A Calibration and Measurement Method for Electrical Measuring Instruments Based on Smart Sensors

Abstract

This invention discloses a calibration and metrology method for electrical measuring instruments based on intelligent sensors, comprising the following steps: synchronously collecting relevant data of the electrical measuring instrument under calibration to construct a metrological state field; generating a cause-of-fact evidence vector and calculating the corresponding evidence confidence vector; inserting controlled disturbance segments to form a calibration sequence, generating a measurement residual sequence and recording it in a structured manner according to error components to obtain the original sample set of cause-of-fact response spectrum; inputting relevant data into an improved TimeMixer model to output a spectrum embedding vector and a set of fractionated error parameters; constructing a cause-of-fact response spectrum and screening stable invariants to generate a minimum calibration generation set; generating an uncertainty budget based on the minimum calibration generation set and determining the dynamic validity period boundary; identifying drift-dominant causes through online self-checking and adaptively updating the calibration results. This invention improves the accuracy and long-term stability of electrical measuring instrument calibration and is suitable for high-precision intelligent metrology scenarios.

Description

Technical Field

[0001] This invention relates to the field of electrical measurement and metrological calibration technology, and in particular to a calibration and metrological method for electrical measuring instruments based on intelligent sensors. Background Technology

[0002] With the widespread application of high-precision electrical measuring instruments in metrology, industrial control, and scientific research, the requirements for the accuracy, long-term stability, and calibration effectiveness of measurement results are constantly increasing. Existing calibration methods for electrical measuring instruments mostly rely on standard source comparison or periodic manual calibration, typically performed under ideal environments or single operating conditions. This makes it difficult to reflect the true metrological characteristics of the instrument under the combined influence of various factors such as environmental changes, switching of operating states, and external electromagnetic disturbances during actual operation. In complex application scenarios, factors such as temperature, humidity, power supply ripple, contact conditions, and electromagnetic interference often exhibit time-varying and coupled characteristics, easily causing measurement error drift and uncertainty accumulation.

[0003] In existing technologies, the acquisition and processing of multi-source state data often employs simple time alignment or post-hoc statistical correction methods. Differences in sampling frequencies, time drift, and missing data from different sensor channels are difficult to compensate for effectively, leading to an unclear correlation between instrument status and measurement output. Furthermore, traditional calibration methods typically focus only on overall error or a single error index, lacking structured modeling of error components such as zero-point bias, proportional gain, noise, and transient recovery, making it difficult to identify the dominant causes of error. For non-stationary and nonlinear measurement residual signals, existing analysis methods based on fixed models or simple regression struggle to accurately characterize disturbance response patterns and cannot assess the impact of different calibration data on the reliability of the results.

[0004] In addition, existing calibration results often lack dynamic validity management and online self-testing mechanisms. Once the instrument's operating environment or status changes, it is difficult to detect and update the calibration parameters in a timely manner, which can easily lead to unreliable measurement results.

[0005] Therefore, how to provide a calibration and measurement method for electrical measuring instruments based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a calibration and metrology method for electrical measuring instruments based on intelligent sensors. This invention constructs a metrology state field by integrating instrument operating status and environmental data collected by multiple types of intelligent sensors and introduces a controlled disturbance calibration mechanism. It combines an improved TimeMixer model to characterize the response relationship between the inducement and error components, and further realizes the screening of stable invariants, the construction of the minimum calibration generation set, and the uncertainty budgeting and dynamic validity period management. It has the advantages of high calibration accuracy, strong adaptability to complex operating conditions, high efficiency of calibration data utilization, and support for online self-testing and drift adaptive updates.

[0007] A calibration and measurement method for electrical measuring instruments based on intelligent sensors according to an embodiment of the present invention includes the following steps: Step 1: Collect the original measurement output sequence, instrument operating status data, and intelligent sensor array data of the electrical measuring instrument being calibrated, and perform time alignment to form a metrological state field; Step 2: Calculate the causal evidence vector from the econometric state field, and calculate the evidence credibility vector based on sampling integrity, time window consistency, and anomalous jump rate; Step 3: Insert controlled disturbance segments to form a calibration sequence, record the disturbance type, intensity level and time window, generate the measurement residual sequence and record it in a structured manner according to the error components, and obtain the original sample set of the cause-error response spectrum; Step 4: Input the measurement residual sequence, the causal evidence vector, the evidence confidence vector, and the controlled perturbation label into the improved TimeMixer model. The improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module. Based on the evidence confidence vector, the mixed output is gated and modulated, and the spectrum embedding vector and the fractional error parameter set are output. Step 5: Construct the cause-error response map by fusing the embedding vector of the fusion map with the original sample set of the cause-error response map, calculate the consistency index and sensitivity index, screen stable invariants and determine the gating compensation term, and construct the minimum calibration generation set; Step 6: Generate an uncertainty budget and a dynamic validity period boundary based on the minimum calibration generation set, the cause-error response spectrum, and the evidence confidence vector, forming a calibration result package; Step 7: Insert an online self-test segment to update the metrological state field and generate an error component change fingerprint. Match the error component change fingerprint with the cause-error response spectrum to determine the dominant cause of drift. Selectively update the gating compensation term or the fractional error parameter set and update the dynamic validity period boundary simultaneously.

[0008] Optionally, step one specifically includes: The original measurement output sequence of the electrical measuring instrument under calibration is collected, and the instrument working status data is collected simultaneously. The instrument working status data includes the range indicator, integration time indicator, reference source access status indicator, and input terminal connection status indicator. Synchronously collect data from an intelligent sensor array, which is generated by various intelligent sensors deployed near key heat sources, input terminals, and power supply and shielding structures inside the electrical measuring instrument being calibrated. The original measurement output sequence, instrument working status data and intelligent sensor array data are written to the acquisition timestamp respectively. Nearest neighbor alignment is used and the corresponding time data is supplemented by keeping the valid value of the previous time when there are missing samples. Perform unit consistency and numerical range normalization processing on the aligned channel data; The aligned original measurement output sequence, instrument operating status data, and smart sensor array data are combined into a metrological state field according to the same time index.

[0009] Optionally, step two specifically includes: Set a sliding time window and divide the measurement state field into windows according to the time index. Each time window corresponds to a set of intelligent sensor array data and instrument working status data within the same time range. Preprocessing is performed on the data from each sensor channel within each time window. The preprocessing includes median filtering of the data within the time window to obtain a smoothed sequence, and marking data points in the smoothed sequence whose deviation from the median within the time window exceeds a preset mutation threshold as mutation points and replacing them with non-mutation points adjacent to the mutation points by interpolation to obtain the causal base sequence. The causal evidence vector is generated window by window based on the causal base sequence. The causal evidence vector includes a temperature level field, a thermal gradient field, a humidity level field, a contact state field, a ripple intensity field, an interference intensity field, and a range switching field. The evidence credibility vector is calculated window by window based on the causal evidence vector. The evidence credibility vector corresponds one-to-one with the causal evidence vector according to the fields. The evidence credibility vector is obtained by combining the sampling integrity sub-credibility, the time window consistency sub-credibility, and the abnormal jump rate sub-credibility. Write the causal evidence vector and evidence credibility vector corresponding to each time window into the time window index position of the econometric state field and establish a correlation with the time index of the original measurement output sequence covered by the current time window.

[0010] Optionally, step three specifically includes: A continuous time window is selected as the insertion interval on the time window index of the metrological state field, and a controlled perturbation segment is applied to the electrical measuring instrument under calibration to form a calibration sequence. For each controlled disturbance segment, a disturbance tag is generated and written into the measurement state field. The disturbance tag includes the disturbance type, intensity level and corresponding start and end time window index, and is associated with the time index of the original measurement output sequence covered by the insertion interval. Within the insertion interval, the adjacent time window before the disturbance is extracted from the original measurement output sequence as the baseline segment, and the mean value of the baseline segment at the corresponding range is used as the reference value. The original measurement output sequence within the controlled disturbance segment covering the time window is sampled point by point and the reference value is subtracted from the reference value to generate the measurement residual sequence. The measurement residual sequence is recorded in a structured manner according to the error components, which include zero-point bias components, proportional gain components, short-term noise components, and transient recovery components. The perturbation labels, measurement residual sequences, and error component structured records are aggregated according to time window indexes to obtain the original sample set of the cause-error response map.

[0011] Optionally, the improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module: The multi-scale sequence construction module divides the measurement residual sequence into residual window sequences according to the time window index. It downsamples the residual window sequences with a set step size to form a first-scale sequence, and takes the window mean of the residual window sequences with a sliding window to form a second-scale sequence. It maps the continuous fields in the causal evidence vector to causal encoding vectors, converts the controlled perturbation label into a perturbation encoding vector, and concatenates it with the causal encoding vector as a condition vector. The condition vector is copied according to the time window index and added element by element to each scale sequence to obtain a conditional multi-scale sequence representation. The decomposition module extracts the trend component by one-dimensional smooth convolution at each scale for the conditional multi-scale sequence representation, and subtracts the trend component element by element from the original multi-scale sequence representation to obtain the residual component. The trend component and the residual component are mapped, and the trend mapping results and residual mapping results at different scales are concatenated in scale order to form the decomposed multi-scale feature sequence. The previously decomposable hybrid module takes the mean of the decomposed multi-scale feature sequence along the time window dimension, concatenates it with the evidence credibility vector, and inputs it into a linear layer to generate a scale score. The scale score is then exponentially calculated and normalized to obtain the scale hybrid weight. The multi-scale feature sequence is weighted and summed according to the scale hybrid weight to obtain a hybrid feature sequence. The hybrid feature sequence is input into a two-layer feedforward network for channel mixing, and the output is added element-wise to the hybrid feature sequence to obtain the hybrid feature sequence after residual connection. The causal evidence vector is passed through a linear layer to generate a causal modulation vector, and then multiplied element-wise with the hybrid feature sequence according to the time window index to obtain the previously decomposable hybrid output. The future multi-predictor hybrid module inputs the past decomposable hybrid outputs into the convolutional predictor and the recurrent predictor respectively to obtain two candidate outputs. The perturbation coding vector corresponding to the controlled perturbation label is concatenated with the time window mean of the past decomposable hybrid outputs and then input into the linear layer to generate the predictor score. The predictor score is exponentially normalized to obtain the predictor hybrid weight. The candidate outputs are weighted and summed according to the predictor hybrid weight to obtain the hybrid output. The evidence confidence vector is input into the linear layer and gating vector is obtained by the Sigmoid function. The gating vector is copied by time window index and multiplied element-wise with the mixed output. The graph embedding vector is obtained by global average pooling and linear mapping. The multi-head linear layer output is divided into a set of quantized error parameters corresponding to the zero-point bias component, proportional gain component, short-term noise component and transient recovery component.

[0012] Optionally, step five specifically includes: Using the time window index as the key, the original sample set of the cause-error response map is aligned and fused with the map embedding vector to construct the cause-error response map. For each response edge, consistency and sensitivity indices are calculated. The consistency index is obtained by reading the same error component parameter of multiple time windows under the same disturbance type and intensity level, averaging the values ​​according to the evidence credibility vector, and summing the deviations of each time window parameter from the center value according to the evidence credibility vector. The sensitivity index is obtained by reading the same error component parameter from low to high intensity level under the condition that the disturbance type remains unchanged, calculating the amplitude of the difference between parameters of adjacent intensity levels, and summing them up by weighting according to the evidence credibility vector. Stable invariants are selected based on consistency and sensitivity indices. The stable invariants are response edges where the consistency index is better than a preset consistency threshold and the sensitivity index falls within a preset sensitivity range. For each stable invariant, the set of quantified error parameters of the corresponding error component is read and averaged according to the evidence credibility vector to obtain the compensation base value. At the same time, the evidence credibility vector of the corresponding cause field is read to generate a gating coefficient. The gating coefficient is obtained by linear mapping of the evidence credibility vector and through the Sigmoid function. The gating coefficient is multiplied by the compensation base value to obtain the gating compensation term. An index is established according to the cause field, error component, disturbance type, and intensity level and written into the cause-error response spectrum. A minimum calibration generation set is constructed based on stable invariants. The minimum calibration generation set is a set of entries that satisfy coverage constraints. The coverage constraints include selecting at least one stable invariant for each error component and at least one intensity level for each disturbance type.

[0013] Optionally, step six specifically includes: Uncertainty allocation is performed on the gating compensation term and the fractional error parameter set based only on the entries contained in the minimum calibration generation set, forming a set of uncertainty budget entries; For each error component, calculate the compensation uncertainty and the induced propagation uncertainty, sum the squares of the compensation uncertainty and the induced propagation uncertainty and take the square root to obtain the combined uncertainty of the current error component, and write the combined uncertainty of each error component into the uncertainty budget item set. A dynamic validity period boundary is generated based on the uncertainty budget item set. The combined uncertainty of the current time window is updated sequentially along the time window index covered by the minimum calibration generation set. The combined uncertainty is compared with the preset allowable error threshold. When the combined uncertainty exceeds the allowable error threshold, the corresponding time window index is marked as the dynamic validity period termination boundary. When the combined uncertainty does not exceed the allowable error threshold, the dynamic validity period termination boundary is updated by recursively pushing forward according to the time window index. When the sampling integrity sub-confidence in the evidence confidence vector of the corresponding time window in the minimum calibration generation set is lower than the preset integrity threshold, the current time window index is marked as the validity period degradation point. The set of uncertainty budget entries and the dynamic validity period boundary are combined to generate a calibration result package, and the calibration result package is associated with the time index of the metrological state field.

[0014] Optionally, step seven specifically includes: An online self-test segment is inserted within the dynamic validity period boundary of the calibration result package, and the original measurement output sequence, instrument working status data and intelligent sensor array data are collected simultaneously to update the metrological state field and update the causal evidence vector and evidence credibility vector. The original measurement output sequence within the online self-test segment coverage time window is subtracted point by point to obtain the self-test quantity residual sequence. The self-test quantity residual sequence is decomposed into zero-point bias component, proportional gain component, short-term noise component and transient recovery component, and then spliced ​​in the order of error components to form the error component change fingerprint. The error component change fingerprint is subtracted from the fractional error parameter set component by component, and the matching degree is obtained by weighting and summarizing according to the gating coefficient. The cause field corresponding to the entry with the best matching degree is selected as the dominant cause of drift. Selectively update the gating compensation term or the fractional error parameter set based on the dominant cause of drift: when the evidence confidence vector of the corresponding item meets the preset confidence threshold and the consistency index is better than the preset consistency threshold, update the compensation base value of the gating compensation term with the self-test component parameter value; otherwise, update the fractional error parameter set with the self-test component parameter value and write back the cause-error response spectrum. After the update is completed, recalculate the uncertainty budget based on the update result and update the dynamic validity period boundary simultaneously to form the updated calibration result package.

[0015] The beneficial effects of this invention are: This invention introduces multiple types of intelligent sensors to synchronously sense the operating status of electrical measuring instruments and environmental factors, constructing a metrological state field. This achieves unified time alignment and structured expression of measurement output and multi-source state data, effectively solving the problem that existing calibration methods struggle to characterize the sources of measurement errors under complex operating conditions. Through structured modeling of controlled disturbance segments and error components, measurement errors are refined into components such as zero-point bias, proportional gain, short-term noise, and transient recovery, significantly improving the interpretability and specificity of error analysis. Furthermore, an improved TimeMixer model is used to model multi-scale, non-stationary measurement residuals, utilizing an evidence credibility gating mechanism to suppress the influence of low-quality data on calibration results, achieving stable extraction of the factor-error response relationship. Based on this, stable invariants are screened using consistency and sensitivity indices, and a minimum calibration generation set is constructed, reducing calibration data redundancy and uncertainty propagation. Simultaneously, uncertainty budgeting and dynamic validity boundaries are introduced to achieve quantitative evaluation and dynamic management of the validity of calibration results. Finally, through online self-checking and drift-dominant factor identification mechanisms, adaptive updating of calibration parameters and long-term stable operation are achieved. This invention improves the accuracy, reliability, and intelligence of electrical measuring instruments in complex environments, and has significant engineering application value and promotional significance. Attached Figure Description

[0016] 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 a calibration and measurement method for electrical measuring instruments based on intelligent sensors proposed in this invention; Figure 2 This is a schematic diagram of a calibration and measurement method for electrical measuring instruments based on intelligent sensors proposed in this invention; Figure 3 This is a framework diagram of the improved TimeMixer model in the calibration and measurement method for electrical measuring instruments based on intelligent sensors proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figure 1-3 A calibration and measurement method for electrical measuring instruments based on intelligent sensors includes the following steps: Step 1: Collect the original measurement output sequence, instrument operating status data, and intelligent sensor array data of the electrical measuring instrument being calibrated, and perform time alignment to form a metrological state field; Step 2: Calculate the causal evidence vector from the econometric state field, and calculate the evidence credibility vector based on sampling integrity, time window consistency, and anomalous jump rate; Step 3: Insert controlled disturbance segments to form a calibration sequence, record the disturbance type, intensity level and time window, generate the measurement residual sequence and record it in a structured manner according to the error components, and obtain the original sample set of the cause-error response spectrum; Step 4: Input the measurement residual sequence, the causal evidence vector, the evidence confidence vector, and the controlled perturbation label into the improved TimeMixer model. The improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module. Based on the evidence confidence vector, the mixed output is gated and modulated, and the spectrum embedding vector and the fractional error parameter set are output. Step 5: Construct the cause-error response map by fusing the embedding vector of the fusion map with the original sample set of the cause-error response map, calculate the consistency index and sensitivity index, screen stable invariants and determine the gating compensation term, and construct the minimum calibration generation set; Step 6: Generate an uncertainty budget and a dynamic validity period boundary based on the minimum calibration generation set, the cause-error response spectrum, and the evidence confidence vector, forming a calibration result package; Step 7: Insert an online self-test segment to update the metrological state field and generate an error component change fingerprint. Match the error component change fingerprint with the cause-error response spectrum to determine the dominant cause of drift. Selectively update the gating compensation term or the fractional error parameter set and update the dynamic validity period boundary simultaneously.

[0019] In this embodiment, step one specifically includes: The original measurement output sequence of the electrical measuring instrument under calibration is collected, and the instrument working status data corresponding to the original measurement output sequence is collected synchronously within the same preset acquisition period. The instrument working status data includes the range indicator, integration time indicator, reference source access status indicator, and input terminal connection status indicator. Synchronously collect data from an intelligent sensor array. The intelligent sensor array data is generated by various intelligent sensors deployed near key heat sources, input terminals, and power supply and shielding structures inside the electrical measuring instrument being calibrated. These include temperature sensors, humidity sensors, contact status sensors for characterizing terminal connection status, and interference sensors for characterizing the external electromagnetic environment. The original measurement output sequence, instrument operating status data and intelligent sensor array data are written with acquisition timestamps respectively. The instrument operating status data and intelligent sensor array data are mapped to the sampling time of the original measurement output sequence according to the timestamp. Nearest neighbor alignment is used and the corresponding time data is supplemented by keeping the valid value of the previous time when there are missing samples. The data from each aligned channel are processed for unit unification and numerical range normalization. Unit unification unifies the dimensions of the output of temperature, humidity and interference sensors. Numerical range normalization linearly scales the data from each channel according to the preset upper and lower limits and truncates out-of-bounds values. The aligned original measurement output sequence, instrument operating status data, and smart sensor array data are combined into a metrological state field according to the same time index.

[0020] In this embodiment, step two specifically includes: Set a sliding time window and divide the measurement state field into windows according to the time index. Each time window corresponds to a set of intelligent sensor array data and instrument working status data within the same time range. Preprocessing is performed on the data from each sensor channel within each time window. The preprocessing includes median filtering of the data within the time window to obtain a smoothed sequence, and marking data points in the smoothed sequence whose deviation from the median within the time window exceeds a preset mutation threshold as mutation points and replacing them with non-mutation points adjacent to the mutation points by interpolation to obtain the causal base sequence. Based on the underlying sequence of the triggering factors, a triggering evidence vector is generated window by window. The triggering evidence vector includes the following fields and is output in the order of the same field: The temperature level field takes the average value of each temperature sensor within the time window and is represented by both the maximum and minimum values. The thermal gradient field takes the maximum difference between the mean values ​​of each temperature sensor within the time window as the thermal gradient characterization. The humidity level field is characterized by the average value and fluctuation range of the humidity sensor within the time window, where the fluctuation range is determined by the difference between the maximum and minimum values ​​within the time window. The contact state field is a contact resistance representation output by a contact state sensor near the input terminal, and the average value and fluctuation amplitude of the contact resistance representation within a time window are used to represent it. The ripple intensity field is the output of the ripple sensor near the power supply and power supply branch. It is characterized by the fluctuation amplitude of the ripple sensor output within the time window and the number of short-period peaks and valleys. The number of short-period peaks and valleys is obtained by statistically analyzing the number of sign changes of adjacent sampling points within the time window. The interference intensity field is the output of the interference sensor near the shielding structure, and is characterized by the maximum value, average value and fluctuation amplitude of the interference sensor output within a time window. The range switching field reads the range indicator from the instrument's working status data. If the range indicator changes within the current time window, the number of changes and the range indicator before and after the change are recorded as the range switching field. The evidence credibility vector is calculated window-by-window based on the causal evidence vector. The evidence credibility vector corresponds one-to-one with the causal evidence vector by field. The evidence credibility vector is generated and combined from the following sub-credibility: The sampling integrity sub-confidence is calculated by comparing the actual number of valid sampling points for each field within the statistical time window with the number of sampling points that should be sampled in the current time window based on the integral time identifier. When there are missing sampling points, the sampling integrity sub-confidence of that field is reduced according to the proportion of missing points. The consistency confidence of the time window is determined by differentiating the evidence values ​​of the same field between the current time window and the previous adjacent time window. If the difference exceeds a preset stability threshold, the consistency confidence of the current field is reduced. The stability threshold is preset for each field and switches according to the range indicator. The sub-confidence of the abnormal jump rate is calculated by the proportion of the number of times the current field is marked as a mutation point within the time window to the total number of sampling points of the current field. When the proportion of mutation points exceeds the preset jump threshold, the sub-confidence of the abnormal jump rate of the field is reduced. Write the causal evidence vector and evidence credibility vector corresponding to each time window into the time window index position of the econometric state field and establish a correlation with the time index of the original measurement output sequence covered by the current time window.

[0021] In this embodiment, step three specifically includes: A continuous time window is selected as the insertion interval on the time window index of the measurement state field. A controlled disturbance segment is applied to the electrical measuring instrument under calibration to form a calibration sequence. The controlled disturbance segment is a repeatable disturbance sequence that presets or injects the equivalent conditions of the input terminal, the access state of the reference source, or the electromagnetic environment within the insertion interval. For each controlled disturbance segment, a disturbance tag is generated and written into the measurement state field. The disturbance tag includes the disturbance type, intensity level and corresponding start and end time window index, and is associated with the time index of the original measurement output sequence covered by the insertion interval. Within the insertion interval, the adjacent time window before the disturbance is extracted from the original measurement output sequence as the baseline segment, and the mean value of the baseline segment at the corresponding range is used as the reference value. The original measurement output sequence within the controlled disturbance segment covering the time window is sampled point by point and the reference value is subtracted from the reference value to generate the measurement residual sequence. The measurement residual sequence is recorded in a structured manner according to the error components. The error components include zero-point bias component, proportional gain component, short-term noise component and transient recovery component. The zero-point bias component is characterized by the average level of the residual sequence, the proportional gain component is characterized by the corresponding relationship between the residual amplitude and the intensity level, the short-term noise component is characterized by the fluctuation amplitude of the residual, and the transient recovery component is characterized by the number of time windows required for the residual to return to the preset stable threshold after the disturbance ends. The perturbation labels, measurement residual sequences, and error component structured records are aggregated according to time window indexes to obtain the original sample set of the cause-error response map.

[0022] In this embodiment, the improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module: The multi-scale sequence construction module divides the measurement residual sequence into residual window sequences according to the time window index. It downsamples the residual window sequences with a set step size of 2 to form the first-scale sequence. It then takes the window mean of the residual window sequences with a sliding window of length 8 to form the second-scale sequence. The continuous fields in the causal evidence vector are input into a single linear mapping layer to obtain the causal encoding vector. The controlled perturbation label is converted into a perturbation encoding vector by a lookup table. The vector is concatenated with the causal encoding vector to form the condition vector. The condition vector is copied according to the time window index and added element by element to each scale sequence to obtain the conditional multi-scale sequence representation. The decomposition module extracts the trend component by using a one-dimensional smooth convolution with a kernel length of 5 in each scale for the conditional multi-scale sequence representation. The original multi-scale sequence representation is subtracted from the trend component element by element to obtain the residual component. The trend component and the residual component are respectively input into a one-dimensional convolutional layer and connected to the ReLU activation function for mapping. The trend mapping results and residual mapping results at different scales are concatenated in scale order to form the decomposed multi-scale feature sequence. The previously decomposable hybrid module takes the mean of the decomposed multi-scale feature sequence along the time window dimension, concatenates it with the evidence credibility vector, and inputs it into a linear layer to generate a scale score. The scale score is then exponentially calculated and normalized to obtain the scale hybrid weight. The multi-scale feature sequence is weighted and summed according to the scale hybrid weight to obtain a hybrid feature sequence. The hybrid feature sequence is input into a two-layer feedforward network for channel mixing, and the output is added element-wise to the hybrid feature sequence to obtain the hybrid feature sequence after residual connection. The causal evidence vector is passed through a linear layer to generate a causal modulation vector, and then multiplied element-wise with the hybrid feature sequence according to the time window index to obtain the previously decomposable hybrid output. The future multi-predictor hybrid module inputs the past decomposable hybrid output into a convolutional predictor consisting of two layers of one-dimensional dilated convolutions and a recurrent predictor consisting of a single layer of gated recurrent units, respectively, to obtain two candidate outputs. The perturbation coding vector corresponding to the controlled perturbation label is concatenated with the time window mean of the past decomposable hybrid output and then input into a linear layer to generate a predictor score. The predictor score is then exponentially normalized to obtain the predictor hybrid weight. The candidate outputs are weighted and summed according to the predictor hybrid weight to obtain the hybrid output. The evidence confidence vector is input into the linear layer and gating vector is obtained by the Sigmoid function. The gating vector is copied by time window index and multiplied element-wise with the mixed output. The graph embedding vector is obtained by global average pooling and linear mapping. The multi-head linear layer output is divided into a set of quantized error parameters corresponding to the zero-point bias component, proportional gain component, short-term noise component and transient recovery component.

[0023] This implementation employs an improved TimeMixer model to perform multi-scale modeling of the measurement residual sequence. It injects causal evidence vectors and controlled disturbance labels as conditions into the multi-scale construction, mixing, and predictor weighting processes to adapt to the nonlinear and non-stationary characteristics of electrical measurement errors under complex operating conditions. Compared to the original TimeMixer model, which primarily targets pure sequence prediction and lacks explicit constraints on metrological causal and disturbance conditions, this model further introduces trend-residual decomposition to separate slow-varying drift from fast disturbance responses. It introduces evidence confidence vectors to generate gating vectors to modulate the mixed output, suppressing the impact of low-confidence information caused by missing sampling, inconsistent time windows, and anomalous jumps. Furthermore, it uses multi-head outputs divided by error components to generate a fractionalized error parameter set, thereby improving the stability and interpretability of causal-error response relationship extraction, reducing the interference of anomalous data on calibration results, and enhancing the reliability of calibration parameters and the effectiveness of subsequent uncertainty budgeting and dynamic validity management.

[0024] In this embodiment, step five specifically includes: Using the time window index as the key, the original sample set of the cause-error response map is aligned and fused with the map embedding vector to construct the cause-error response map. The cause-error response map uses the cause field and error component as nodes, and "cause field - error component" as response edge. The edge attributes of the response edge record the controlled perturbation label, evidence credibility vector and fractional error parameter set, and the map embedding vector is written as the map attribute associated with the corresponding time window index. For each response edge, calculate the consistency index and sensitivity index: The consistency index is obtained by reading the same error component parameter of multiple time windows under the same disturbance type and intensity level, averaging the values ​​according to the evidence credibility vector, and summing the deviations of each time window parameter from the center value according to the evidence credibility vector. The sensitivity index is obtained by reading the same error component parameter from low to high intensity level under the condition that the disturbance type remains unchanged, calculating the amplitude of the difference between parameters of adjacent intensity levels, and summing them up by weighting according to the evidence credibility vector. Stable invariants are selected based on consistency and sensitivity indices. The stable invariants are response edges where the consistency index is better than a preset consistency threshold and the sensitivity index falls within a preset sensitivity range. For each stable invariant, the set of quantified error parameters of the corresponding error component is read and averaged according to the evidence credibility vector to obtain the compensation base value. At the same time, the evidence credibility vector of the corresponding cause field is read to generate a gating coefficient. The gating coefficient is obtained by linear mapping of the evidence credibility vector and through the Sigmoid function. The gating coefficient is multiplied by the compensation base value to obtain the gating compensation term. An index is established according to the cause field, error component, disturbance type, and intensity level and written into the cause-error response spectrum. A minimum calibration generation set is constructed based on stable invariants. The minimum calibration generation set is a set of entries that satisfy the coverage constraint. The coverage constraint includes selecting at least one stable invariant for each error component and at least one intensity level for each disturbance type. The stable invariants are sorted from best to worst according to the consistency index and from high to low according to the evidence credibility vector. The corresponding time window index, controlled disturbance label and gated compensation term are selected and recorded in sequence until the coverage constraint is satisfied, thus obtaining the minimum calibration generation set.

[0025] In this embodiment, step six specifically includes: Using the time window index in the minimum calibration generation set as the key, the inducement field, error component, controlled disturbance label, and associated gating compensation term and fractional error parameter set corresponding to the time window index are read from the inducement-error response spectrum. The evidence confidence vector corresponding one-to-one with the inducement field is also read. Uncertainty allocation is performed on the gating compensation term and fractional error parameter set based only on the entries contained in the minimum calibration generation set to form an uncertainty budget entry set. The uncertainty budget entry set records the uncertainty entries of the zero-point bias component, proportional gain component, short-term noise component and transient recovery component according to the error component. For each error component, a component uncertainty entry is generated. The gate coefficient in the gated compensation term corresponding to the error component is used as a weight to weight and summarize the dispersion of the compensation base value in the minimum calibration generation set to obtain the compensation uncertainty. The sensitivity index corresponding to the error component in the cause-error response spectrum is used as the disturbance propagation coefficient and multiplied by the fluctuation amplitude of the evidence value of the same cause field in the current metrological state field within the coverage time window of the minimum calibration generation set to obtain the cause propagation uncertainty. The compensation uncertainty and the cause propagation uncertainty are summed by squares and the square root is taken to obtain the combined uncertainty of the current error component. The combined uncertainty of each error component is written into the uncertainty budget entry set. A dynamic validity period boundary is generated based on the uncertainty budget item set. The combined uncertainty of the current time window is updated sequentially along the time window index covered by the minimum calibration generation set. The combined uncertainty is compared with a preset allowable error threshold. When the combined uncertainty exceeds the allowable error threshold, the corresponding time window index is marked as the dynamic validity period termination boundary. When the combined uncertainty does not exceed the allowable error threshold, the dynamic validity period termination boundary is updated by recursively pushing forward according to the time window index. When the sampling integrity sub-confidence in the evidence confidence vector of the corresponding time window in the minimum calibration generation set is lower than the preset integrity threshold, the current time window index is marked as the validity period degradation point, which is used to shrink the dynamic validity period boundary in advance. The calibration result package is generated by combining the set of uncertainty budget items with the dynamic validity period boundary. The calibration result package includes: a minimum calibration generation set index table, gating compensation items corresponding one-to-one with the minimum calibration generation set index table, a set of fractional error parameters, the combined uncertainty of each error component, and the start and end time window index of the dynamic validity period. The calibration result package is then associated with the time index of the metrological state field.

[0026] In this embodiment, step seven specifically includes: An online self-test segment is inserted within the dynamic validity period boundary of the calibration result package, and the original measurement output sequence, instrument working status data and intelligent sensor array data are collected simultaneously to update the metrological state field and update the causal evidence vector and evidence credibility vector. Using the mean of the original measurement output sequence of the adjacent time window before the online self-test segment as the reference value, the original measurement output sequence within the time window covered by the online self-test segment is subtracted point by point to obtain the self-test quantity residual sequence. The self-test quantity residual sequence is decomposed into zero-point bias component, proportional gain component, short-term noise component and transient recovery component according to the error component structured recording method, and then spliced ​​in the error component order to form the error component change fingerprint. The time window index of the minimum calibration generation set is used to locate the cause-error response spectrum entry, read the corresponding gating compensation item and the fractional error parameter set, subtract the error component change fingerprint from the fractional error parameter set component by component, and sum the results according to the gating coefficient to obtain the matching degree. The cause field corresponding to the entry with the best matching degree is selected as the drift-dominant cause. Selectively update the gating compensation term or the fractional error parameter set based on the dominant cause of drift: when the evidence confidence vector of the corresponding item meets the preset confidence threshold and the consistency index is better than the preset consistency threshold, update the compensation base value of the gating compensation term with the self-test component parameter value; otherwise, update the fractional error parameter set with the self-test component parameter value and write back the cause-error response spectrum. After the update is completed, recalculate the uncertainty budget based on the update result and update the dynamic validity period boundary simultaneously to form the updated calibration result package.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to the DC voltage calibration business of a metrology and testing institution. The object to be calibrated was a 6½-digit benchtop digital multimeter (the electrical measuring instrument being calibrated), which was used alternately in a constant-temperature laboratory and an open workstation. Common problems included measurement error drift caused by the superposition of temperature and humidity fluctuations, power supply ripple, electromagnetic interference, and changes in the contact status of the input terminals. Traditional methods often only detect deviations when calibration is due, lacking quantifiable validity period judgment and online correction methods. In this embodiment, the original measurement output sequence, instrument operating status data, and intelligent sensor array data are synchronously connected to the acquisition terminal. The sampling rate is set to 10 s / s, and the integration time is set to 1 PLC. The instrument operating status data fixedly includes the range indicator, sampling rate or integration time indicator, reference source access status indicator, and input terminal connection status indicator. The intelligent sensor array is configured with two temperature sensors near key heat sources inside the machine, one temperature sensor and one contact status sensor (outputting contact resistance) near the input terminals, one ripple sensor near the power supply and power supply branch, one interference sensor near the shielding structure, and one humidity sensor inside the machine. All sensors have a uniform sampling period of 1 second. After writing timestamps to all channels, they are aligned to the sampling time of the original measurement output sequence according to nearest neighbor. If there are missing samples, they are filled with valid values ​​from the previous moment. The temperature unit is uniformly ℃, the humidity unit is uniformly %RH, and the interference and ripple are expressed in the sensor amplitude output unit. The numerical range is linearly scaled according to the upper and lower limits of the range, and out-of-range values ​​are truncated. Finally, the data are combined according to the same time index to form the measurement state field.

[0028] When the metrological state field enters the stage of generating causal evidence and credibility, the sliding time window has a window length of 8 sampling points (0.8s) and a step size of 4 sampling points (0.4s). Within each time window, each channel is first subjected to 3-point median filtering to obtain a smooth sequence. Then, a mutation point is marked with a mutation threshold of "deviation from the median within the window exceeding 3 times the standard deviation". The causal evidence vector fields are fixed as temperature level, thermal gradient, humidity level, contact state, ripple intensity, interference intensity, and range switching. The field values ​​are executed according to the claims. The credibility vector is composed of three sub-credibility combinations: sampling integrity, time window consistency, and abnormal jump rate. The sampling integrity missing threshold is 90%. The time window consistency stability threshold is fixed by field as follows: temperature 2℃, humidity 5%RH, contact state 10%, ripple intensity 15%, and interference intensity 15% (the corresponding threshold group is switched when switching ranges). The abnormal jump rate threshold is 20%. The data is written window by window into the metrological state field and associated with the original measurement output time index.

[0029] The calibration sequence is formed by inserting controlled disturbance segments, with the insertion interval selected as a continuous time window covering at least 20 seconds. The disturbance types are fixed as "input equivalent condition switching" and "power supply ripple injection," with fixed intensity levels of low, medium, and high. Input equivalent condition switching is achieved by switching the equivalent resistance network using a programmable switch, while power supply ripple injection is achieved through a programmable ripple injection unit. Each disturbance segment generates a disturbance label (disturbance type, intensity level, and start / end time window index) and writes it into the metrological state field. The measurement residual sequence uses the average of the original measurement outputs of adjacent time windows before the disturbance as a reference value, and is obtained by subtracting each sampling point within the disturbance coverage time window. The error components are recorded with a unified structure: the zero-point bias component is characterized by the residual mean, the proportional gain component by the correspondence between the residual amplitude and the intensity level, the short-term noise component by the residual fluctuation amplitude, and the transient recovery component by the number of time windows required for the residual to return to the preset stable threshold after the disturbance ends. The stable threshold is fixed at ±1µV, thus obtaining the original sample set of the cause-error response spectrum.

[0030] The modeling stage adopts the improved TimeMixer model described in claim 5, with the following fixed values: the multi-scale sequence construction module uses only two scales, the first scale being a downsampled sequence with a stride of 2, and the second scale being a window mean sequence with a window length of 8; the continuous field of the causal evidence vector is projected to 16 dimensions through a single linear mapping layer, and the controlled perturbation label is embedded through a lookup table to obtain a 16-dimensional perturbation encoding vector. The two are concatenated to form a 32-dimensional conditional vector, which is then broadcast and added element-wise to the two-scale sequences; the decomposition module uses a one-dimensional smooth convolution with a kernel length of 5 to extract the trend component and subtracts it to obtain the residual component. The trend and residual are concatenated after being mapped by one-dimensional convolution and ReLU, respectively; the channel expansion factor of the two-layer feedforward network of the previously decomposable hybrid module is set to 2, and it is combined with the input... The input elements are summed one by one to form a mixed feature sequence after residual connection. Then, the causal evidence vector is linearly generated into a causal modulation vector and multiplied element by element to obtain the past decomposable mixed output. The future multi-predictor mixing module is fixed to two branches. The convolutional predictor is a two-layer one-dimensional dilated convolution (convolutional kernel length 3, dilation rate 1 and 2 respectively), and the recurrent predictor is a single-layer gated recurrent unit (hidden state dimension 32). The predictor score is exponentially normalized to obtain the mixed weight and then weighted and summed on the candidate output. The evidence confidence vector is passed through a linear layer and Sigmoid to obtain a gated vector and multiplied element by element on the mixed output. Finally, the graph embedding vector is output through global average pooling and linear mapping, and the fractionalized error parameter set is output through a multi-head linear layer divided according to four error components.

[0031] When constructing the cause-error response map, the embedding vector of the fusion map is aligned with the original sample set using a time window index. The edge attributes of the response edges record the controlled perturbation label, the evidence credibility vector, and the set of fractional error parameters. Stable invariant selection uses a fixed threshold: the consistency threshold is ±15%, and the sensitivity range is 5%–30%. The gating coefficient is obtained by linear mapping and Sigmoid from the evidence credibility vector. The gating compensation term is formed by "gating coefficient × compensation base value" and an index of "cause field - error component - perturbation type - intensity level" is established. The minimum calibration generation set coverage constraint is fixed to select at least one stable invariant for each error component and at least one intensity level for each perturbation type. Items with higher consistency indicators and higher evidence credibility are included in the set first. The full sample of 240 items is compressed to 48 items, a compression ratio of approximately 80%. Step 6: Uncertainty budgeting is calculated solely based on the minimum calibration generation set, with a fixed allowable error threshold of ±5µV. The dynamic validity period boundary terminates at the time window when the combined uncertainty first exceeds the allowable error threshold. Simultaneously, when the confidence level of the sampled complete sub-components falls below 90%, it is marked as a validity period degradation point, and the validity period is shortened prematurely. Step 7: The online self-test segment length is fixed at 60s. The self-test residual generation method is consistent with Step 3. The self-test component parameter values ​​are concatenated to form an error component change fingerprint. The dominant cause of drift is determined by the matching degree of component-by-component difference and weighted summation according to the gating coefficient of the cause-error response spectrum entries. When the evidence confidence vector meets the confidence threshold of 0.8 and the consistency index is better than the threshold, only the gating compensation term's compensation base value is updated; otherwise, the fractionalized error parameter set is updated, and the dynamic validity period boundary is recalculated synchronously.

[0032] Three comparison methods were set up and data standards were unified. The traditional periodic calibration method is a conventional offline calibration: multi-point comparison is performed using a traceable standard source in a stable environment, zero-point bias and proportional gain calibration parameters are calculated and fixed into the instrument, and only fixed parameters are corrected within the usage cycle until the next cycle. This method does not collect environmental and status data during operation, nor does it have online self-testing and dynamic validity period evaluation. The linear temperature drift compensation method only uses the temperature mean to make a linear correction to the output, and the temperature drift coefficient is obtained and fixed by one offline fitting. The LSTM residual prediction method trains a single LSTM to predict the residuals and corrects them by regression, without introducing controlled disturbance label conditions or evidence credibility gating. The four methods were evaluated under 12 typical operating conditions (constant temperature laboratory, near switching power supply, 1m RF source, high temperature and humidity, low temperature, range switching, loose terminals, ripple injection, shielding gap interference, 8h and 24h continuous operation in open station), with indicators including RMS error, expanded uncertainty U95, and dynamic validity period, and the drift identification time was statistically analyzed during continuous operation in open station.

[0033] Table 1 Summary of Comprehensive Indicators for Each Method

[0034] As shown in Table 1, the RMS error of this invention is reduced by approximately 57.7% compared to traditional periodic calibration, and by approximately 38.9% compared to LSTM residual prediction; U95 is reduced by approximately 67.1% compared to traditional periodic calibration, indicating that under the same disturbance conditions, by screening stable invariants through consistency and sensitivity and introducing evidence credibility gating, uncertainty propagation can be significantly converged; the dynamic validity period is improved by approximately 85.3%, consistent with the "uncertainty budget driven by minimum calibration set + validity period degradation point", so that the validity period is determined by a quantifiable uncertainty threshold rather than an empirically fixed period; the drift identification time is shortened to about 8 minutes, which comes from the gated weighted matching of the error component change fingerprint formed by online self-testing and the cause-error response spectrum, which can locate the drift to the specific cause field and trigger selective updates, thereby reducing the duration of inaccuracy. In summary, this invention achieves a closed loop of "multi-source state synchronous perception - controlled disturbance modeling - credibility gating learning - minimum calibration set budget - online self-testing and self-updating" in real metrology applications, significantly improving calibration accuracy, long-term stability and engineering usability under complex operating conditions.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A calibration and measurement method for electrical measuring instruments based on intelligent sensors, characterized in that, Includes the following steps: Step 1: Collect the original measurement output sequence, instrument operating status data, and intelligent sensor array data of the electrical measuring instrument being calibrated, and perform time alignment to form a metrological state field; Step 2: Calculate the causal evidence vector from the econometric state field, and calculate the evidence credibility vector based on sampling integrity, time window consistency, and anomalous jump rate; Step 3: Insert controlled disturbance segments to form a calibration sequence, record the disturbance type, intensity level and time window, generate the measurement residual sequence and record it in a structured manner according to the error components, and obtain the original sample set of the cause-error response spectrum; Step 4: Input the measurement residual sequence, the causal evidence vector, the evidence confidence vector, and the controlled perturbation label into the improved TimeMixer model. The improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module. Based on the evidence confidence vector, the mixed output is gated and modulated, and the spectrum embedding vector and the fractional error parameter set are output. Step 5: Construct the cause-error response map by fusing the embedding vector of the fusion map with the original sample set of the cause-error response map, calculate the consistency index and sensitivity index, screen stable invariants and determine the gating compensation term, and construct the minimum calibration generation set; Step 6: Generate an uncertainty budget and a dynamic validity period boundary based on the minimum calibration generation set, the cause-error response spectrum, and the evidence confidence vector, forming a calibration result package; Step 7: Insert an online self-test segment to update the metrological state field and generate an error component change fingerprint. Match the error component change fingerprint with the cause-error response spectrum to determine the dominant cause of drift. Selectively update the gating compensation term or the fractional error parameter set and update the dynamic validity period boundary simultaneously.

2. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step one specifically includes: The original measurement output sequence of the electrical measuring instrument under calibration is collected, and the instrument working status data is collected simultaneously. The instrument working status data includes the range indicator, integration time indicator, reference source access status indicator, and input terminal connection status indicator. Synchronously collect data from an intelligent sensor array, which is generated by various intelligent sensors deployed near key heat sources, input terminals, and power supply and shielding structures inside the electrical measuring instrument being calibrated. The original measurement output sequence, instrument working status data and intelligent sensor array data are written to the acquisition timestamp respectively. Nearest neighbor alignment is used and the corresponding time data is supplemented by keeping the valid value of the previous time when there are missing samples. Perform unit consistency and numerical range normalization processing on the aligned channel data; The aligned original measurement output sequence, instrument operating status data, and smart sensor array data are combined into a metrological state field according to the same time index.

3. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step two specifically includes: Set a sliding time window and divide the measurement state field into windows according to the time index. Each time window corresponds to a set of intelligent sensor array data and instrument working status data within the same time range. Preprocessing is performed on the data from each sensor channel within each time window. The preprocessing includes median filtering of the data within the time window to obtain a smoothed sequence, and marking data points in the smoothed sequence whose deviation from the median within the time window exceeds a preset mutation threshold as mutation points and replacing them with non-mutation points adjacent to the mutation points by interpolation to obtain the causal base sequence. The causal evidence vector is generated window by window based on the causal base sequence. The causal evidence vector includes a temperature level field, a thermal gradient field, a humidity level field, a contact state field, a ripple intensity field, an interference intensity field, and a range switching field. The evidence credibility vector is calculated window by window based on the causal evidence vector. The evidence credibility vector corresponds one-to-one with the causal evidence vector according to the fields. The evidence credibility vector is obtained by combining the sampling integrity sub-credibility, the time window consistency sub-credibility, and the abnormal jump rate sub-credibility. Write the causal evidence vector and evidence credibility vector corresponding to each time window into the time window index position of the econometric state field and establish a correlation with the time index of the original measurement output sequence covered by the current time window.

4. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step three specifically includes: A continuous time window is selected as the insertion interval on the time window index of the metrological state field, and a controlled perturbation segment is applied to the electrical measuring instrument under calibration to form a calibration sequence. For each controlled disturbance segment, a disturbance tag is generated and written into the measurement state field. The disturbance tag includes the disturbance type, intensity level and corresponding start and end time window index, and is associated with the time index of the original measurement output sequence covered by the insertion interval. Within the insertion interval, the adjacent time window before the disturbance is extracted from the original measurement output sequence as the baseline segment, and the mean value of the baseline segment at the corresponding range is used as the reference value. The original measurement output sequence within the controlled disturbance segment covering the time window is sampled point by point and the reference value is subtracted from the reference value to generate the measurement residual sequence. The measurement residual sequence is recorded in a structured manner according to the error components, which include zero-point bias components, proportional gain components, short-term noise components, and transient recovery components. The perturbation labels, measurement residual sequences, and error component structured records are aggregated according to time window indexes to obtain the original sample set of the cause-error response map.

5. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, The improved TimeMixer model includes a multi-scale sequence construction module, a decomposition module, a past decomposable mixing module, and a future multi-predictor mixing module. The multi-scale sequence construction module divides the measurement residual sequence into residual window sequences according to the time window index. It downsamples the residual window sequences with a set step size to form a first-scale sequence, and takes the window mean of the residual window sequences with a sliding window to form a second-scale sequence. It maps the continuous fields in the causal evidence vector to causal encoding vectors, converts the controlled perturbation label into a perturbation encoding vector, and concatenates it with the causal encoding vector as a condition vector. The condition vector is copied according to the time window index and added element by element to each scale sequence to obtain a conditional multi-scale sequence representation. The decomposition module extracts the trend component by one-dimensional smooth convolution at each scale for the conditional multi-scale sequence representation, and subtracts the trend component element by element from the original multi-scale sequence representation to obtain the residual component. The trend component and the residual component are mapped, and the trend mapping results and residual mapping results at different scales are concatenated in scale order to form the decomposed multi-scale feature sequence. The previously decomposable hybrid module takes the mean of the decomposed multi-scale feature sequence along the time window dimension, concatenates it with the evidence credibility vector, and inputs it into a linear layer to generate a scale score. The scale score is then exponentially calculated and normalized to obtain the scale hybrid weight. The multi-scale feature sequence is weighted and summed according to the scale hybrid weight to obtain a hybrid feature sequence. The hybrid feature sequence is input into a two-layer feedforward network for channel mixing, and the output is added element-wise to the hybrid feature sequence to obtain the hybrid feature sequence after residual connection. The causal evidence vector is passed through a linear layer to generate a causal modulation vector, and then multiplied element-wise with the hybrid feature sequence according to the time window index to obtain the previously decomposable hybrid output. The future multi-predictor hybrid module inputs the past decomposable hybrid outputs into the convolutional predictor and the recurrent predictor respectively to obtain two candidate outputs. The perturbation coding vector corresponding to the controlled perturbation label is concatenated with the time window mean of the past decomposable hybrid outputs and then input into the linear layer to generate the predictor score. The predictor score is exponentially normalized to obtain the predictor hybrid weight. The candidate outputs are weighted and summed according to the predictor hybrid weight to obtain the hybrid output. The evidence confidence vector is input into the linear layer and gating vector is obtained by the Sigmoid function. The gating vector is copied by time window index and multiplied element-wise with the mixed output. The graph embedding vector is obtained by global average pooling and linear mapping. The multi-head linear layer output is divided into a set of quantized error parameters corresponding to the zero-point bias component, proportional gain component, short-term noise component and transient recovery component.

6. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step five specifically includes: Using the time window index as the key, the original sample set of the cause-error response map is aligned and fused with the map embedding vector to construct the cause-error response map. For each response edge, consistency and sensitivity indices are calculated. The consistency index is obtained by reading the same error component parameter of multiple time windows under the same disturbance type and intensity level, averaging the values ​​according to the evidence credibility vector, and summing the deviations of each time window parameter from the center value according to the evidence credibility vector. The sensitivity index is obtained by reading the same error component parameter from low to high intensity level under the condition that the disturbance type remains unchanged, calculating the amplitude of the difference between parameters of adjacent intensity levels, and summing them up by weighting according to the evidence credibility vector. Stable invariants are selected based on consistency and sensitivity indices. The stable invariants are response edges where the consistency index is better than a preset consistency threshold and the sensitivity index falls within a preset sensitivity range. For each stable invariant, the set of quantified error parameters of the corresponding error component is read and averaged according to the evidence credibility vector to obtain the compensation base value. At the same time, the evidence credibility vector of the corresponding cause field is read to generate a gating coefficient. The gating coefficient is obtained by linear mapping of the evidence credibility vector and through the Sigmoid function. The gating coefficient is multiplied by the compensation base value to obtain the gating compensation term. An index is established according to the cause field, error component, disturbance type, and intensity level and written into the cause-error response spectrum. A minimum calibration generation set is constructed based on stable invariants. The minimum calibration generation set is a set of entries that satisfy coverage constraints. The coverage constraints include selecting at least one stable invariant for each error component and at least one intensity level for each disturbance type.

7. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step six specifically includes: Uncertainty allocation is performed on the gating compensation term and the fractional error parameter set based only on the entries contained in the minimum calibration generation set, forming a set of uncertainty budget entries; For each error component, calculate the compensation uncertainty and the induced propagation uncertainty, sum the squares of the compensation uncertainty and the induced propagation uncertainty and take the square root to obtain the combined uncertainty of the current error component, and write the combined uncertainty of each error component into the uncertainty budget item set. A dynamic validity period boundary is generated based on the uncertainty budget item set. The combined uncertainty of the current time window is updated sequentially along the time window index covered by the minimum calibration generation set. The combined uncertainty is compared with the preset allowable error threshold. When the combined uncertainty exceeds the allowable error threshold, the corresponding time window index is marked as the dynamic validity period termination boundary. When the combined uncertainty does not exceed the allowable error threshold, the dynamic validity period termination boundary is updated by recursively pushing forward according to the time window index. When the sampling integrity sub-confidence in the evidence confidence vector of the corresponding time window in the minimum calibration generation set is lower than the preset integrity threshold, the current time window index is marked as the validity period degradation point. The set of uncertainty budget entries and the dynamic validity period boundary are combined to generate a calibration result package, and the calibration result package is associated with the time index of the metrological state field.

8. The calibration and measurement method for electrical measuring instruments based on intelligent sensors according to claim 1, characterized in that, Step seven specifically includes: An online self-test segment is inserted within the dynamic validity period boundary of the calibration result package, and the original measurement output sequence, instrument working status data and intelligent sensor array data are collected simultaneously to update the metrological state field and update the causal evidence vector and evidence credibility vector. The original measurement output sequence within the online self-test segment coverage time window is subtracted point by point to obtain the self-test quantity residual sequence. The self-test quantity residual sequence is decomposed into zero-point bias component, proportional gain component, short-term noise component and transient recovery component, and then spliced ​​in the order of error components to form the error component change fingerprint. The error component change fingerprint is subtracted from the fractional error parameter set component by component, and the matching degree is obtained by weighting and summarizing according to the gating coefficient. The cause field corresponding to the entry with the best matching degree is selected as the dominant cause of drift. Selectively update the gating compensation term or the fractional error parameter set based on the dominant cause of drift: when the evidence confidence vector of the corresponding item meets the preset confidence threshold and the consistency index is better than the preset consistency threshold, update the compensation base value of the gating compensation term with the self-test component parameter value; otherwise, update the fractional error parameter set with the self-test component parameter value and write back the cause-error response spectrum. After the update is completed, recalculate the uncertainty budget based on the update result and update the dynamic validity period boundary simultaneously to form the updated calibration result package.

Images