibration criterion is set based on the traceable reference surface temperature, the corrected surface temperature, the reliable coverage range, and the total compensation amount; if the temperature is not obtained, the criterion is set to not meet the preset update conditions. S6. According to the criterion, if the preset update conditions are met, the output parameters of the amplitude-limited update error component or the non-negative operating condition weight are updated; if not, the parameters are frozen.

[0021] In this specific embodiment, S1 includes: The data acquisition and correction device establishes a working condition record using sampling time identifier, sensor channel identifier, and measured object identifier. The acquisition circuit of the contact surface temperature sensor reads the raw temperature reading at a 1-second sampling cycle. The ambient temperature sensor, flow rate sensor, fluid temperature sensor, radiation temperature sensor, and contact pressure acquisition channel are triggered by the same clock. The data of each channel are first aligned within a 0.2-second tolerance according to the timestamp, and then single-point pulses are eliminated through range check and three-point median filtering. If any necessary channel is missing, the working condition record is marked as invalid and network inference for the current sampling cycle is stopped. The rate of temperature change is calculated based on the difference between the original temperature readings at adjacent valid sampling times. , The rate of change of the original temperature reading over time, expressed in °C / s. This indicates the raw temperature reading at the current sampling time, in °C. This represents the original temperature reading at the previous valid sampling time, in °C. This indicates the current sampling time, in seconds. This indicates the previous valid sampling time, in seconds. When the sampling interval exceeds 0.8–1.2 seconds, the difference result is not used and an interval anomaly flag is written. When the most recent valid rate of change is no more than 3 seconds from the current time and has been reused no more than twice, it can be used as a temporary input with an anomaly flag to continue the inference for this cycle. However, this flag is passed to S5 and S6 along with the operating condition record and the online update condition is forcibly set to unmet, freezing the parameters for this cycle. When the time age exceeds 3 seconds or has been reused more than twice, the operating condition record is set to invalid, the network inference and reliable correction for this cycle are stopped, only the previous valid corrected surface temperature is retained as an unreliable diagnostic value, and the reason for the sampling interval exceeding the limit is passed to S5 and S6 to freeze the execution parameters. Local convection state records include fluid temperature, flow velocity, and convective heat transfer coefficient. Fluid temperature and flow velocity are obtained synchronously from the corresponding sensors, while the convective heat transfer coefficient is obtained from the wind tunnel calibration curves of the same batch of test subjects according to the fluid medium type, flow velocity range, and sensor installation direction. The calibration curves record flow velocities of 0–20 m / s and convective heat transfer coefficients of 2–300 W / (m²). 2• K), curve version and validity date; when the flow rate exceeds the calibration range, the flow rate and convective heat transfer coefficient sent to the network are truncated to the corresponding boundary. At the same time, the original flow rate, flow direction, the over-limit range of the relative boundary, the calibration curve version and the extrapolation flag are written into the operating condition record. Only when S4 hits the same medium, same flow direction, same parameter version and covers the extrapolation residual library of the original flow rate, is it allowed to form a reliable coverage range of the extrapolation operating condition. Otherwise, the operating condition is marked as unreliable. The ambient radiation temperature is preferentially read from the radiation temperature sensor calibrated by the blackbody furnace. When this channel is not configured, it is calculated based on the absolute temperature and emissivity of each surrounding surface within the field of view. , This represents the equivalent radiant temperature of the surrounding area, expressed in Kelvin (K). Indicates the index of the surrounding surface. This represents the number of surrounding surfaces within the field of view, and is a positive integer. Indicates the first The viewing angle coefficients of the surrounding surfaces to the outer surface of the sensor are dimensionless and the sum of the viewing angle coefficients is 1. They are calculated by the radiation preprocessing model bound to the sensor mounting geometry and verified by the viewing angle coefficients. Indicates the first The emissivity of the surrounding surface is a dimensionless number from 0 to 1, taken from the emissivity parameter table in the batch measurement record or material test report of the same batch after the emissivity measuring instrument has been calibrated, and is kept with the material batch and expiration date; Indicates the first The absolute temperature of the surrounding surface, in K, is obtained by adding 273.15 to the Celsius temperature acquired by the corresponding surface temperature channel within a 0.2s synchronization tolerance. Only channels with valid calibration certificates and whose range covers the current temperature are used. If any source version is missing, expired, or inconsistent with the current geometry and material version, the radiation status will be marked as unavailable. The correction for this cycle will also be stopped if the denominator is less than 0.05. The contact state is defined by contact pressure, effective contact area, and contact thermal resistance. The contact pressure is measured by a thin-film pressure sensor in the range of 0–2 MPa. The effective contact area is calibrated by the indentation on the sensor mounting surface. The contact thermal resistance is obtained by dividing the temperature difference at the contact interface by the heat flux density through the interface at three steady-state temperature difference points under the same mounting structure. The material and structural parameters record the sensor package length, width, thickness, depth of the temperature sensing element from the contact surface, thermal conductivity and volumetric heat capacity of each layer, as well as the thickness, thermal conductivity and volumetric heat capacity of the measured object. The parameter values ​​are taken from the material test report with batch number or the test results of the same batch of samples, and are consistent with the material version of the finite element model. Perform the operation on each continuous feature according to the range of offline training samples. It is truncated to 0-1, while retaining the uncrunted normalized value obtained from the same formula before the truncation is performed; Indicates the truncated first... The normalized chemical condition characteristics are dimensionless numbers. Indicates the working condition characteristic index. Indicates the first The original operating condition parameters should be in units consistent with the units used to collect these parameters. Indicates the first offline training sample The lower limit of the parameter, Indicates the first offline training sample The upper limit of the parameter is consistent with the unit of the original parameter and is given by the training data version table; when any untruncated normalized value is less than 0 or greater than 1, the training range out-of-bounds flag and out-of-bounds direction are written for the corresponding continuous feature; when the upper and lower limits are equal, the term is set to 0 and written to the constant feature flag; the category feature is encoded one-hot using a fixed dictionary, and categories outside the dictionary are entered into the unknown category position. Before being put into use, each sensing channel undergoes zero-point and span calibration. The temperature channel is fitted with a correction coefficient using no less than three temperature points. The pressure, flow rate, and radiation temperature channels are respectively calibrated using certified standards. The operating condition record saves the calibration coefficient, calibration date, standard certificate identification, and allowable drift. When the sampled value exceeds the hardware range, calibration validity period, or allowable drift, the corresponding feature is set to missing. Abnormal measurements after truncation are not treated as normal network inputs. The field dictionary of the working condition feature vector is fixed according to the model version. The field dictionary records the Chinese name, index, original unit, lower normalization limit, upper normalization limit, missing value rule and data source of each feature. When the model is loaded, the hash value of the field dictionary is compared with the network input definition. If the hash value is inconsistent, the inference is rejected and the most recent model version with consistent fields is called, thereby preventing feature misalignment caused by changes in material parameters or sensor channel order. The acquisition module assembles a working condition feature vector by sequentially using fixed fields such as the original temperature reading, time rate of change, ambient temperature, fluid temperature, flow rate, convective heat transfer coefficient, ambient radiation temperature, contact pressure, contact area, contact thermal resistance, sensor packaging and temperature sensing element structural parameters, sensor material thermal properties, and the thermal properties of the measured material. The module then writes the truncated normalized vector, the untruncated normalized vector, the unnormalized physical quantity, dimensions, timestamp, parameter table version, feature-by-feature training range out-of-bounds flag, extrapolation flag, interval anomaly flag, and validity flag into the working condition feature record. Valid working condition feature records serve as network input for S2, while invalid records retain the previous valid corrected surface temperature and output an incomplete data state.

[0022] In this specific embodiment, S2 includes: The model training module reads the feature field definitions, material batch parameters, and boundary condition versions used by S1 to establish a three-dimensional finite element transient conjugate heat transfer model that includes the measured body, contact layer, sensor package, temperature sensing element, surrounding fluid domain, and radiation boundary. The mesh is refined to 0.05–0.2 mm around the contact layer and temperature sensing element, and the time step is adaptively selected in the range of 0.01–0.5 s. The mesh independence criterion is that the temperature peak difference of the temperature sensing element under two consecutive mesh levels does not exceed 0.2 °C. Transient solid domain according to Please solve. This represents the volumetric heat capacity of a material, expressed in J / (m³). 3 ·K), This indicates the temperature of the solid domain, expressed in Kelvin (K). The simulation time is expressed in seconds. This represents the first-order partial derivative of the solid domain temperature with respect to simulation time, expressed in K / s. The thermal conductivity of a material is expressed in W / (m·K). This represents the volumetric heat source intensity, expressed in W / m³. 3 For non-self-heating test objects and passive sensor regions, Set to 0; when there is electric heating or internal heating, The value obtained by dividing the non-negative heating power measured by the calibrated power channel by the corresponding heating region volume in the finite element model is taken, and the power channel calibration version, heating region geometry version and applicable time are written into the boundary condition table. The simulation sample is invalid when the source is missing or the power is negative. The differential operator represents the gradient or divergence with respect to spatial coordinates; the fluid domain adopts the incompressible energy equation and couples the temperature and heat flux continuity conditions on the outer surface of the sensor; the net radiative heat flux of the radiation boundary is calculated using the emissivity and viewing angle coefficient of each surface. The simulation operates within the undisturbed surface temperature range of -60 to 600℃. Latin hypercube combinations are performed based on ambient temperature, fluid temperature, flow velocity, radiation temperature, contact pressure, contact area, contact thermal resistance, package dimensions, and the thermal properties of the materials on both sides. For each combination, a baseline model without the sensor and a coupled model with the sensor installed are solved separately. The difference between the undisturbed surface temperature and the temperature sensing element temperature at any given moment is recorded as follows: , This indicates the total error label, in degrees Celsius (°C). This indicates the undisturbed surface temperature at the measuring point when no sensor is installed, in °C. This indicates the temperature of the sensing element after the sensor is installed, in °C. The post-processing module extracts the heat terms in a fixed order: heat flow conducted at the contact boundary, heat flow convectively from the sensor surface, heat flow radiated from the sensor surface, and transient heat storage of the temperature sensing element. It then uses the local temperature sensitivity calculated by the benchmark model under the same operating conditions to convert the four heat terms into initial equivalent temperature differences. The keys for sensitivity recording include temperature zone, material group of the measured object, packaging structure version, and contact pressure range. The value fields include thermal resistance sensitivity in K / W units, effective range, and simulation version. If the key is not hit, the adjacent parameter conditions are rerun instead of using cross-material extrapolation. The post-processing module first obtains the reference contribution through a minimum deviation decomposition with signed constraints, specifically: The constraints are as well as ; This represents a vector consisting of four reference contributions, with the vector elements in degrees Celsius (°C). This means selecting the candidate vector that minimizes the objective function among those that satisfy the constraints. Indicates the first Reference contribution of thermal circuits, in °C. This indicates the first term converted from the corresponding directional heat term. The initial equivalent temperature difference is expressed in °C. This represents the error type index, with values ​​of 1, 2, 3, and 4 corresponding to contact and conduction, convection, radiation, and dynamic hysteresis, respectively. Indicates the first The sign of the direction determined by the heat flow at the class boundary is either 1 or -1, and is a dimensionless number. The total error label mentioned above in this step is expressed in °C; the equality constraint makes the sum of the four reference contributions equal to the total error label. When the constraint decomposition has no feasible solution, the solver fails to converge, or the absolute value of the sum of the four initial equivalent temperature differences is less than 0.01℃ while the absolute value of the total error is not less than 0.02℃, the finite element time step is successively reduced to half of the previous value and the heat term is re-extracted, with a maximum of 3 recalculations. Samples that still fail to obtain a feasible solution after 3 recalculations are excluded from the training, validation, and test datasets, and the working condition combination identifier, number of failures, final time step, and reason for exclusion are written into the abnormal sample table. For near-zero samples with a total error absolute value of less than 0.02℃, the constraint decomposition is still performed, but labels are not generated with a zero denominator ratio. Based on the reference contribution, construct reference operating condition weights and unweighted component labels, specifically as follows: , as well as ; Indicates the first The smoothed amplitude of the reference contribution, in °C. The aforementioned first Reference contribution of thermal circuits, in °C. This indicates taking the absolute value. To prevent smoothing temperature differences with a reference weight of zero, this specific implementation uses 0.01℃. Indicates the first The reference operating condition weights for class errors are dimensionless numbers greater than 0 and whose sum of the four terms is 1. This represents the weighted normalized summation index. This represents the unweighted component label used for supervised learning at the error component output, expressed in °C; hence, This ensures that the composite value of each component label corresponds to the total error label, and that the training label relationship is consistent with the weighted composite relationship in the inference phase. The multi-output backpropagation neural network receives the operating condition feature vector S1 and employs three shared fully connected hidden layers with 64, 32, and 16 neurons respectively. The hidden layers use hyperbolic tangent activation and are applied with a 0.05 modulus during training. Weight decay; the shared hidden layer connects four linear error component output terminals and a four-channel operating condition weight module. The error component output terminals output four predicted error components in °C. The operating condition weight module receives at least the normalized temperature difference between the original temperature reading and the ambient temperature, the normalized time change rate, the convective heat transfer state quantity, the radiative heat transfer state quantity, and the contact heat transfer state quantity. The working condition weighting module for the first Execution of unconstrained output from each channel and ,in Represents the natural logarithm function. This represents an exponential function with the natural constant as its base. Indicates the first The non-negative intermediate weights corresponding to the class errors are dimensionless numbers. Indicates the first The network output of each weighted channel is a dimensionless number. Indicates the first The non-negative operating condition weights for class-type errors are dimensionless numbers, and the sum of the four weights is 1. The meaning is the same as the error type index in the component label; each inference outputs the four unconstrained channels at the same sampling time. Four non-negative intermediate weights Four normalized weights The four prediction error components, sampling time identifier, working condition record identifier, model version, and field dictionary version are jointly written into the S2 inference output record, and this record is used as the direct input of S3 and the read-only input for S6 historical sample playback; normalized weights are not allowed. Back-reaming unconstrained output ; denominator is less than At that time, all four weights were set to 0.25 and the weight degradation flag was recorded; Training loss is defined as , This represents the total training loss of the network, which is the dimensionless normalized loss. This represents the component supervision loss that simultaneously directly supervises the predicted values ​​of four error components and the weights of four operating conditions. This represents the mean square consistency loss between the combined four prediction error components (weighted according to non-negative operating conditions) and the total error label. This represents the square hinge loss when the predicted component direction violates the corresponding boundary heat flow direction. and All error terms with the dimension of squared temperature difference are divided by the total error label variance in the training set. Dimensionless This represents the variance of the total error labels in the training set, in °C. 2 Offline verification uses independent operating condition samples covering all temperature zones and four error directions to determine the positive lower limit of variance; in this specific implementation, 0.01℃ is used. 2 The variance threshold record saves the lower limit, sample size, data version, applicable temperature zone, and calculation date; only when... Not lower than the stated 0.01℃ 2 The loss is calculated only when the positive lower bound and the variance threshold record is consistent with the training data version. If the loss is below the positive lower bound, equal to 0, or the record version is inconsistent, training is stopped, the current dataset is marked as insufficient variance, and the training set is re-divided after expanding the working condition samples of different temperature zones and boundary conditions. The calculation is not continued with a denominator of 0 or near zero. The dimensionless weight representing the consistency loss is determined by a grid search of the validation set; The dimensionless weights representing the symbolic constraint loss are determined by a grid search on the same validation set; Consistency loss by calculate, This represents the dimensionless mean square consistency loss between the total compensation and the total error label. This represents the number of samples in the current training batch, and is a positive integer. This represents the sample index of the training batch, ranging from 1 to... , This represents the error type index, which can be 1, 2, 3, or 4. Indicates the first The first sample The non-negative operating condition weights for class-type errors are dimensionless numbers. Indicates the first The first sample Class prediction error components, in °C. Indicates the first The total error label for each sample is obtained from the finite element simulation in this step, in °C. This represents the variance of the total error labels in the training set, in °C. 2 ; Component monitoring loss , as well as calculate; This represents the direct monitoring term of the error component, which is a dimensionless number. This represents the direct monitoring item for the operating condition weights, which is a dimensionless number. This represents the number of samples in the current training batch, and is a positive integer. Indicates the batch sample index. Indicates the first The first sample Class prediction error components, in °C. This represents the first sample obtained by decomposing it according to the aforementioned constraints. Unweighted component labels, in °C. This indicates the weight module's output for this sample. The weights for different operating conditions are dimensionless. This indicates that the first sample is constructed from the reference contribution. The reference working condition weights are dimensionless numbers. The dimensionless coefficients representing the weights of the direct supervision terms are determined by a grid search of the validation set; thus, the reference load weights are directly incorporated into the network training objective. Square hinge loss according to calculate, This represents the number of samples in the current training batch, and is a positive integer. This represents the sample index in the training batch. Indicates the first The first sample The sign of the direction of heat flow at the boundary is 1 or -1, and is a dimensionless number. Indicates the network's response to the first... The first sample Predicted values ​​of error components, in °C. This indicates a comparison between 0 and real numbers. And select the value that is not less than another value. This represents the real number input to the function, with units of AND and DOWN. The same; therefore, the square of the quantity inside the square brackets is the same as... Both are ℃ 2 ; The simulation samples were divided into a 70% training set, a 15% validation set, and a 15% test set according to the finite element working condition combination identifier. All moments of the same time process were included in only one set. The training used an Adam optimizer with a batch size of 128 and an initial learning rate of 0.001. Training was stopped when the validation loss did not decrease for 20 consecutive rounds. The parameters of the shared hidden layer with the lowest validation loss, the error component output, and the weight module were saved. The test set reported the root mean square error of the four components, the root mean square error of the total compensation, and the boundary sign violation rate, respectively. The traceable reference surface temperature test is conducted in a calibrated constant temperature bath, dry block furnace, or controlled heating platform. The calibration certificate identification, validity period, and expanded uncertainty of the reference sensor are saved with the sample. The test data are organized into operating condition units according to temperature zone, the direction of temperature difference between the original temperature reading and the ambient temperature, the convective heat transfer coefficient range, and the contact state range. The unit key, characteristic boundary, center vector, number of samples, residual sample version, and validity date are written into the operating condition unit table. Each unit retains at least 30 synchronous stable samples, and the training calibration samples and the coverage range verification samples are separated according to the test round. Within each operating condition unit, the total measured error is calculated as follows: Calculate the paired residuals according to calculate, This represents the total measured error, expressed in °C. This indicates the traceable reference surface temperature, in °C. This indicates the synchronized original temperature reading, in °C. This represents the paired residual, in °C. Indicates the first The non-negative operating condition weights for class-type errors are dimensionless numbers. Indicates the first The predicted value at the output of each error component, in °C. This indicates four error types; samples with a synchronization time difference exceeding 0.2s or a reference measurement uncertainty exceeding the allowable value for this temperature range are not included in the calibration. Keeping the shared hidden layer unchanged, the scale and bias of the four error component outputs are fitted for each working condition unit, and the bias of the weight module output is calibrated when necessary. The fitting adopts constrained least squares with a ridge penalty of 0.01, so that each scale is between 0.5 and 1.5 and each bias is between -5 and 5℃. Simulation verification samples that did not participate in network training in the same working condition unit are also input into the calibration process. The weight bias constraint term is the mean square error between the calibrated weights and the corresponding reference working condition weights, which is 0.05 times. Calibration is only accepted when the median absolute value of the measured paired residual decreases, the root mean square error of the measured total compensation of the holdout set does not increase, the root mean square errors of the four types of components of the simulation holdout set do not increase, and the mean absolute error of the working condition weights does not increase. The decomposition verification threshold is determined by an independent simulation verification set that does not overlap with the training samples and calibration samples before the first offline parent version is frozen: the root mean square error of the four types of components and the mean absolute error of the working condition weights of the first offline parent version are calculated separately for temperature zone and working condition unit. Each threshold is taken as 1.05 times the corresponding statistic. The threshold units for the four categories are in °C, and the weight thresholds are dimensionless. The threshold record saves the four categories of values, weight values, independent validation set version, number of samples, temperature zone, operating condition unit key, generation date, and the identifier of the first offline parent version, and remains unchanged within the same parent version chain. The re-verification cycle of offline calibration parameters is determined by the equipment calibration and material batch re-verification plan. In this specific implementation, it is 12 months. The verification time and verification expiration time are written when the verification is completed. If any of the material batch, sensor packaging or temperature sensing element structure, feature field dictionary, or corresponding residual library generation changes, the original record is immediately invalidated without waiting for the expiration time. The offline calibration parameter record is composed of the verified network version, operating condition unit key, four sets of scales and biases, weight calibration parameters, four categories of component errors, weight errors, verification time, verification expiration time, re-verification cycle source, and applicable scope. It is output to S3 along with the current four error components and non-negative operating condition weights.

[0023] In this specific embodiment, S3 includes: The compensation calculation module reads the four prediction error components and the four unconstrained weight channels from the S2 inference output record. The system reads four non-negative operating condition weights, their sampling time identifiers, and model versions, while simultaneously reading the operating condition unit table, offline calibration parameter records, and the currently active online sub-version pointer maintained by S6. If the sampling time identifier, operating condition record identifier, model version, or field dictionary version of the S2 inference output record are inconsistent with any of the following, the compensation for this cycle is stopped, and no other cycles are used. Replacement; when consistent, the applicable working condition unit is located using the S1 working condition feature vector; before location, unnormalized physical quantities, untruncated normalized values, and training range out-of-bounds flags are read item by item, and the training range and the applicable range of the current parameter version are verified first; except for the velocity branch with a qualified extrapolation residual library and a clear verification path as described in S1, if any continuous feature goes out of bounds, the reliable compensation calculation for this cycle is stopped, the most recent valid correction value is retained as the unreliable diagnostic value, written into the out-of-bounds state, and this state is passed to S4, S5, and S6; only after all features have passed the verification are the truncated values ​​compared item by item. Normalization features and upper and lower bounds of the operating condition unit; when a unit is hit, the current active online sub-version that has the same operating condition unit key, has been submitted by S6, has an effective sampling period no later than the current period, has not expired, and whose parent offline version is still valid is read first, and the scale, bias, and weight output calibration bias of the sub-version and its bound residual library and threshold record are called; only when there is no qualified active online sub-version, the latest offline calibration parameters of the unit that have been verified and have not expired are called; when no unit is hit, no online extrapolation is performed, the most recent valid correction value is retained and written to the state of being out of applicable range; For the The prediction error component is obtained by applying the scale and bias of the current operating cell. , Indicates the first Error components after calibration, in °C. This indicates the first operating condition unit. The scale for the error components is a dimensionless number ranging from 0.5 to 1.5. This indicates the first error component output given by the S2 error component output terminal. Class prediction error components, in °C. This indicates the first operating condition unit. The bias of the error components, in °C and ranging from -5 to 5 °C. Indicates four error type indices; scale and bias are both taken from parameter records formed by S2 and verified offline; When the characteristic vector of a working condition falls into the boundary region of two or more adjacent working condition units simultaneously, according to Calculate the normalized distance to the center of each adjacent cell. Indicates the current operating condition up to the [number]th [number]. The Euclidean distance between the centers of adjacent working condition units is a dimensionless number. Indicates the adjacent working condition cell index, ranging from 1 to... positive integers, This indicates the number of adjacent working condition units participating in the interpolation. This represents the normalized chemical condition characteristic index. The number of features involved in the unit partitioning is a positive integer. Indicates the current operating condition. The normalized features are dimensionless and are directly read from the normalized feature vector of S1 output. Indicates the first in the working condition unit table The first adjacent working condition unit center The normalized coordinates of the terms are dimensionless numbers; Interpolation coefficients of adjacent working condition units are based on calculate, Indicates the first The interpolation coefficients for each adjacent working condition element are dimensionless numbers ranging from 0 to 1, and the sum of all coefficients is 1. This represents the index of adjacent working condition cells in the summation. Indicates the number of adjacent working condition units. Indicates up to the The dimensionless distance between the centers of adjacent working condition units This represents a dimensionless smoothing quantity to prevent the center distance from being zero; if a certain distance is less than... If so, then only the unit parameter is called and its interpolation coefficient is set to 1; Before adjacent working condition units participate in interpolation, a common version gating is performed. Since online sub-versions only apply to a single working condition unit, this boundary interpolation branch explicitly disables all online sub-versions, retaining only offline parameters belonging to the same offline release version and consistent in parent version chain, feature field dictionary version, and residual library generation. If the latest parameters of each unit are inconsistent, the most recent common offline version that can simultaneously cover all adjacent units, has passed offline verification, and whose current sampling time is no later than the verification expiration time is retrieved in reverse order of verification time, and the interpolation is recalculated using this common version. If no common version exists, the current period's reliable compensation is stopped, and the most recent valid correction value is retained as unreliable. The diagnostic value outputs a status indicating no version-consistent parameter and passes it to S4, S5, and S6. When a common version exists, the scale, bias, and weighted output calibration bias are weighted and summed according to the interpolation coefficients of adjacent units. The interpolated scale is still truncated to 0.5–1.5, and the error component bias is still truncated to -5–5℃. When the current sampling time is later than the verification expiration time, or when the material batch, sensor structure, feature field dictionary, or residual library generation changes as described in S2, the parameter immediately becomes invalid and is deleted from the candidate set. After deletion, the common version search is re-executed. When there is no common valid version among all adjacent units, the current parameter is frozen and a status indicating no valid calibration parameter is output. Before calculating the component contribution, the interpolated first... Each weight output calibration bias is applied to the network weight channel, according to... , and Calculate sequentially; Indicates the calibrated number Each weighted channel outputs a dimensionless number. Indicates the weight module of S2 The unconstrained outputs of each channel are dimensionless numbers. This indicates the first element obtained by directly reading from the current cell or by interpolation from adjacent cells. The calibration bias at each weighted output terminal is a dimensionless number. Indicates the calibrated number Class non-negative intermediate weights are dimensionless numbers. Indicates the calibrated number The weights for different operating conditions are dimensionless and the sum of the four terms is 1. Indicates a weighted normalized summation index; the denominator is less than All four calibration weights were set to 0.25 and the degradation status was recorded. The total compensation is obtained by summing the contributions of the four error components, specifically according to... and Calculate and correct surface temperature according to calculate, Indicates the first The contribution of the error components, in °C. This indicates the number of elements after processing with the current operating condition unit weight calibration parameters. The weights for non-negative operating conditions are dimensionless numbers. Indicates the first Error components after calibration, in °C. This represents the total compensation amount, in degrees Celsius (°C). This represents the index for four error types. This indicates the corrected surface temperature, in °C. This indicates the current raw temperature reading, in °C. The compensation calculation module writes the original temperature reading, four calibrated error components, four non-negative operating condition weights, four contribution amounts, total compensation amount, corrected surface temperature, all operating condition unit keys involved in interpolation, common parameter versions, parent version chains, feature field dictionary versions, residual library generations, and interpolation status into the compensation result record. The contribution amount and corrected surface temperature are passed to S4 for dominant error diagnosis and reliable coverage interval calculation, while the total compensation amount is simultaneously passed to S5 for online calibration criterion calculation.

[0024] In this specific embodiment, S4 includes: The diagnostic module reads the four contribution values ​​from the S3 compensation result record and first verifies the decomposition verification record of the current parameter version. This record must indicate that the root mean square error of the four components in the simulation reserve set and the mean absolute error of the load case weights are all no higher than the corresponding offline verification threshold generated by S2 using the independent simulation verification set. The temperature zone, load case unit key, independent verification set version, and parent version chain in the threshold record must be consistent with the current parameter version. If verification fails, the diagnostic module first searches for the most recent offline version that has passed decomposition verification and covers the current load case. If no such fallback candidate version exists, the dominant error sorting and reliable coverage interval output are immediately stopped, and the current correction value is marked as diagnostic only. If the version is deemed unreliable, write the offline version status that has not passed decomposition verification and covers the current operating condition. Prohibit the continued use of the verification-failed version and pass this status to S5 to set the update condition to unmet, and to S6 to perform parameter freezing and handle the lack of applicable version. If a rollback candidate version exists, return the rollback version identifier, the current original temperature reading, and the operating condition feature vector to S3. Instruct S3 to recalculate the four calibrated error components, four non-negative operating condition weights, four contribution amounts, total compensation amount, and corrected surface temperature using the rollback version within the same sampling period, overwriting the original compensation result record and verifying that the parameter version in the record is the rollback version. S3 completes the recalculation and version consistency. After verification, S4 writes the original failed version to the persistent disable table and passes the failed version identifier, rollback version identifier, and recalculated verification result to S6. The S6 parameter management module clears the pointer to the currently active online sub-version of the failed version or its sub-version using atomic transactions, sets the rollback offline version as the currently active offline version and the calling version for subsequent sampling periods, and returns the transaction result to S4. Only when the persistent disable write, active pointer clearing, and rollback offline version switching are all successful does S4 read the four recalculated contribution values ​​and corrected surface temperature. If any transaction fails, the version is inconsistent, or the recalculation fails, the dominant error sorting and trusted coverage interval input are stopped. If the result of this cycle is not found, mark it as unreliable and pass the failure reason to S5 to set the online update condition to not be met, and pass it to S6 to freeze the parameters; only when the decomposition verification passes or the above rollback recalculation and atomic switching all pass, are they arranged in descending order of absolute value of contribution, and the contact and heat conduction error, convection error, radiation error or dynamic hysteresis error corresponding to the first position is determined as the dominant error type; when the difference between the absolute values ​​of two contribution quantities does not exceed 0.01℃, the first position is determined according to the fixed order of contact and heat conduction, convection, radiation and dynamic hysteresis, and at the same time, the parallel mark and the second largest contribution quantity are saved in the diagnostic record to avoid frequent switching of types due to numerical noise at adjacent sampling times; For each operating condition unit, paired residual samples that were not involved in the offline calibration parameter calculation are reserved from the traceable reference surface temperature test data according to the test round, and stored in the residual library according to the operating condition unit key, parameter version, and verification time; records are excluded only according to predefined auditable data quality conditions that are unrelated to the residual result value, including invalid certificates, excessive synchronization time difference, excessive sensor range, acquisition channel failure, or deviation from test procedures; records with residual absolute values ​​exceeding three times the median absolute deviation are only marked with a tail sample, are not deleted due to the residual amplitude itself, and continue to be retained in the empirical quantile calculation and independent coverage verification population; marked records can only be excluded when there is fault evidence independent of the residual value, and the fault evidence, exclusion reason, and approval record are saved; a subset division of the credible coverage interval is only entered when the total number of valid quality samples is not less than 40, and when it is less than 40, the nearest neighboring units with the same temperature zone, the same temperature difference direction, and the same parameter version are merged and written into the merge source; For the convection condition with the extrapolation flag written in S1, the extrapolation residual library established by the controlled wind tunnel extrapolation test is called separately. This library is indexed by fluid medium, flow direction, untrunculated original velocity range, over-limit amplitude range relative to the 20m / s boundary, convection calibration curve version, and calibration parameter version. Each index unit stores no less than 40 independent paired residual samples that did not participate in offline parameter fitting. According to the mutual exclusion rules described later in this step, the samples are divided into subsets of quantiles and independent verification subsets of no less than 20 samples each. The independent verification subset covers the current original velocity. Only when all indexes are hit, the parameter versions are consistent, the number of samples in both subsets meets the threshold, and the independent verification coverage rate passes the threshold described later in this step, is the empirical quantile of the corresponding extrapolation residual used to form the interval. Otherwise, the extrapolation condition status is set to unreliable and the output of reliable coverage interval is stopped. In this specific implementation, the target coverage probability is set to 0.95. The effective residual absolute values ​​are sorted from smallest to largest, and then... and Determine the empirical quantiles. This represents the empirical quantile of the absolute value of the residual corresponding to the target coverage probability, in °C. Indicates the sorted order of the first... The absolute value of the residual at position, in °C. Indicates the ordinal position of the empirical quantiles and restricts them to 1 to 2. between, This represents the target coverage probability, a dimensionless number between 0 and 1, and is taken as 0.95 in this specific implementation. This represents the number of valid residual samples, which is a positive integer. This indicates the rounding up operation; Trusted coverage intervals by generate, This indicates the reliable coverage range of the corrected surface temperature, in °C. This represents the corrected surface temperature obtained from S3, in °C. The empirical quantile of the working condition matching residual library is expressed in °C; the upper and lower limits of the interval, the version of the residual library used, the number of samples, the target coverage probability, and the extrapolation flag are written into the coverage result record so that S5 can verify the source of the interval. When the current operating condition falls within the applicable range of two or more adjacent operating condition units, the empirical quantiles of each unit under the same parameter version are read, arranged in descending order of value, and the first empirical quantile is used to expand the reliable coverage interval, with the first value being used as the half width of the interval; when the number of samples in any unit is insufficient or the residual library version is behind the current calibration parameter version, its status is recorded as unverifiable and the reliable coverage interval is stopped from being output, without replacing it with the quantiles of narrower units, in order to maintain conservative coverage of the boundary operating condition; The diagnostic module simultaneously verifies the applicable range of the operating condition feature vector recorded in the current offline calibration parameter version, the verification expiration time, and the immediate failure event specified in S2. When the operating condition feature vector in S1 exceeds the applicable range of the current version, the current sampling time is later than the verification expiration time, or when there are specified changes in the material batch, sensor structure, feature field dictionary, or residual library generation, the corresponding out-of-range or parameter failure status is written into the coverage result record, the interval calculated in this step is stopped from being marked as a reliable coverage interval, and the event is passed to S6 to trigger version rollback judgment. The residual library construction module establishes a five-level index based on temperature zone, temperature difference direction, convective heat transfer coefficient range, contact state range, and parameter version. Only one time-aligned residual record is retained for the same test round. The corresponding residual library is regenerated after each offline calibration parameter version change, without mixing the residuals under the old parameters with the residuals under the new parameters. After generation, the file integrity is verified by the number of records, minimum value, maximum value, median, and empirical quantile summary. To verify the experience coverage, the residual library is sorted in ascending order by the start time of the test round. If the start times are the same, it is sorted in lexicographical order by the test round identifier. The residual library is then divided into odd and even subsets based on the sorting number: odd-numbered numbers are included in the subset for quantile calculation, and even-numbered numbers are included in the subset for independent verification. The version of this partitioning rule is saved with the residual library. All samples from each test round can only be included in one subset, and the number of valid samples in each subset must not be less than 20. If either subset has fewer than 20 samples, the nearest adjacent operating condition units with the same temperature range, the same temperature difference direction, and the same parameter version are merged. If the coverage is still insufficient after re-dividing independently according to the same sorting rules, the coverage status is set to unverifiable, the output of the reliable coverage interval is stopped, and the S5 update condition is set to unsatisfied. After the threshold is met, the interval half-width is obtained by using quantiles to obtain a subset, and the proportion of the reference surface temperature falling into the interval is statistically analyzed using the independent verification subset. When the independent verification proportion is lower than the target coverage probability minus 0.02, the working condition unit is marked as insufficient coverage and the output of the reliable interval is stopped. It will be enabled again after supplementing test samples or re-calibrating offline. This verification is not avoided by narrowing the target temperature range. The contribution ranking record saves the signed numerical value, absolute value, ranking number and error type code of the four contribution values. The dominant error type is traceable along with the compensation result record identifier. When the same dominant error type appears in three consecutive sampling times, a stable dominant flag is generated. If three consecutive sampling times are not reached, the current ranking result is still output but no maintenance alarm is triggered, so that the diagnostic conclusion is independent of the single temperature correction. The upper and lower limits of the interval remain unchanged within the applicable temperature range of -60 to 600℃. If either end exceeds the temperature range, the value is not truncated. Instead, the reason for exceeding the temperature range is written into the trusted state and the trusted output is stopped. This process retains the corrected surface temperature and interval value for diagnosis, while avoiding the possibility of the interval width becoming smaller after truncation, which may erroneously trigger the S5 online update condition. When the version identifiers of the residual sample library, calibration parameter records, and operating condition unit table cannot form the same parent version chain, the diagnostic module sets the version consistency status to failure and rejects interval calculation; when the versions are consistent and the current operating condition is located at the boundary of adjacent units, the overwrite result record simultaneously saves all candidate unit keys, their respective empirical quantiles, and the finally adopted unit key, so that the conservative expansion process using the maximum empirical quantile can be verified. After each residual library update, the diagnostic module performs a replay verification, reading the original temperature readings, reference surface temperatures, and corresponding total compensation amounts of the reserved reference samples one by one, and recalculating the absolute value of the residuals, empirical quantiles, and interval hit results. If the replay results are inconsistent with the residual library summary, or if the sample association identifier is duplicated or missing, the residual library version is locked, the previous verified version is retained, and the reason for the lock is sent to the parameter management module to prevent unverified intervals from participating in online updates. The overwrite result record also saves the calculation program version and record generation time, and calculates the hash value of the content after writing; when S5 reads, it first verifies the hash value and the record association identifier. If the verification fails, the overwrite result is regarded as unreliable and the previous record is retained only for comparison. It is not allowed to enter the criterion calculation. The dominant error type, the contribution identifier after sorting, the upper and lower limits of the credible coverage interval, the empirical quantile, the residual library version, the operating condition unit key, the number of samples, and the credible status together constitute the diagnosis and coverage result record; the credible record is input into S5 to form the online calibration criterion, and the unreliable or missing interval record makes S5 directly set the criterion to not meet the preset update conditions, while retaining the dominant error type for maintaining the diagnosis.

[0025] In this specific embodiment, S5 includes: When the online calibration criterion module obtains reference temperature measurement channel data, it reads the traceable reference surface temperature, calibration certificate identifier, certificate validity period, reference sensor serial number, measurement uncertainty, and sampling timestamp, and verifies their consistency with the certificate registration form. If the certificate identifier exists, the current date is within the validity period, the serial number is consistent, and the expanded uncertainty is not greater than the offline allowable value for the corresponding temperature zone, the reference traceability status is set to valid; otherwise, it is set to invalid and the reason for failure is saved. The reference temperature measurement channel and the contact surface temperature channel are paired with a synchronization tolerance of 0.2s. The reference value is converted to ℃ using the correction value and unit specified in the calibration certificate. The S3 correction result and the S4 coverage result are read at the same sampling time. The reference measurement stability judgment is calculated from the last event when the reference device setting value changes, the heating power of the measured object changes, the installation contact state changes, or the fluid boundary conditions change. The stability duration and reference temperature change rate threshold of each temperature zone are determined by the offline reference test procedure. In this specific implementation, the absolute value of the reference temperature change rate does not exceed 0.02℃ / s within the most recent consecutive 10s before the stable state is set to valid. The stability threshold record saves the duration, change rate threshold and unit, test procedure version, offline verification data version, number of samples, applicable temperature zone and generation date. If there is no paired sample, the stable state is invalid, the stability threshold record is missing or inconsistent with the current temperature zone and parameter version, the S4 interval state is unreliable, or the current operating condition has an extrapolation flag, no updateable criterion is generated and the current reference record is retained for offline verification. Coverage range width according to calculate, Indicates the width of the coverage area, in degrees Celsius (°C). This represents the upper limit of the S4 trusted coverage area, in degrees Celsius (°C). This represents the lower limit of the S4 trusted coverage interval, in °C. If the upper limit is less than the lower limit or the width exceeds the maximum value allowed by the current temperature zone residual library, the interval status will be set to invalid. The width threshold is determined by offline verification data under the same target coverage probability and is saved with the residual library version. Component consistency residuals calculate, This represents the component consistency residual, in °C. This represents the synchronously acquired and traceably valid reference surface temperature, in °C. This indicates the synchronized original temperature reading, in °C. This represents the total compensation amount output by S3, in °C. This residual is defined in the same way as the paired residual during offline calibration of S2, and its absolute value is used to determine whether the current component synthesis result is still within the range of offline verification. The online calibration criterion record is composed of the reference traceability status, reference measurement stability status, stability threshold record version, coverage interval width, component consistency residual, number of hits of working condition unit, current parameter version, certificate verification result and data synchronization status. When a valid reference value is obtained, each field is passed to S6 for item-by-item comparison with preset update conditions. When no reference value is obtained or any required status is invalid, the update condition field of the criterion is set to not met and the reason for not meeting is noted. The current reference value is not replaced by the historical reference value. The width threshold is determined by the 95th empirical quantile of the width of the offline verification interval under the same temperature zone and the same target coverage probability. The consistency threshold is determined by the 95th empirical quantile of the absolute value of the residual left by the corresponding working condition unit. The threshold record saves the value in ℃, the residual library version, the number of samples, the calculation date and the applicable working condition unit. S5 only calls the threshold record that is consistent with the current parameter version. If no consistent record is found, the update condition is set to not met. The criterion module performs an integrity check on each field. The reference traceability status must be a valid Boolean value, the coverage interval width and component consistency residual must be finite values, and the number of hits in the working condition unit must be a positive integer. When non-numerical, infinite values, missing fields, or broken record association identifiers occur, threshold comparison is not performed and the data integrity status is set to failure to prevent abnormal data from being interpreted as meeting the update conditions. When there are two or more reference temperature measurement channels at the same sampling time, the certificate, range, and synchronization verification are completed first. Then, the consistency test is performed on each pair of channels that have passed the verification. If the difference between any pair of reference temperatures exceeds the sum of the expanded uncertainties of the pair of channels, the reference consistency status is set to failure and updates are prohibited, regardless of whether the two uncertainty values ​​are the same. The original record is written to the audit log for review. After all channel pairs have passed the consistency test, they are sorted in ascending order of expanded uncertainty values, and the channel at the top of the list is selected as the criterion. If the values ​​are the same, the first channel is determined by the dictionary order of the reference sensor serial number. The online calibration criterion record, along with the reference sample identifier, S3 compensation result record identifier, and S4 coverage result record identifier, is written to the append-only calibration audit log. The log saves the threshold version and parameter version used when the criterion was formed. This record serves as the only input that allows triggering online parameter updates in S6, thereby keeping the parameters frozen for sampling periods lacking traceable reference surface temperatures.

[0026] In this specific embodiment, S6 includes: The parameter management module uses the S5 appending only the current cycle update condition field in the audit log as the sole authorized entry point for online updates, and does not re-form alternative criteria with fewer conditions in S6. Only when this field is satisfied will the following be verified item by item before entering any update calculation: reference traceability status is valid, reference measurement stability status is valid, synchronization status is valid, data fields are complete, coverage interval width is not greater than the width threshold of the current temperature zone, absolute value of component consistency residual is not greater than the consistency threshold of the current operating condition unit, operating condition feature vector falls within the applicable range of only one offline calibration parameter, S4 trust status is valid, threshold record is consistent with the current parameter version, residual library version is consistent and there is no extrapolation status flag. When all verifications pass, the Boolean result of the current cycle preset update condition is set to satisfied and subsequent updates will proceed. If any field is missing, invalid, or version inconsistent, the Boolean result will be set to unsatisfied before entering update calculation, the current cycle parameters will be immediately frozen, the consecutive unsatisfied count will be incremented by 1, and the rejection field and reason will be written to the running status record. The width threshold and consistency threshold are both determined by the 95% empirical quantile of offline verification data and bound to the parameter version. When the preset update conditions are met, with the goal of reducing the squared residual of dimensionless components, only the output scale, bias, or output bias of the four error components associated with the current working condition unit are updated, while the shared hidden layer remains frozen; for each parameter to be updated, first press... Establish dimensionless parametric coordinates, and according to Establish dimensionless residuals; Represents the dimensionless coordinates of the parameter to be updated. This indicates the current physical value of the parameter. The scale and weight output biases are dimensionless, while the error component biases are in °C. This represents the baseline value of the parameter in the corresponding offline parent version, with units equal to or equal to the base value. same, This indicates that the parameter is a positive normalized scale given by offline validation, with units equal to... same, This represents the dimensionless component consistency residual. This represents the component consistency residual calculated by S5 based on the reference temperature, the original temperature, and the total compensation, in °C. This represents the consistency threshold bound to the current operating condition unit and parameter version, expressed in °C and taking a value greater than 0. Dimensionless parametric coordinates according to Update, the updated physical parameters are as follows recover; This represents the updated dimensionless parameter coordinates. This represents the dimensionless parameter coordinates before the update. This represents the online learning rate, a dimensionless number ranging from 0.001 to 0.01 determined through offline validation. Indicates a dimensionless target right The gradient of is a dimensionless number. This represents the allowable single update magnitude of the parameter in dimensionless coordinates, expressed as a dimensionless number divided by the corresponding physical update limit. get, This indicates that the dimensionless input is being used. Restricted to the lower bound With the upper realm Truncation operations between This represents a dimensionless real number input to be truncated. This indicates the dimensionless lower bound used in the truncation operation. This indicates that the dimensionless upper bound used in the truncation operation is not less than 1. , Indicates the updated physical parameters, units are 1 and 2. and Consistent; The single update magnitude of the error component scale is limited to 0.01 and the cumulative range is maintained between 0.5 and 1.5. The single update magnitude of the error component bias is limited to 0.05℃ and the cumulative range is maintained between -5℃ and 5℃. The single update magnitude of the bias at the output end of the operating condition weight module is limited to 0.01. The parameter management module deduplicates samples according to the sampling time identifier and only accumulates reference samples that are valid in certificate, valid in synchronization status, and whose operating condition unit key is consistent with the current unit. When the current operating condition unit has less than 10 accumulated valid reference samples, it does not generate or accept candidate online updates, keeps the old parameters, freezes the update of this cycle, and records the number of insufficient samples. After accumulating 10, it recalculates the root mean square of the residuals of the most recent 10 valid reference samples using the parent version and the candidate version respectively. When the number of samples reaches 10, the parameter management module only accepts candidate updates that reduce the absolute value of the residuals of the current paired samples and whose root mean square of the residuals of the most recent 10 samples of the candidate version is not higher than the corresponding value of the parent version. Otherwise, it cancels the candidate update and keeps the old parameters. After the load condition weight module is updated, the four load condition weights are recalculated according to the S2 non-negative transformation and the normalization rule of summing to one. Any weight less than 0 or the sum of the four weights deviating from 1 by more than 1 will result in a load condition weight being applied. The update is deemed invalid and the old weight parameters are restored. The online update parameters are checked and then appended with the parent version identifier, working condition unit key, update sample identifier, update time, cumulative update count and verification statistics to form an online sub-version that only applies to the single working condition unit. Before the online candidate sub-version is put into the next sampling cycle, an independent sample that has not participated in the current gradient update, has the same unit key as the current working condition, and whose certificate and synchronization status are valid is selected from the reserved reference sample of the parent version; for each reserved sample, the same inference output record retained by S2 is read according to the sampling time identifier and the working condition record identifier, and the output of the four unconstrained weight channels is directly obtained. With four prediction error components, candidate scales, biases, and weight output calibration biases are applied to recalculate the total S3 compensation and paired residuals, without inversely inferring from historical normalized weights. If the inference record is missing, the identifier is inconsistent, or the model version does not belong to the candidate parent version chain, the sample is deemed invalid and the reason is recorded. A candidate residual library is established according to the candidate sub-version identifier. At the same time, the root mean square error of the four components and the mean absolute error of the load weights are recalculated using the simulation reserved samples of the same working unit. The simulation reserved samples do not overlap with the network training samples and the updated samples. The candidate residual library must not use the residual values ​​of the parent version. The total number of its quality valid samples is not less than 40 and is independent of the updated samples according to the test rounds. Then, according to the same sorting and parity mutual exclusion rules of S4, it is divided into subsets and independent verification subsets of no less than 20 samples each. Then, the empirical quantiles, reliable coverage intervals and independent coverage rates are recalculated, and the online calibration criteria are re-formed according to the same rules of S5. If the total number of samples is less than 40 or any subset is less than 20, the candidate verification is directly determined to be unsuccessful, the candidate sub-version is withdrawn and the original activity pointer is maintained. Samples from adjacent working units are not used. Supplement the online sub-version of the single working condition unit; the candidate sub-version is only submitted when the criteria still meet the preset update conditions, the independent coverage rate is not less than the target coverage probability minus 0.02, the root mean square of the residuals is not higher than the parent version, the root mean square errors of the four types of components are not higher than the parent version, and the average absolute error of the working condition weights is not higher than the parent version; when submitting, the sub-version identifier, working condition unit key, parent offline version, candidate residual library version, threshold record version, verification statistics, and next sampling period number are written to the current active online sub-version pointer in an atomic transaction, so that it is called by S3 first from the next sampling period. S3, S4, and S5 all verify the corresponding parameters, residual library, and threshold according to the actual active version in the compensation result record; if any field of the transaction fails to write or the candidate verification fails, the original active pointer is not changed and the candidate sub-version is revoked. The revocation reason, candidate update amount, candidate residual library statistics, decomposed verification statistics, and the parent version to be maintained are recorded in the audit log; When the preset update conditions are not met, the parameter management module freezes the shared hidden layer, error component output parameters, and operating condition weight parameters. If there are offline calibration parameters that have been called before the current sampling time, have passed offline verification, and whose applicable scope covers the current operating conditions, then the current version will continue to be called, so that S3 maintains the output corrected surface temperature and S4 maintains the output reliable coverage range when the residual library is valid. At the same time, the reason for freezing and the number of consecutive non-compliance are written into the running status record. The parameter management module saves the version identifier, applicable range of the operating condition feature vector, verification time, verification expiration time, residual library version, and parent version for each offline calibration parameter version. At the end of each sampling cycle, it directly reads the Boolean result of the preset update condition for this cycle, formed after the first segment of this step complies with the S5 update condition field and completes all verifications. Only when this result is satisfied will the consecutive non-satisfaction count be cleared to zero. If any condition is not satisfied, the count is incremented by 1 and the update for this cycle is frozen. When the number of hits of the operating condition unit is not equal to 1, the preset update condition is explicitly set to unsatisfied and the number of hits and the The reasons are recorded in the running status log; missing or invalid reference values, invalid reference measurement stable state, synchronization failure, missing data fields, version inconsistency, lack of qualified residual library for extrapolated operating conditions, invalid reliable state, and failure to meet any threshold will also increase the count by 1. If the count has already been increased by the first segment in the same sampling period, it will not be accumulated again in this segment. The preset number of samplings is used to evaluate 1 to 10 candidate continuous periods one by one in an independent offline playback set. The playback set includes normal stable segments, reference failure segments, version out-of-bounds segments, and online sub-version abnormal segments. The selected segments have a false rollback rate of no more than 1%. The minimum number of candidate iterations for anomalies lasting no more than 5 seconds is selected in this specific implementation with a 1-second sampling period of 5. The rollback strategy version record saves the selected number of iterations, sampling period, false rollback rate threshold, anomaly duration threshold, replay data version, number of samples, calculation date, and applicable parameter version. During runtime, the record consistent with the current parameter version is read. If the record is missing or the version is inconsistent, it is immediately frozen and the count is not cleared. When the working condition feature vector exceeds the applicable range of the current version, the current sampling time is later than the verification expiration time, the immediate failure event specified in S2 occurs, the preset number of consecutive samplings given by the rollback strategy version record does not meet the preset update conditions, or the online calibration criterion changes from met to unmet after the current active online sub-version is called by S3, the online update parameters are immediately revoked, the pointer of the current active online sub-version is cleared with an atomic transaction, and the offline parameter version that has passed offline verification the most recently, has not yet reached the verification expiration time, and whose applicable range covers the current working condition is retrieved in reverse order of verification time. After the retrieval is successful, its error component output parameters and non-negative working condition weight parameters are restored, it is set as the current active offline version, and it is called by S3 from the next sampling period. If there are no applicable offline calibration parameters that cover the current operating conditions and have passed offline verification, the parameter management module freezes all network parameters, writes a status flag indicating that the current operating conditions exceed the trusted coverage range to the compensation result record, and commands S4 to stop outputting the calculated range as a trusted coverage range. The final output record saves the original temperature reading, the current corrected surface temperature value, the trusted status of the correction value, the dominant error type, the parameter version, the freeze or rollback status, and the audit log index. When the trusted status of the correction value is negative, only the value is retained for diagnosis and not as a trusted temperature measurement result.

[0027] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0028] This invention combines the interpretable error component generated by transient conjugate heat transfer simulation with the measured total error generated by traceable reference surface temperature test. This allows the BP neural network to utilize the influence of different heat transfer paths on the temperature sensing element readings, and to correct the deviation between the simulation model and the field object through paired residual correction. In this way, it can form a compensation amount corresponding to the current operating condition under different temperature zones, environments and installation conditions.

[0029] This invention restricts neural network updates to parameters or operating condition weights at the output of error components, and uses the reliable coverage interval and component consistency residuals as update gates. When the reference is unavailable, the verification conditions are not met, or the applicable range of parameters cannot cover the current operating conditions, online updates are stopped and freeze, undo, or offline version rollback is performed, thereby suppressing compensation drift while preserving the model's nonlinear correction capability.

Claims

1. A surface temperature measurement correction method based on thermal simulation and BP neural network, characterized in that, include: S1. Collect the raw temperature readings and their rate of change over time from the contact surface temperature sensor, along with the ambient temperature, local convection state, surrounding radiation temperature, contact state, and structural parameters of the sensor and the measured object, to form a working condition feature vector. S2. Input the operating condition feature vector into a multi-output BP neural network trained based on transient conjugate heat transfer simulation data and calibrated with traceable reference surface temperature test data. Output the contact and heat conduction, convection, radiation and dynamic hysteresis error components and their corresponding non-negative operating condition weights. The transient conjugate heat transfer simulation data is obtained in the following way: a finite element transient conjugate heat transfer model is established, which includes the measured body, contact layer, sensor package, temperature sensing element, surrounding fluid and radiation boundary. The ambient temperature, fluid state, radiation boundary, contact state and material and structural parameters are combined and set within the measured surface temperature range of −60 to 600℃. The undisturbed surface temperature without the sensor and the temperature sensing element temperature after the sensor is installed are calculated separately, and the difference between the undisturbed surface temperature and the temperature sensing element temperature is used as the total error label. Extract the heat flow conducted at the contact boundary, the heat flow convected on the sensor surface, the heat flow radiated on the sensor surface, and the transient heat storage of the temperature sensing element. Calculate and normalize the equivalent temperature difference corresponding to each heat flow according to the preset boundary heat flow separation order to form component labels for the contact and heat conduction error components, convection error components, radiation error components, and dynamic hysteresis error components. Make the composite value of each component label correspond to the total error label. The multi-output BP neural network includes a shared hidden layer that receives the operating condition feature vector, four error component output terminals that output four error components respectively, and an operating condition weight module that outputs the non-negative operating condition weights. The operating condition weight module generates at least four non-negative numbers based on the difference between the original temperature reading and the ambient temperature, the time change rate, the convective heat transfer state quantity obtained by normalizing the local convection state, the radiative heat transfer state quantity obtained by normalizing the ambient radiation temperature, and the contact heat transfer state quantity obtained by normalizing the contact state, and normalizes the sum of the four non-negative numbers to one. The training loss includes the component loss between each error component and its corresponding component label, the consistency loss between the error synthesized by each error component according to the non-negative operating condition weight and the total error label, and the sign constraint loss for violating the corresponding boundary heat flow direction. The traceable reference surface temperature test data is divided into multiple operating condition units according to the temperature zone, the direction of the temperature difference between the original temperature reading and the ambient temperature, the convective heat transfer coefficient range, and the contact state range. Within each operating condition unit, the difference between the traceable reference surface temperature and the synchronously acquired original temperature reading is taken as the total measured error. The total compensation amount is obtained by summing each error component according to the non-negative operating condition weights, and the difference between the total measured error and the total compensation amount is taken as the pairing residual. Under the condition of keeping the shared hidden layer unchanged, at least one of the bias, scale and non-negative operating condition weights of the output terminals of each error component is calibrated according to the pairing residuals to obtain the offline calibration parameters corresponding to each operating condition unit. S3. Combine each error component according to the non-negative working condition weight to obtain the total compensation amount and the contribution amount of each error component. Add the total compensation amount to the original temperature indication to obtain the corrected surface temperature. S4. Determine the dominant error type by sorting the absolute values ​​of the contribution, and determine the reliable coverage range of the corrected surface temperature based on the measured residual distribution of the working condition matching. S5. When a traceable reference surface temperature is obtained, set the online calibration criteria based on the traceable reference surface temperature, the corrected surface temperature, the reliable coverage range, and the total compensation amount; if it is not obtained, set the criteria to not meet the preset update conditions. S6. According to the criteria, if the preset update conditions are met, the output parameters of the error component or the non-negative operating condition weights are updated by limiting the amplitude; if not, the parameters are frozen.

2. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 1, characterized in that, S1 includes: The original temperature reading is obtained by the sensor acquisition circuit, and the rate of change of time is determined by the difference between the readings at adjacent sampling times and the sampling time interval. The ambient temperature is obtained by an ambient temperature sensor; the local convection state, characterizing at least one of fluid temperature, flow velocity, and convective heat transfer coefficient, is obtained by a flow velocity sensor, a fluid temperature sensor, or heat transfer calculation based on flow boundary conditions; the ambient radiation temperature is obtained by a radiation temperature sensor or by calculation of ambient surface temperature and emissivity; the contact state is characterized by at least one of contact pressure, contact area, and contact thermal resistance; and the sensor package size, temperature sensing element location, material thermal conductivity and heat capacity, as well as the thickness, thermal conductivity, and heat capacity of the measured object are used as the material and structural parameters. The parameters are normalized according to the value range of each parameter in the offline training samples to form the working condition feature vector.

3. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 1, characterized in that, S3 includes: applying the bias and scale corresponding to the current operating condition unit to each error component, multiplying the calibrated error component by the corresponding non-negative operating condition weight to obtain the contribution of the error component, and summing the contribution of the four error components to obtain the total compensation amount; when the operating condition feature vector is in the boundary region of multiple adjacent operating condition units, interpolating the offline calibration parameters of each adjacent operating condition unit according to the normalized distance from the operating condition feature vector to the center of each adjacent operating condition unit, and using the interpolated parameters to calculate the total compensation amount.

4. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 3, characterized in that, S4 includes: The contribution quantities are arranged in descending order of absolute value, and the error component corresponding to the first contribution quantity is determined as the dominant error type. The residual samples not used for obtaining the offline calibration parameters are reserved from the traceable reference surface temperature test data of each operating condition unit. The empirical quantiles of the absolute values ​​of the residual samples are determined according to the preset target coverage probability. The reliable coverage interval is obtained by subtracting and adding the empirical quantiles from the corrected surface temperature. When the current operating condition falls within the applicable range of two or more adjacent operating condition units, the empirical quantiles of the adjacent operating condition units are arranged in descending order of value, and the reliable coverage interval is expanded using the empirical quantile at the top of the list.

5. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 2, characterized in that, S5 includes: verifying the calibration certificate identifier, validity period, and measurement uncertainty of the traceable reference surface temperature to obtain the reference traceability status; calculating the difference between the upper and lower limits of the reliable coverage interval to obtain the coverage interval width; calculating the difference between the traceable reference surface temperature and the original temperature reading, and subtracting the total compensation amount to obtain the component consistency residual; and combining the reference traceability status, the coverage interval width, and the component consistency residual to form the online calibration criterion.

6. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 5, characterized in that, S6 include: The preset update conditions are: the reference traceability status is valid, the coverage interval width is not greater than the width threshold determined by offline verification data, the absolute value of the component consistency residual is not greater than the consistency threshold determined by offline verification data, and the operating condition feature vector falls within the applicable range of only one offline calibration parameter. When the preset update conditions are met, the output bias, scale, or non-negative operating condition weight of the error component associated with the applicable range is updated with the goal of reducing the component consistency residual, and the update amount of each update is limited to the update range of the corresponding parameter determined by offline verification data. When the preset update conditions are not met, the update stops, and if there is an offline calibration parameter that has been called before the current sampling time and whose applicable range covers the current operating condition, the parameter continues to be called, while maintaining the output of the corrected surface temperature and the reliable coverage interval.

7. The surface temperature measurement correction method based on thermal simulation and BP neural network according to claim 2, characterized in that, For each offline calibration parameter version, set a version identifier, applicable range of operating condition feature vector, and verification time. When the operating condition feature vector exceeds the applicable range of the current version, the preset number of consecutive samplings determined by offline verification data fails to meet the preset update condition, or the online calibration criterion changes from meeting the preset update condition to not meeting the preset update condition after calling online update parameters, cancel the error component output parameters or non-negative operating condition weights after online update, revert, and call the most recently verified offline calibration parameter whose applicable range covers the current operating condition according to the version identifier. If there is no offline calibration parameter whose applicable range covers the current operating condition, freeze the parameters, first output the status identifier that the current operating condition exceeds the reliable coverage range, and stop outputting the interval calculated in step S4 as the reliable coverage interval.

Images