X-ray weighing self-calibration method and system based on physical constraints and timing compensation

Abstract

The application discloses an X-ray weighing self-calibration method and system based on physical constraints and timing compensation, generates physical constraint source domain data according to Beer-Lambert attenuation relationship and equipment, material and environmental parameters, trains a physical basic estimator, and trains a causal timing residual estimator with the difference between a reference mass flow and a physical basic mass flow. When deployed on site, the above-mentioned main model is frozen, dimensionless attenuation features are constructed using an empty load response benchmark and a real-time response, the physical basic mass flow and the timing residual are added, and only a small amount of reference measurement segments are used to fit installation level proportion and bias parameters under the cumulative mass consistency constraint. The communication, spike, saturation and empty load gating state limit benchmark and parameters are written, candidate versions are switched after being written in isolation, independently verified and secondly integrity checked, and instantaneous mass flow and cumulative mass are output.

Description

Technical Field

[0001] This invention relates to the fields of non-contact continuous weighing, industrial sensor calibration, soft measurement and edge computing technologies, specifically to an X-ray weighing self-calibration method and system for weighing bulk materials based on physical constraints and time-series compensation.

[0002] Bulk materials refer to materials that can be transported by belt conveyors or other continuous conveying mechanisms and have a calibrable transmission attenuation relationship under selected X-ray energy spectrum, material thickness and detector dynamic range, including coal, ore and other granular or lumpy materials that meet the above measurement conditions. Background Technology

[0003] Existing X-ray belt scales or nuclear belt scales estimate the load per unit length based on the attenuation signal of X-rays after passing through the conveyor mechanism and material. This is then combined with the conveyor speed to calculate the mass flow rate, which is integrated to obtain the cumulative mass. Engineering calibration typically employs a fixed no-load response benchmark, static or dynamic physical calibration, fixed empirical curves, linear or nonlinear regression, or training a machine learning model using multi-source sensor data. Some devices update the no-load response benchmark when the conveyor mechanism is unloaded, or use belt power, weighing output, standard materials, or source strength feedback to correct the benchmark and slope; other devices use neural networks or multi-sensor fusion to compensate for measurement errors.

[0004] When X-ray source output, detector response, installation geometry, transport speed, environment, and material composition all change, fixed models, single-reference calibrations, or purely data-driven regression models struggle to characterize material attenuation, slowly varying hardware drift, and installation-level scaling bias. Changes in X-ray source-material-detector geometry, detector response, collimation conditions, or electrical range alter the scaling and bias of the original calibration curve; direct migration across equipment or installations results in cumulative quality bias.

[0005] Source aging and detector drift alter the no-load response. A fixed no-load response baseline cannot keep up with this change; when updating only based on the high-order value of the received response, thin or low-load samples may be misclassified as no-load and included in the no-load response baseline update. There may be missing measurement periods for moisture, particle size, ash content, or material type. Using a single statistic to fill in the gaps can confuse measured values ​​with filled values; changes in the missing mechanism alter the model input distribution.

[0006] The number of reference tags for target installations is small and time-dependent. When fine-tuning a high-degree-of-freedom model on a small number of target installation samples, equipment bias, tag noise, and local load distribution all contribute to parameter updates. Industrial field reference quality tags come from traceable dynamic scales, static weighing, batch totals, or other reference measuring devices. The data collection process requires organizing materials, synchronizing time, and verifying reference quantities. It is difficult to continuously obtain a large number of tags covering all operating conditions in the field; when only a small number of segments are available, training a high-degree-of-freedom model from scratch or fine-tuning the entire model will cause the parameters to be dominated by local operating conditions. Summary of the Invention

[0007] (a) Technical issues

[0008] This invention provides an X-ray weighing self-calibration method and system based on physical constraints and time-series compensation. It aims to solve engineering problems in existing X-ray weighing technologies, such as installation migration deviation, erroneous update of no-load reference, confusion due to missing auxiliary variables, overfitting with few samples, and risks associated with version updates. Through physical constraint pre-training and master model freezing, strict no-load gating and causal time-series residual compensation, fitting acceptance isolation, and two-stage version submission, it achieves accurate, stable, and traceable continuous weighing self-calibration of bulk materials.

[0009] (II) Technical Content

[0010] To solve the above-mentioned technical problems, the technical solution of the present invention is: an X-ray weighing self-calibration method based on physical constraints and time-series compensation, applied to the weighing of bulk materials, wherein the bulk materials are materials that have a calibrable transmission attenuation relationship with the material mass surface density under at least one X-ray energy response channel, including a model building stage and an on-site reasoning and adaptation stage;

[0011] The model construction phase includes: generating physically constrained source domain data based on the Beer-Lambert attenuation relationship and the ranges of the X-ray source, receiver detector, installation dimensions, and material parameters; calculating the receiver digital response at each source domain sampling time according to the relationship that the receiver digital response equals the product of the no-load response reference and the natural exponential attenuation term, where the exponent of the natural exponential attenuation term is the negative of the product of the effective attenuation coefficient and the unit length load, and calculating the source domain reference mass flow rate according to the product of the unit length load and the delivery speed; and changing the source output, detector gain and bias, effective attenuation coefficient, material state, installation ratio and bias, environmental disturbances, missing auxiliary variables, and communication. At least one of the anomalies is used to form source domain samples across equipment, installation sections, and operating conditions; using dimensionless logarithmic decay characteristics, the product of the dimensionless logarithmic decay characteristics and the conveying speed, and equipment operating status as physical inputs, and using the source domain reference mass flow rate as the fitting target, the physical basis estimator parameters are obtained and the source domain physical basis mass flow rate is obtained; using the difference between the source domain reference mass flow rate and the source domain physical basis mass flow rate as the residual label, a causal temporal residual estimator is trained; using source domain validation fragments that are not involved in fitting and training, the main model composed of the physical basis estimator and the causal temporal residual estimator is validated, and the weights of the validated main model are set to read-only and the model building stage is ended;

[0012] After entering the field inference and adaptation phase, the edge processing device loads and freezes the read-only master model, updates the installation calibration parameters only based on the field reference metrology segment, and performs the following steps:

[0013] S1. Obtain the digital response of the receiving end after X-rays pass through the conveying mechanism and the bulk material it conveys, the conveying speed, and the operating status of at least one device, synchronize them according to the timestamp, and generate data quality markers based on the communication status, range status, and time continuity.

[0014] S2. Identify packet loss, duplication, saturation, spikes, or time anomalies in the digital response of the receiving end to obtain an effective response after anomaly protection; generate an idle gating value based on communication effectiveness, the closeness of the effective response to the idle response reference, attenuation, short window stability, and transmission speed stability; update the idle response reference only when the idle gating value meets the update conditions, freeze the idle response reference when the update conditions are not met, and prohibit abnormal data from participating in the update of the idle response reference and the update of installation calibration parameters;

[0015] S3. Calculate the dimensionless logarithmic decay characteristic based on the no-load response benchmark and the current effective response. Input the dimensionless logarithmic decay characteristic, the product of the dimensionless logarithmic decay characteristic and the conveying speed, and the equipment operating status into the physical basis estimator to obtain the physical basis mass flow rate.

[0016] S4. The multi-source features at the current time and before the current time are combined into a causal time window, which is then input into the causal time series residual estimator to obtain the residual compensation amount relative to the physical foundation mass flow rate. The physical foundation mass flow rate is added to the residual compensation amount to obtain the main model mass flow rate.

[0017] S5. Using at least one of the dimensionless proportional parameter corresponding to the target equipment or target installation section and the bias parameter with the dimension of mass flow rate, construct the installation calibration parameters, and perform affine calibration and non-negative truncation on the mass flow rate of the master model according to the following formula:

[0018]

[0019] Wherein, the subscript d represents the target device or target installation section. This represents the mass flow rate of the master model input to the calibration layer at the t-th valid sampling time. It is a dimensionless proportionality parameter. A bias parameter with the same dimensions as mass flow rate. The final instantaneous mass flow rate after affine calibration and non-negative truncation is represented; the proportional parameter not involved in the field estimation is set to 1, and the bias parameter not involved in the field estimation is set to 0; the final instantaneous mass flow rate is integrated according to the actual time interval between adjacent effective moments to obtain the cumulative mass;

[0020] Furthermore, mutually isolated reference fitting segments and validation segments are obtained through the reference measurement interface; each segment contains at least two valid sampling times with strictly increasing timestamps, and the sum of the actual time intervals between adjacent valid sampling times is greater than 0; under the frozen state of the master model composed of the physical basis estimator and the causal temporal residual estimator, candidate installation calibration parameters are fitted with cumulative mass consistency loss as a necessary constraint; candidate installation calibration parameters are generated only when both the total reference mass value and the total candidate predicted mass value meet the preset minimum mass threshold (1 kg), and the dimensionless proportional parameter is limited to a preset range, which is 0.85 to 1.15; when there is only a cumulative mass label, the bias parameter is fixed at 0, and when there is an instantaneous reference mass flow rate, the absolute value of the bias parameter does not exceed a preset proportion of the median mass flow rate of the loaded master model, which is 5%;

[0021] Candidate installation calibration parameters and their associated configurations are grouped into a candidate version package, and integrity verification is performed on the candidate version package; the candidate version is independently accepted using the verification fragments that were not involved in the fitting; if the acceptance is successful, the running pointer is switched at the boundary of the complete inference cycle to make the candidate version effective; if the acceptance or integrity verification fails, the original effective version is retained.

[0022] Specifically, a candidate version package is composed of the main model version identifier, the idle response baseline update strategy version, candidate installation calibration parameters, the idle response baseline snapshot at the time of candidate formation, standardized statistics, data quality rules, version number, verification pass flag, data length, and a 32-bit cyclic redundancy check (CRC) code. The verification pass flag is initially set to 0, and the 32-bit CRC code covers all candidate version package fields except for the check code field itself. After the initial overall write, the data length, check code, and verification pass flag (0) are read back and verified. After the initial readback and verification pass, independent acceptance is performed using verification segments that were not involved in fitting. The cumulative quality relative error of the candidate versions does not exceed 1.0% and is not greater than that of the effective version. Within a 60-second loaded window, the cumulative quality relative error of the candidate versions is within 1.0%. The proportion of samples where the non-negative truncation becomes 0 does not exceed 1%, the rate of change of the current effective idle response benchmark relative to the snapshot of the idle response benchmark when the candidate is formed does not exceed 4% during acceptance, the verification segment does not contain communication invalidity, saturation, spikes, duplicate frames, time anomalies, or previous value retention flags exceeding 5 output cycles, and the low, medium, and high load layers each contain at least one verification segment, then the acceptance is passed; after acceptance, the verification pass flag is set to 1, the checksum is recalculated according to the complete candidate version packet after setting to 1, and then written and read back as a whole again; the run pointer is switched only at the boundary of the complete inference cycle if the data length, checksum, and verification pass flag 1 are consistent in the second readback; the original effective version is maintained if the acceptance fails or any readback verification fails;

[0023] It also includes risk control steps: statistically analyzing risk indicators according to a preset statistical period; generating risk events when the proportion of abnormal frames exceeds a preset threshold, the installation calibration correction magnitude exceeds a preset empirical quantile, the main model mass flow rate or dimensionless logarithmic decay characteristic exceeds the distribution range of the reference fitting segment or source domain training set, the relative change rate of the idle response benchmark exceeds a preset threshold, or the effective version integrity verification information is inconsistent with the non-volatile version storage area record; risk events trigger reference segment review or new calibration segment selection, and freeze the write enable of installation calibration parameters; after communication anomalies cause the previous value to remain for more than a preset number of periods, pausing the cumulative quality increment and outputting an abnormal status code; risk events themselves do not directly modify the installation calibration parameters;

[0024] Specifically: Risk indicators are calculated every 60 seconds; a risk event is generated when the proportion of abnormal frames exceeds 1%, the installation calibration correction exceeds the 95th empirical quantile of the correction of the reference fitted segment, the main model quality flow exceeds the minimum to maximum range of the reference fitted segment, the dimensionless logarithmic decay feature exceeds the minimum to maximum range of the source domain training set, the relative change rate of the idle response benchmark exceeds 4% within the 60-second maintenance window, or the data length or 32-bit cyclic redundancy check code of the effective version is inconsistent with the record in the non-volatile version storage area; the risk event triggers the reference segment review or the selection of a new calibration segment, and freezes the write enable of the installation calibration parameters; when communication anomalies cause the previous value to be held, it can be held for a maximum of 5 output cycles, and after 5 output cycles, the cumulative quality increment is paused and an abnormal status code is output; the risk event itself does not directly modify the installation calibration parameters.

[0025] Further, S1 and S2 include: recording device identifier, installation identifier, material configuration, device time, edge reception time, frame sequence number, and data quality marker for each frame of data; verifying frame length, checksum, time monotonicity, range, and saturation state; aligning the receiver response, transmission speed, source state, and environmental state according to the timestamp, and generating communication invalid marker, spike marker, duplicate frame marker, previous value hold marker, saturation marker, and time anomaly marker; data with communication invalid, spike, duplicate frame, saturation, time anomaly, or timeout previous value hold markers are not included in the no-load response benchmark update, candidate installation calibration parameter fitting, and candidate version acceptance.

[0026] Furthermore, S2 includes: providing positive value protection for the current effective response after anomaly protection.

[0027]

[0028] In the formula, This indicates the digital response of the aggregated receiver after communication protection, and the unit is digital response unit; This represents the response after positive value protection, with the same units; the constant 1 is the minimum calculated value of the digital code at the receiving end, used to limit the domain of logarithmic operations;

[0029] Calculate the non-negative conservative attenuation and normalized relative deviation based on the effective response after positive protection and the previous no-load response benchmark:

[0030]

[0031] In the formula, Indicates the index of the current output cycle; Indicates time The non-negative conservative decay is a dimensionless quantity. It represents the relative deviation of the current protection response from the previous no-load response benchmark, and is a dimensionless quantity; This represents the baseline of the no-load response at the previous moment; This indicates that the current state is in the protection response phase; This is the logarithmic molecule protection constant, and its unit is consistent with the digital response. Represents the natural logarithm; A vertical bar indicates taking the larger of the values ​​within the parentheses; a vertical bar indicates taking the absolute value.

[0032] Define the basic online update gate as:

[0033]

[0034] In the formula, This indicates that the basic online update gate value is being implemented. Indicates the output cycle index; Indicates the number of samples initiated; Indicates a non-negative conservative decay; Indicates the attenuation threshold; Indicates relative deviation; Indicates the relative deviation threshold;

[0035] When gating is successful, update the no-load response baseline according to the following formula; when gating is unsuccessful, maintain the no-load response baseline from the previous moment:

[0036]

[0037] In the formula, and These represent the current and previous time-to-time no-load response baselines, respectively, in digital response units. Take 0.985; This indicates that the current protection response is active. When the threshold value is 0, the current baseline remains the previous value; when the threshold value is 1, the update is performed exponentially.

[0038] The current valid response positive protection value represents the larger of the receiver's valid digital response after communication verification, spike protection, and saturation check, and 1. The previous time-of-flight no-load response reference is the reference value read from the non-volatile memory at the beginning of the t-th sampling period. The non-negative conservative attenuation is the dimensionless logarithmic attenuation of the reference before the update relative to the current protection response, and the normalized relative deviation is the dimensionless proportion of the difference between the two relative to the previous time-of-flight reference. The basic online update gating value is the write enable flag issued by the processor to the no-load response reference storage address; a value of 1 indicates that a reference update is performed, while a value of 0 indicates that writing is prohibited and the previous time-of-flight reference is maintained. The starting sample number represents the number of consecutive valid no-load samples required for initialization. The two gating thresholds limit the non-negative conservative attenuation and the normalized relative deviation, respectively. The reference update coefficient is a slowly varying weight greater than 0 and less than 1, and the time index represents the valid sampling sequence number. Communication invalidity, packet loss, timeout, previous value retention, saturation, or spike marking all cause the write enable flag to be 0.

[0039] Furthermore, in an embodiment where the output period is 1 second and the conveying mechanism is a belt conveyor, the no-load response benchmark is initialized using the median of explicit no-load valid samples, and the initial sample number, gating threshold, and update coefficient are:

[0040]

[0041]

[0042] In the formula, Indicates the initial no-load response baseline; to This indicates the positive protection response obtained continuously during the explicit no-load phase; Taking 60 corresponds to a 60-second explicit no-load stable duration under a 1-second output cycle; Represents the median operator; This represents the baseline update coefficient, set to 0.985; the output period is... seconds, Divide 60 by The rounded-up value of the obtained numerical value, where This indicates the time interval between adjacent output moments, in seconds.

[0043] Furthermore, assuming the aforementioned basic online update gating condition is met, the product of the communication validity condition, response proximity condition, low attenuation condition, short window normalized fluctuation condition, stable transmission speed condition, and external no-load confirmation condition is set as the multi-source extended gating factor:

[0044]

[0045] In the formula, Indicates the multi-source extended gating factor; , , , , and These represent, in order, communication, near-response, low attenuation, short-window fluctuation, stable transmission speed, and external confirmation conditions;

[0046] The product of the basic online update gate value and the multi-source extended gate factor is used as the final gate value:

[0047]

[0048] When no external no-load confirmation signal is configured, the external no-load confirmation condition is set to 1; the final gate value enters the update state after 5 consecutive output cycles of 1, and exits the update state when any gate condition is not met for 2 consecutive output cycles; the previous value caused by communication abnormality is maintained for a maximum of 5 consecutive output cycles, and the final gate value is forcibly set to 0 when it exceeds 5 output cycles.

[0049] Furthermore, the anomaly protection includes: identifying spikes using the moving median and absolute median difference of the most recently communicated valid raw response values.

[0050]

[0051]

[0052] In the formula, Indicates time The absolute median difference; Indicated by time A set of causal valid indexes ending with a specific endpoint; Represents a set The sample index within; Indicates index The corresponding original digital response from the receiving end; This indicates the current original digital response from the receiving end; Indicates the index of the current output cycle; and These represent the causal window middle and absolute median difference, respectively; peak value. The debugging data can be selected, preferably 3 to 8; It is a numerically stable quantity; To quantify the amount of compensation.

[0053] Furthermore, S3 constructs a dimensionless logarithmic attenuation characteristic based on the following X-ray attenuation relationship:

[0054]

[0055]

[0056] In the formula, Indicates time The receiver's digital response, in units of digital response units; This indicates the no-load response reference used at the same time. Represents the natural exponential function; This indicates the load per unit length, expressed in kilograms per meter. Indicates the sampling time index; Indicates time Dimensionless logarithmic decay characteristic; This represents the effective attenuation coefficient, expressed in meters per kilogram. Representing the natural logarithm, this formula converts the device response scale into a dimensionless decay principal quantity;

[0057] The physical input consists of the dimensionless logarithmic decay characteristic, the product of the dimensionless logarithmic decay characteristic and the conveying speed, and the equipment operating status. A regularized physical estimator is then used to obtain the physical baseline mass flow rate.

[0058]

[0059]

[0060] In the formula, Indicates time The physical input vector; the constant 1 is the intercept term; and The product represents the attenuation-velocity coupling amount; Indicates the conveying speed; Indicates the tube voltage; Indicates tube current; Indicates the source casing temperature; Represents environmental observations; Indicates the granularity of the proxy quantity; Indicates missing auxiliary quantity markers. This represents the parameter that minimizes the objective function; Represents the optimization variable. The cap at the top indicates the parameters of the physical basis estimator obtained through the objective function; Represents the reference mass flow vector of the source domain training samples, in kilograms per second; Represents the source domain physical input matrix. The tilde at the top indicates the matrix after standardization using training statistics from the source domain; the double vertical lines with the subscript 2 indicate the Euclidean L2 norm, and the 2 in the upper right corner indicates the square of the L2 norm; the L2 regularization coefficient... Take 10.0; Indicates by parameters Determined physical basis mapping; Indicates time The physical input vector; Indicates time The physical basis of mass flow rate is expressed in kilograms per second.

[0061] In this context, the material load in the attenuation relationship represents the mass per unit conveying length calculated based on the effective conveying width and installation geometry. The effective attenuation coefficient absorbs the influence of material composition, X-ray energy spectrum, installation geometry, and effective conveying width. Pipe voltage, pipe current, temperature, humidity, material or composition coding, and missing auxiliary variable markers constitute the equipment auxiliary status. The physical input vector consists of a constant term, a dimensionless logarithmic attenuation characteristic, the product of the characteristic and the conveying speed, the conveying speed, and the equipment auxiliary status. The source domain normalization design matrix is ​​composed of the physical inputs at each source domain sampling time. The source domain reference mass flow vector is composed of the corresponding traceable reference measurement values. The physical estimation parameters are obtained by fitting with a L2 regularization term. The physical estimation function maps the current physical input to a physical basic mass flow with mass flow dimension. The coefficients corresponding to the dimensionless logarithmic attenuation characteristic and the product of the conveying speed are set to non-negative.

[0062] Furthermore, the auxiliary variable is ash content; when ash content cannot be continuously obtained online, ash content statistics from historically confirmed data of the training domain or the current material configuration are used for filling, and the filled values ​​and corresponding missing labels are jointly input into the physical basis estimator and the causal time-series residual estimator. The relationship between the missing labels and the filling is as follows:

[0063]

[0064]

[0065] In the formula, Indicates time Gray missing marker; This represents the ash content value fed into the model; Indicates time The effective ash content measurement value; This represents the median ash content calculated solely from historical values ​​confirmed in the source domain training set or the current material configuration.

[0066] Furthermore, the causal temporal residual estimator only accesses the feature window at the current time and times prior to the current time. The training objective is the residual between the reference mass flow rate and the physical basis mass flow rate, and the residual compensation amount and the master model mass flow rate are obtained according to the following formula:

[0067]

[0068] In the formula, Indicates time Unmodeled mass flow residuals, in kilograms per second. This represents the residual estimate; The parameter is A causal time-series residual estimator; Indicated by time The endpoint is and the length is Standardized feature windows; Take 60, which corresponds to 60 seconds; and These represent the standard deviation and mean of the source domain residual labels, respectively. This represents the mass flow rate of the main model. This represents the mass flow rate estimate of the main model; This represents the physical base mass flow rate. The residual label is fixed as the reference mass flow rate minus the physical base mass flow rate.

[0069] The standardized feature sequence consists of multi-source features from the current time and previous times. The window length represents the number of consecutive valid sampling times or the number of masked historical times. The causal temporal residual function takes the standardized feature sequence as input and outputs standardized residuals. The standard deviation and mean of the residual target are used as the scaling parameter and center parameter, respectively, to restore the standardized residuals to a residual compensation quantity with mass flow dimension. The physical basis mass flow is added to the residual compensation quantity to form the master model mass flow at the t-th valid sampling time. The multi-source features include at least one of the following: receiver response, dimensionless logarithmic decay feature, transmission speed, the product of the feature and the transmission speed, source state, environmental state, equipment auxiliary state, missing marker, and physical basis mass flow.

[0070] Furthermore, the causal temporal residual estimator is a temporal convolutional network. Each residual block includes two layers of causal one-dimensional convolution, nonlinear activation, Dropout, and residual connections. The dilation rate of each residual block increases. When the number of channels is mismatched, a 1×1 convolutional projection is used. The operational relationship of the residual blocks is as follows:

[0071]

[0072] In the formula, and They represent the first Input and output characteristics of each residual block; Take 0 to -1, number of residual blocks Take 4; This represents a causal convolution branch with a kernel width of 3 and dilation rates of 1, 2, 4, and 8 respectively. This indicates a shortcut branch for dimension matching; This represents the linear rectified activation function. The number of network channels is set to 48, the random inactivation rate is set to 0.05, and the time window step size is set to 4.

[0073] The residual block input represents the temporal features received by the i-th residual block, the main transform branch represents the transform consisting of the two layers of causal one-dimensional convolution, nonlinear activation and Dropout, and the shortcut branch is an identity mapping when the number of channels is consistent and a 1×1 convolution projection when the number of channels is inconsistent. The output of the main transform branch is added to the output of the shortcut branch and then processed by ReLU to obtain the input of the next residual block. Each residual block only uses the input at the current time and before the current time.

[0074] Furthermore, the low-degree-of-freedom installation calibration parameters include at least a proportional parameter and an offset parameter corresponding to the target device or target installation section, and the final instantaneous mass flow rate is obtained according to the following formula:

[0075]

[0076] In the formula, This represents the final instantaneous mass flow rate, expressed in kilograms per second. This represents the calibrated estimated value; This represents the mass flow rate output by the frozen master model. This represents the main model estimate; Indicates installation The dimensionless proportionality parameter is limited to 0.85 to 1.15; Indicates installation The bias parameter, in kilograms per second; This indicates non-negative truncation, which is fixed only when using cumulative quality labels. =0; when instantaneous reference mass flow rate is available, The absolute value is not greater than 5% of the median mass flow rate of the loaded principal model in the fitted segment;

[0077] Using the predicted mass flow rate of the master model within each reference measurement segment as input to the candidate installation calibration parameters, the candidate calibration output and the candidate predicted cumulative mass are calculated according to the following formulas:

[0078]

[0079]

[0080] In the formula, Indicates the time within the reference segment The candidate instantaneous mass flow rate estimates, in kilograms per second; This indicates that the mass flow rate estimate output by the main model is frozen at the same time, in kilograms per second. and These represent the candidate installation ratio parameter and the offset parameter, respectively; Indicates the candidate version in the reference fragment The predicted cumulative mass is in kilograms. Represents the set of valid sampling times; Represents a set The summation time index within; This represents the actual time interval between adjacent valid sampling moments, in seconds. The fitting of candidate installation calibration parameters requires the following cumulative quality consistency loss as a necessary term:

[0081]

[0082] In the formula, This represents the cumulative quality estimate of the candidates; This represents the cumulative relative deviation of the fragment's quality and is a dimensionless quantity. Indicates the reference cumulative mass, in kilograms; stable denominator Take the larger of the reference cumulative mass multiplied by 10 to the power of -6 and 1 kilogram, in kilograms; the denominator is actually the reference cumulative mass and... The larger of the values ​​in the range; stop generating candidate parameters when the total cumulative mass value or the predicted total cumulative mass value is less than 1 kg.

[0083] Furthermore, the installation calibration correction magnitude is calculated according to the following formula; a set of verification residuals is constructed using verification segments that did not participate in the fitting of candidate installation calibration parameters, wherein the verification residuals are the difference between the reference mass flow rate and the final predicted mass flow rate:

[0084]

[0085]

[0086] In the formula, the calibration correction magnitude represents the absolute value of the difference between the final instantaneous mass flow rate and the master model mass flow rate, in kilograms per second; the validation residual represents the difference between the reference mass flow rate and the final instantaneous mass flow rate at the same validation time, in kilograms per second; the subscript t represents the validation time index. Based on the set of absolute values ​​of validation residuals that did not participate in the fitting of candidate installation calibration parameters, the prediction interval is constructed according to the following formula:

[0087]

[0088] The calibration correction magnitude is the absolute value of the difference between the final instantaneous mass flow rate and the master model mass flow rate, and has the dimension of mass flow rate. The validation residual is the difference between the reference mass flow rate and the final instantaneous mass flow rate at the same validation moment, and has the dimension of mass flow rate. The empirical quantile is calculated from the absolute value of the validation residual that did not participate in the fitting of the candidate installation calibration parameters, and the upper and lower limits of the resulting prediction interval have the same dimension as the mass flow rate. Risk indicators are calculated every 60 seconds: the calibration correction magnitude exceeds the 95% empirical quantile of the correction magnitude of the reference fitted segment, the master model mass flow rate exceeds the range from the minimum to the maximum value of the reference fitted segment, the dimensionless logarithmic decay feature exceeds the range from the minimum to the maximum value of the source domain training set, the proportion of abnormal frames exceeds 1%, the relative change rate of the idle response benchmark exceeds 4% within the 60-second maintenance window, or the data length or 32-bit cyclic redundancy check code of the effective version is mismatched, triggering a risk event. Before obtaining the reference measurement value, the low, medium and high load layers are divided according to the 33.3% and 66.7% empirical quantiles of the current predicted mass flow rate, and reference measurement candidate segments with valid data quality labels and a start time interval of not less than 60 seconds are selected in each load layer.

[0089] Furthermore, the X-ray response channel is a dual-energy or multi-energy detection channel. Multiple dimensionless logarithmic decay features are constructed based on the no-load response benchmark and the current effective response of each energy channel. These multiple dimensionless logarithmic decay features are input into the physical basis estimator, which calculates the physical basis mass flow rate. Each energy channel that constructs the dimensionless logarithmic decay feature has a calibrable relationship with the surface density of the target bulk material.

[0090] Another aspect of the present invention provides an X-ray weighing self-calibration system based on physical constraints and timing compensation, comprising:

[0091] A conveying mechanism that carries and continuously conveys bulk materials, an X-ray source and a receiving detector arranged on opposite sides of the conveying mechanism, a conveying speed sensor that measures the conveying speed of the conveying mechanism, a reference measurement interface connected to a reference measurement device, and an edge processing device that is connected to the receiving detector, the conveying speed sensor and the reference measurement interface respectively.

[0092] The edge processing device includes a processor, a memory, and a communication interface. The memory includes a volatile working memory area and a non-volatile version memory area. The volatile working memory area is configured with an input circular buffer. The non-volatile version memory area is configured with non-overlapping active version logical address areas, candidate version logical address areas, historical version logical address areas, and a running pointer record area. The active version logical address area stores the no-load response benchmark, physical basis estimation parameters, causal time-series residual model, installation calibration parameters, and data quality rules. The candidate version logical address area stores candidate installation calibration parameters and their associated master model version identifier, gating threshold, standardized statistics, data quality rules, and checksum.

[0093] The processor executes program instructions from the memory, forming:

[0094] The acquisition synchronization and anomaly protection unit receives the timestamped digital response, transmission speed and equipment auxiliary status of the receiving end through the communication interface and writes them into the input circular buffer. It marks packet loss, duplicate frames, spikes, saturation and time anomalies, and prevents frames with anomaly marks from being written into the idle response benchmark and candidate parameter fitting data.

[0095] The gating and no-load response reference update unit generates a write enable flag based on the communication status, the closeness of the received response to the no-load response reference at the previous moment, the non-negative conservative attenuation, the normalized relative deviation, and the stability of the transmission speed. It only modifies the stored value of the no-load response reference when the write enable flag is valid.

[0096] The physical basis estimation unit constructs a dimensionless logarithmic decay characteristic and its product with the conveying speed based on the no-load response benchmark, the current effective response and the conveying speed, and calculates the physical basis mass flow rate.

[0097] The causal temporal residual unit reads the current time and the multi-source feature windows before the current time from the input circular buffer and calculates the residual compensation amount with the dimension of mass flow rate;

[0098] The fusion and installation calibration unit adds the physical basis mass flow rate and the residual compensation amount, multiplies the resulting master model mass flow rate by the proportional parameter of the target equipment or target installation section and adds the bias parameter, compares it with 0 and takes the larger value as the final instantaneous mass flow rate, and integrates it according to the actual time interval between adjacent effective moments to form the cumulative mass.

[0099] The candidate parameter calculation and verification unit calculates candidate proportional parameters and candidate bias parameters using field reference fitting segments while the main model parameters are frozen. It writes the candidate version package into the candidate version logical address area and performs integrity verification. Then, it uses the verification segments that did not participate in the fitting to recalculate the cumulative quality error, load coverage, and parameter boundaries of the candidate version and the effective version.

[0100] The version switching and rollback unit only switches the running pointer after a complete inference cycle ends and before the version snapshot is read in the next inference cycle when both the integrity verification and the verification fragment verification of the candidate version package pass. If the verification fails, the running pointer is kept pointing to the original effective version. If the running is abnormal, the running pointer is switched to the most recently verified historical version.

[0101] At the beginning of each inference cycle, the processor reads the running pointer and latches the corresponding version identifier. During this cycle, the idle gating, physical basis estimation, timing residual calculation, installation calibration, and cumulative mass increment all use the same version snapshot. During the candidate version writing, the communication interface continuously receives sensor frames, and the candidate version writing does not overwrite the effective version logical address area or clear the input ring buffer.

[0102] Furthermore, the system adopts a device-edge-platform architecture; the device includes the X-ray source, the receiving detector, the conveying mechanism, and the conveying speed sensor; the edge includes an edge processing device and runs a verified read-only master model; the platform includes a physical constraint source domain construction and offline pre-training module, a model and configuration version storage module, a log risk and quality verification recording module, and a metrological traceability and manual review module. The platform sends the read-only master model and configuration with version identifier and verification code to the edge. The edge continues to perform sensor acquisition, real-time inference, and anomaly protection when the platform is offline or intermittently connected.

[0103] Furthermore, the acquisition synchronization and anomaly protection unit, the gating and idle response benchmark update unit, and the candidate parameter calculation and verification unit share communication invalid flags, spike flags, saturation flags, duplicate frame flags, time anomaly flags, and previous value retention flags; when any anomaly flag is valid, the gating and idle response benchmark update unit outputs a prohibited update status, and the candidate parameter calculation and verification unit disables the write enable of the candidate installation calibration parameter address; the previous value retention lasts for a maximum of 5 consecutive output cycles, and after 5 output cycles, the cumulative quality increment is paused and an anomaly status code is output.

[0104] Furthermore, the version switching and rollback unit assembles a candidate version package from the main model version identifier, the idle response baseline update strategy version, installation calibration parameters, idle response baseline snapshot during candidate formation, standardized statistics, data quality rules, version number, verification pass flag, data length, and 32-bit cyclic redundancy check code. The check code covers all fields except the check code field itself. The candidate version package is first written to the candidate version logical address area with the verification pass flag 0 and read back for verification. After the verification fragment passes, the flag is set to 1, the check code is recalculated, and the package is written back to the whole package for the second read back for verification. The running pointer is switched only when the data length, check code, and flag 1 are consistent in the second read back. The non-volatile version storage area retains at least one historical version with a flag of 1 and whose length and check code have passed verification. When the effective version mismatches, the most recent valid historical version is selected in descending order of version number. If no valid historical version exists, the metering output is frozen and an invalid status code is recorded.

[0105] (III) Technical Effects

[0106] The advantages of this invention compared to the prior art are:

[0107] 1. Installation-level few-sample adaptation mechanism based on physical constraint source domain generation and master model freezing: During the model construction phase, physical constraint source domain data is generated according to X-ray attenuation relationships and equipment, installation, material, environmental, and anomaly parameters. The source domain is used to train a high-degree-of-freedom physical fundamental estimator and a causal temporal residual estimator. The target installation segment only reads the master model validated and solidified by the source domain. After obtaining 5, 10, 15, or 20 reference measurement segments on-site, the processor keeps the master model weights read-only and only calculates the installation-level proportional parameters and bias parameters. The installation parameter fitting is simultaneously constrained by parameter value boundaries, cumulative quality consistency loss, isolation of fitted and validated segment indices, and candidate version acceptance conditions. A small number of on-site reference segments cannot rewrite the temporal convolution kernel or physical fundamental coefficients; equipment proportional differences and bias differences are confined within the installation-level parameter storage area; cumulative quality constraints limit the accumulation of unidirectional instantaneous errors in continuous integration. This mechanism organizes physical constraint source domain training, target installation low-degree-of-freedom calibration, and running versions into a defined measurement chain.

[0108] 2. Cumulative quality target directly constrains integral deviation: Both installation parameter fitting and candidate acceptance use cumulative quality formed according to actual time intervals. Instantaneous errors in the same direction caused by proportional parameters or bias parameters will be retained in the cumulative amount of segments and cannot be masked by averaging positive and negative instantaneous errors; the 1 kg stable denominator and the minimum effective duration condition of segments eliminate the zero denominator.

[0109] 3. The no-load response baseline and installation parameters adopt independent write permissions: the basic gating and multi-source extension gating only control the no-load response baseline; the installation parameters are only generated from the reference fitted segment. In the event of communication abnormalities, loaded thin layers, or unstable speed, the baseline address remains the previous value; fluctuations in the no-load response will not directly change the installation ratio or bias.

[0110] 4. Two-stage submission of candidate versions: After the verification flag changes from 0 to 1, the 32-bit cyclic redundancy check code covering all stored fields is recalculated, and a second overall write and read-back are performed. The run pointer only points to versions with a verification flag of 1 and which pass the second integrity check. Verification mismatches caused by candidate write interruptions or flag changes will not enter the real-time metering chain.

[0111] 5. Separation of source-end status input and source-end control: Tube voltage, tube current, and source casing temperature are input into physical input or risk assessment, but no adjustment commands are output to the X-ray source from the edge. This signal direction maintains the separation of the weighing estimation chain and the X-ray source control chain, avoiding the technical route of maintaining stability by adjusting the X-ray source through feedback.

[0112] 6. Suppression of Common Proportional Gain Variation by Dimensionless Attenuation: The processor constructs a dimensionless logarithmic attenuation using the unloaded response reference and the loaded received response of the same receiving channel. When the X-ray source output intensity, detector gain, or signal conditioning gain produces approximately the same proportional change to the unloaded and loaded responses, this common proportionality factor cancels out in the response ratio. The online unloaded response reference tracks the slowly varying response according to basic gating and multi-source extended gating, which reduces the impact of absolute digital response level variations on attenuation calculation. The received response first performs positive value protection, then non-negative conservative attenuation, and finally limits the logarithmic output; zero response, negative response, or loaded response higher than the unloaded response will not produce a non-finite attenuation value. During evaluation, the source intensity coefficient, detector gain, and unloaded response reference are changed respectively, and the attenuation deviation, physical basis mass flow rate deviation, and number of non-finite outputs are recorded.

[0113] 7. Missing Value Awareness Encoding Maintains Temporal Input Chain Continuity: When ash, moisture, or particle size auxiliary quantities are missing, the processor uses the median of the source domain training set to generate a filler value, while simultaneously setting a corresponding binary missing value label. The filler value maintains a fixed input vector dimension at each sampling time, and the missing value label enables the causal temporal residual estimator to distinguish between valid measurements and filler values. This process prevents missing values ​​from propagating into the convolutional window as non-finite values ​​and limits the interpretation of filler values ​​as residual offsets resulting from actual coal quality measurements.

[0114] 8. Abnormal Frame Isolation Limits Fault Propagation to Metering Status: Spike discrimination uses a causal window consisting of a maximum of 11 valid communication responses preceding the current time. Spike discrimination is not performed if there are fewer than 5 valid responses in the window. Responses that meet the absolute median difference threshold are replaced by the median value in the window. Previous value retention is performed for a single packet loss cycle, with a maximum of 5 consecutive previous value retention cycles. If there is still no valid response in the 6th output cycle, quality flow rate values ​​are stopped. Duplicate frames do not re-enter the cumulative quality integral. Frames marked with communication invalidity, saturation, timestamp anomalies, or previous value retention are not included in the no-load response benchmark update, master model training, installation parameter fitting, and candidate version acceptance. This limits the impact of single-frame communication faults on the no-load response benchmark, installation parameters, and cumulative quality register. Evaluation metrics include anomaly detection rate, on-load write error rate, invalid frame ratio, repeated cumulative count, and continuous output interruption count.

[0115] 9. The time correspondence between actual time interval integration and cumulative quality is maintained: The cumulative quality of valid segments is calculated based on the actual time interval between adjacent valid sampling moments. Frames with duplicate or reversed timestamps are not included in the cumulative quality; segments with adjacent time intervals greater than 1 second are marked as time anomalies and are not included in idle response benchmark updates, installation parameter fitting, or candidate version acceptance. Both fitted and validation segments must contain at least two valid sampling moments with strictly increasing timestamps and a valid duration greater than 0 seconds. This process ensures that candidate parameter loss and acceptance error correspond to the actual valid sampling duration.

[0116] 10. Non-negative truncation and cycle boundary switching maintain consistent output state: The installation-level affine calibration output undergoes non-negative truncation, preventing negative bias or residual overcompensation from generating negative mass flow, and the cumulative mass register will not decrease in reverse due to negative prediction values. The candidate version only switches the run pointer between adjacent full-second inference cycles after the verification flag is set to 1, the 32-bit cyclic redundancy check code is recalculated, and both the second overall write and second readback pass. If candidate acceptance fails, the write is interrupted, or the second verification fails, the original effective version continues to be used to avoid mixing two sets of installation parameters in the same inference cycle. Attached Figure Description

[0117] Figure 1 It adopts a three-layer collaborative architecture of device end—edge end—platform end.

[0118] Figure 2 This involves a process for pre-training the source domain under physical constraints, fitting parameters for the on-site target domain, and independent verification.

[0119] Figure 3 The branches are: physical foundations, missing perception coding, and causal temporal residuals.

[0120] Figure 4 This includes no-load response baseline gating and serial port anomaly protection logic.

[0121] Figure 5 This provides a closed loop for installation-level small-sample calibration, candidate acceptance, version submission, and rollback.

[0122] Figure 6 The calibration results of reference sampling points for lightweight evaluation configuration under controlled simulator mismatch conditions.

[0123] Figure 7 Results of small-sample adaptation and selection of representative reference sampling points for controlled generation of on-site target domains. Detailed Implementation

[0124] The present invention will now be described in further detail with reference to the accompanying drawings.

[0125] Reference Appendix Figure 1 To be continued Figure 7 The device receives responses, transmission speeds, and status variables, which are then sent to the edge. The edge sequentially performs data acquisition protection, gating updates, physical and residual calculations, installation calibration, cumulative integration, and version control. The platform issues a read-only master model and configuration, and receives metering results, risk events, acceptance results, and version hashes.

[0126] The equipment sends the transmission response of the receiver detector, the speed value of the conveying speed sensor, the status values ​​of the environment or state sensors, and the reference segments of the reference measurement interface to the edge end. The X-ray source and the receiver detector are located on opposite sides of the conveying mechanism; the receiving response is below the upper limit of saturation when unloaded, and above the lower limit of noise when the maximum design material layer is reached.

[0127] The platform-side physical constraint source domain construction and offline pre-training module constructs physical constraint source domain data and trains a read-only master model. The model and configuration version storage module saves the model and configuration versions. The log and quality verification record module archives logs, risk and quality records. The metrology traceability and manual review module saves metrology traceability and manual review results. The platform distributes the read-only master model and configuration that have passed integrity and version verification to the edge. The edge uploads metrology results, idle update records, anomaly markers, risk events, reference metrology fragment summaries, candidate acceptance results, and version hashes.

[0128] I. X-ray Measurement Chain and Acquisition Conditions

[0129] The X-rays emitted from the X-ray source pass through the conveyor belt and the bulk material on it before reaching the detector. Single-energy transmission, dual-energy transmission, or detection methods with multiple energy channels can be used. At least one response channel must have a calibrable monotonic relationship with the material's surface density. Dual-energy implementations are constructed separately... and The attenuation characteristics of the two channels are then input into the physical basis estimator.

[0130] When using dual-energy or multi-energy detection channels, each energy channel used to construct the dimensionless logarithmic attenuation characteristic should have a calibrable relationship with the surface density of the target bulk material within the selected energy spectrum, material layer range, and detector linear range; channels that do not meet this condition are not used as the main quantity for measuring attenuation.

[0131] When the equipment is put into operation, the source operating parameters should be selected based on the maximum areal density and the linear range of the detector: the response under no-load conditions should be lower than the upper limit of saturation, and the response under the thickest loaded condition should be higher than the lower limit of noise, while retaining drift margin. The tube voltage, current, integration time, collimator, shielding, and interlocks should be configured according to the actual equipment's rated range and written into the installation configuration; these safety and hardware parameters are determined by the equipment manufacturing data and on-site radiation safety requirements and are not automatically changed by the learning model.

[0132] The sampling interval can be fixed or vary with timestamps. A preferred embodiment outputs measurement results every 1 second; for detectors with higher initial sampling frequencies, robust 1-second statistics can be generated at the edge first. Cumulative quality must use the actual time difference. This avoids integrating based on a fixed number of samples even after packet loss.

[0133] II. Time Synchronization and Quality Marking

[0134] The acquisition module records the equipment identifier, installation identifier, material configuration, equipment time, edge reception time, sequence number, and quality marker for each frame. Different sensors are first sorted by timestamp, and then aligned using nearest neighbor, window aggregation, or a preservation strategy; data exceeding the allowable alignment error retains missing or invalid markers, and unconditional interpolation is prohibited. Reference metrology segments need to be estimated and their time offset relative to the X-ray data recorded separately.

[0135] The system must generate at least the following flags: communication valid (comm_valid), spike (spike_flag), held (previous value held_flag), saturation (saturation_flag), time gap (time_gap), and gate state (gate_state). Data marked as communication invalid, saturated, or held for an extended period must not be used for no-load response benchmarking. Update, install calibration parameter fitting or candidate acceptance.

[0136] The data acquisition synchronization and anomaly protection unit, gating and idle response benchmark update unit, candidate parameter calculation and verification unit, and version switching and rollback unit share communication invalid, spike, duplicate frame, saturation, time anomaly, and previous value retention flags. When any anomaly flag is valid, the processor disables the write enable for both the idle response benchmark and candidate installation calibration parameters; the previous value is retained for a maximum of 5 consecutive output cycles, after which the cumulative quality increment is paused and an anomaly status code is output.

[0137] III. Communication Anomaly Protection and Idle Response Baseline Update

[0138] (a) Serial port packet loss and spike protection

[0139] Let the nearest The median of the valid raw response values ​​of the communications is , for:

[0140]

[0141] In the formula, Indicates time The absolute median difference, in digital response units; Indicated by time The set of causal valid indexes ending with the number of valid points currently available and 11, with a size of not less than 5; Represents a set The sample index within; Indicates index The corresponding original digital response from the receiving end; This represents the median operator.

[0142] The current original response is identified as a spike candidate when the following formula is satisfied:

[0143]

[0144] In the formula, This indicates the current original digital response from the receiving end; Indicates the index of the current output cycle; and These represent the causal window middle and absolute median difference, respectively; peak value. The selection can be made on the debugging data, preferably 3 to 8, and 3 is used in this embodiment; 1.4826 is the normal distribution uniformity scaling coefficient; numerical stability quantity Take 10 to the power of negative 6 digital response units; quantize the compensation amount. Take one digital response unit; if it is strictly greater than the threshold on the right side of the formula, set a spike marker, and use... Replace the current response.

[0145] Among them, the peak candidates can be selected from Replacement or simply marking is then handled by a degradation strategy; the protection output serves as the receiver response for subsequent physical and temporal models. For packet loss, the system can handle up to [number missing] packets per second. The previous value is retained and set within each output cycle. ;Exceed The affected metering output will be frozen or switched to a conservative estimate with an alarm. This applies to both short-term hold and long-term packet loss. Writing to the no-load response baseline is prohibited.

[0146] In the peak criterion, the lower limit of numerical stability with the same dimensions as the response is... To avoid excessively small absolute median difference, the lower limit of quantization noise is... This is used to avoid misinterpreting normal quantitative fluctuations as spikes; both are saved with the device configuration version.

[0147] (II) Initialization of no-load response baseline and gating update

[0148] Explicit no-load data collection is performed before each new device or installation segment goes online. In the preferred embodiment, the initial data collection is taken as... The median of the valid, unsaturated, and positively protected no-load responses is used as the initial value:

[0149] Before calculating the initial value and the runtime gating, positive value protection is first applied to the current receiver response after communication verification and anomaly protection:

[0150]

[0151] In the formula, This indicates the digital response of the aggregated receiver after communication protection, and the unit is digital response unit; This represents the response after positive value protection; constant 1 is the minimum calculated value of the digital code at the receiving end, used to limit the domain of logarithmic operations.

[0152] From this formula onwards, the current response in the subsequent gating and no-load response benchmark update formulas refers to the effective response after positive value protection. The effective response is expressed as a dimensionless digital response, and the two positive value protection constants in the formula are both dimensionless constants.

[0153]

[0154] In the formula, Indicates the initial no-load response baseline; to This indicates the positive protection response obtained continuously during the explicit no-load phase; Taking 60 corresponds to 60 seconds in a 1-second output cycle; Represents the median operator; This represents the baseline update factor, which is set to 0.985.

[0155] After initialization, calculate the non-negative conservative attenuation and normalized relative deviation based on the current protection response and the previous no-load response benchmark:

[0156]

[0157] In the formula, Indicates the index of the current output cycle; Indicates time The non-negative conservative decay is a dimensionless quantity. It represents the relative deviation of the current protection response from the previous no-load response benchmark, and is a dimensionless quantity; This represents the baseline of the no-load response at the previous moment; This indicates the current positive protection response; 10 to the power of negative 6 is the logarithmic numerator protection constant, whose unit is the same as that of the digital response. Represents the natural logarithm; The vertical bar indicates taking the larger value among the quantities within the parentheses; the vertical bar indicates taking the absolute value.

[0158] The basic online update gate is strictly defined as:

[0159]

[0160] In the formula, This indicates that the basic online update gate value is being implemented. Indicates the output cycle index; Indicates the number of samples initiated; Indicates a non-negative conservative decay; Indicates the attenuation threshold; Indicates relative deviation; This represents the relative deviation threshold; basic gating only occurs when the output cycle index reaches [a certain threshold]. The value is 1 when both inequalities are true, and 0 in other cases. The communication, spike, saturation, repeated frames, time anomalies and previous value preservation flags are determined separately by the multi-source extended gating described later.

[0161] The basic online update mode uses the following set of preferred parameters with a 1-second output cycle:

[0162]

[0163] Specifically, positive value protection prevents non-positive responses from entering logarithmic operations; the conservative attenuation is constrained to be non-negative by the maximum values ​​of the inner and outer levels; the relative deviation is normalized to the larger of the previous moment's idle response reference and 1. When the gate value is 1, the processor sends a write enable to the idle response reference storage address and calculates the new value according to the reference update formula; when the gate value is 0, the write enable remains disabled, and the previous moment's reference in the storage address remains unchanged. The gate value is set to 0 when the sampling sequence number does not reach the number of starting samples, any threshold condition is not met, or communication is invalid, saturated, spikes, or timeouts occur.

[0164] Sampling interval is Seconds, start sample count Divide 60 by The rounded-up value of the obtained value corresponds to the 60-second explicit no-load stability time, where This represents the time interval between adjacent output moments, in seconds. When there is status information such as short window, speed control, or external no-load confirmation, multi-source extended gating can be enabled in addition to basic gating, and the relative deviation threshold of the response can be calculated based on the same explicit no-load sequence. Idle attenuation threshold Short-window normalized fluctuation threshold and belt speed stability threshold ; and Take the 99th percentile of the statistic corresponding to at least 60 explicitly unloaded valid samples. Take 0.035, The value is set to 0.02, meaning that the relative rate of change of the conveying speed between adjacent output cycles does not exceed 2%.

[0165] When multi-source extension configuration is enabled, the multi-source extension gating factor is calculated under the premise that the basic online update gating is established; when no external no-load confirmation signal is configured, the external confirmation condition is set to 1.

[0166]

[0167] In the formula, the response is close to the threshold. Take the 99th empirical quantile of the explicit no-load normalized response deviation; short window fluctuation threshold. Take the 99th empirical quantile of the normalized absolute median difference of the explicit 30-second no-load window. Represents the empirical quantile operator, tail probability Take 0.01; Indicates the explicit unloaded sample at time [time]. The aggregation response after communication protection; Indicates the explicit no-load initialization baseline; This represents the set of 29 positive protection responses that are currently active, not saturated, and have not experienced spikes. Represents a set The absolute median difference; ( ) represents a set the median; Takes a single digital response unit; the vertical bar indicates an absolute value. This indicates taking the larger value among the quantities within the parentheses; both thresholds are calculated using at least 60 externally or manually verified explicit unloaded samples.

[0168] The tail probability parameter in the above quantile formula takes a value greater than 0 and less than 1, and its configuration value and determination data are saved with the gating configuration version.

[0169] The multi-source extended gating factor and the final gating value are written as follows:

[0170]

[0171]

[0172] in: This indicates that communication is valid and no spikes, saturation, or timeouts have occurred. express ; express ; This indicates that the short window fluctuation does not exceed ; This indicates that the conveyor belt is running and the speed variation does not exceed [a certain value]. ; The optional external material flow switch, video, upstream coal feeding signal, or manual no-load confirmation is set to 1 if not configured. Indicates the final gating value; This indicates that the basic online update gate value is being implemented. This represents the multi-source extended gating factor; all three are binary quantities. The gating factor enters the update state after five consecutive output cycles of 1; it exits the update state when any condition fails for two consecutive output cycles.

[0173] The final gate value enters the update state after five consecutive output cycles of 1; it exits the update state when any gate condition fails for two consecutive output cycles. The gate symbol in the following formula... This indicates the update gating actually used in the current configuration. The basic configuration uses the basic online update gating value, and the multi-source extended configuration uses the final gating value. The idle response baseline is updated only when the corresponding update gating is valid, using the following formula:

[0174]

[0175] In the formula, and These represent the current and previous time-to-time no-load response baselines, respectively, in digital response units. This indicates the actual update gating value used in the current configuration, when multi-source extended gating is enabled. ; Take 0.985; This indicates the current positive protection response. When the gating value is 0, the current baseline remains the previous value; when the gating value is 1, it is updated exponentially.

[0176] For implementations with multi-source extension mode or output cycle variation, the baseline update coefficient... According to the output cycle With slowly varying time constants Exponential relationship conversion, slow-varying time constant Fixed at 66.17 seconds, so that the output cycle is 1 second. The value equals 0.985; when gating is not valid, the idle response baseline remains at the previous time value. If the cumulative rate of change of a candidate idle response baseline relative to the window start point exceeds 4% within any 60-second maintenance window, the system cancels the candidate baseline within that window, restores the window start point baseline, and generates a risk event. The idle status label during the physical constraint source domain generation process is only used to evaluate gating accuracy and write error rate and is not used as online input.

[0177] The basic gating system uses a fixed starting sample count of 60, a non-negative conservative attenuation threshold of 0.035, a normalized relative deviation threshold of 0.04, and a baseline update coefficient of 0.985 at a 1-second output cycle. The starting sample count of 60 corresponds to a 60-second continuous and effective explicit no-load initialization window. When the output cycle changes, the starting sample count is calculated by dividing 60 seconds by the output cycle and rounding up. The non-negative conservative attenuation threshold and the normalized relative deviation threshold remain unchanged, and the baseline update coefficient is calculated based on a slow-varying time constant of 66.17 seconds. Samples used for initializing and verifying the gating system must be communicatively valid, unsaturated, free of spikes, and confirmed to be no-load by manual or external signals. Samples with load, repeated frames, time anomalies, or previous value hold-up are not included in the initialization or threshold verification. If the relative change rate of the no-load response baseline exceeds 4% within any 60-second maintenance window, the candidate updates within that window are cancelled and the window's starting baseline is restored. Sensor configuration, output cycle, and gating configuration versions are saved together.

[0178] IV. Physical basis estimation, causal temporal residuals and installation-level calibration

[0179] (a) Physical basis mass flow rate estimation

[0180] Within the range of the narrow-beam approximation and the detector's linear response, the received response can be expressed as:

[0181]

[0182] In the formula, Indicates time The receiver's digital response, in units of digital response units; This indicates the no-load response reference used at the same time. Represents the natural exponential function; The effective attenuation coefficient is expressed as a unit of meters per kilogram, which is affected by material composition, source state and installation geometry. This indicates the load per unit length, expressed in kilograms per meter.

[0183] The material load in the above attenuation relationship represents the mass per unit conveying length calculated based on the effective conveying width and installation geometry, with units of mass per length. The effective attenuation coefficient absorbs the effects of material composition, X-ray energy spectrum, installation geometry, and effective conveying width. Therefore, the product of the conveying speed and the mass per unit conveying length has the dimension of mass flow rate.

[0184] Therefore, a dimensionless attenuation characteristic is constructed:

[0185]

[0186] In the formula, Indicates the sampling time index; Indicates time Dimensionless logarithmic decay characteristic; and These represent the no-load response reference and the receiver digital response, respectively, both in digital response units. This represents the effective attenuation coefficient, expressed in meters per kilogram. This indicates the load per unit length, expressed in kilograms per meter. This represents the natural logarithm. This formula converts the device response scale into a dimensionless decay principal quantity.

[0187] Ideal quality flow rate meets ,therefore It has a physically monotonic relationship with mass flow rate. The effective attenuation coefficient varies with material type, composition, source condition, and installation geometry. The current physical input is written as:

[0188]

[0189] In the formula, Indicates time The physical input vector; the constant 1 is the intercept term; Indicates logarithmic decay; and The product represents the attenuation-velocity coupling amount; Indicates the conveying speed; Indicates the tube voltage; Indicates tube current; Indicates the source casing temperature; Represents environmental observations; Indicates the granularity of the proxy quantity; This indicates missing auxiliary labels. Each continuous variable is standardized using source domain training statistics.

[0190] Physical basis estimation can be performed using a regularized linear model:

[0191]

[0192] In the formula, This represents the parameter that minimizes the objective function; Represents the optimization variable. This represents the parameters of the physical basis estimator obtained through the objective function; Represents the reference mass flow vector of the source domain training samples, in kilograms per second; Represents the source domain physical input matrix. This represents the matrix after standardization using the source domain training statistics; Represents the square of the Euclidean L2 norm; L2 norm regularity coefficient Take 10.0; Indicates by parameters Determined physical basis mapping; Indicates time The physical basis of mass flow rate is expressed in kilograms per second.

[0193] The standardized design matrix in the physical estimation formula consists of the physical inputs at each time point in the source domain, the source domain reference mass flow vector consists of the corresponding traceable reference values, the regularization coefficient controls the L2 constraint strength of the physical estimation parameters, and the physical estimation function is used to map the current physical inputs to the physical base mass flow.

[0194] It can also be used with The corresponding coefficients should be set with non-negativity or monotonicity constraints. For scenarios with limited computing power or small data volume, physical ridge / robust regression can be used as the master estimator. When the relationship between physical features and mass flow is non-linear and explicit preservation of physical principal quantities is still required, generalized additive, piecewise linear, or tree models can be used to fit the physical features, but they must not be deleted. , Measurement chain for belt speed and cumulative integration.

[0195] (ii) Multi-source coding lacking perception

[0196] Taking ash auxiliary amount as an example, for Define missing markers:

[0197]

[0198] In the formula, Indicates time The ash content is marked as missing; 1 is used when ash content is missing, and 0 is used when ash content is available. Moisture, particle size, or environmental quantities use the same binary marking rules.

[0199] Calculate fill values ​​using only historically confirmed data from the training domain or the current material configuration. And construct:

[0200]

[0201] In the formula, This represents the ash content value fed into the model; Indicates time The effective ash content measurement value; This represents the median ash content calculated solely from historical values ​​confirmed in the source domain training set or the current material configuration; For gray missing data markers, a segmented approach can prevent missing values ​​from propagating further after multiplication.

[0202] The model jointly receives the filled ash input and missing markers to distinguish between true measurements and statistical filled values. This method can be used for moisture, particle size, ash content, or material composition. If an auxiliary quantity has never been available on the deployed equipment, that numerical channel should be deleted or the missing marker fixed; unobservable true values ​​should not be used as online inputs.

[0203] (III) Causal Temporal Residual Compensation

[0204] The time-series residual model only uses current and historical samples:

[0205]

[0206] In the formula, Indicates time Unmodeled mass flow residuals, in kilograms per second. This represents the residual estimate; The parameter is A causal time-series residual estimator; Indicated by time The endpoint is and the length is Standardized feature windows; Take 60, which corresponds to 60 seconds; and These represent the standard deviation and mean of the source domain residual labels, respectively. This represents the mass flow rate of the main model. This represents the mass flow rate estimate of the main model; This represents the physical base mass flow rate. The residual label is fixed as the reference mass flow rate minus the physical base mass flow rate.

[0207] The standardized feature sequence in the time-series residual formula consists of multi-source features from the current time and previous times, with the window length being the number of consecutive or masked historical time points. The residual network outputs standardized residuals, and the standard deviation and mean of the residual target are used as scaling and centering statistics to restore the standardized residuals to residual compensation quantities with mass flow dimension.

[0208] The training objective residual is defined as the difference between the source domain reference mass flow rate and the source domain physical basis mass flow rate. The raw sensor input is first estimated using the physical basis, and the residual model only calculates the portion not covered by the physical basis estimate. Causal time window length. Take 60 1-second sampling points; window features include receiver response, source state, temperature and humidity, transport speed, material code, missing marker, dimensionless logarithmic decay feature, the product of the decay feature and the transport speed, physical basic mass flow rate, and the median and absolute median difference of the 30-second short window formed by the current time and the previous 29 valid sampling points.

[0209] The preferred time series estimator is TCN. Each residual block contains two layers of causal one-dimensional convolution, nonlinear activation, Dropout, and residual connections. The kernel size is preferably 3, and the dilation rate is... Increment; used when the number of channels does not match. Convolutional projection. A reproducible implementation is configured with a 60-second history window, 48 channels, 4 expanded residual blocks, and a linear residual output head. For lower edge computing power, a 30-second window, 24 channels, and 3 residual blocks can be used.

[0210] The operational relationships of a single causal extended residual block are as follows:

[0211]

[0212] In the formula, and They represent the first Input and output characteristics of each residual block; Take 0 to -1, number of residual blocks Take 4; This represents a causal convolution branch with a kernel width of 3 and dilation rates of 1, 2, 4, and 8 respectively. This indicates a shortcut branch for dimension matching; This represents the linear rectified activation function, with 48 network channels, a random inactivation rate of 0.05, and a time window step size of 4.

[0213] In the residual block relation, the input of the i-th residual block is the temporal feature tensor; the main transformation branch consists of two layers of causal one-dimensional convolution, nonlinear activation and Dropout; the shortcut branch uses identity mapping when the number of channels is consistent, and uses 1×1 convolution projection when the number of channels is inconsistent.

[0214] Causal time-series residual estimators can employ TCN, Ridge based on lag statistics, state-space models, gated recurrent networks, or other time-series regressors that only access current and historical information. Each implementation maintains the causal input, residual objective, and physical main chain; this embodiment uses TCN.

[0215] (iv) Installation-level small sample calibration

[0216] The equipment or installation section d first uses a low-degree-of-freedom affine calibration layer to obtain:

[0217]

[0218] In the formula, Indicates installation At any moment Affine-calibrated but not yet non-negative cutoff mass flow rate estimate; This indicates that the mass flow rate estimate output by the main model is frozen at the same time. Indicates installation The corresponding dimensionless proportionality parameter; Indicates installation The corresponding bias parameter, in kilograms per second; subscript Indicates the target device or target installation section, subscript Indicating the effective sampling time, the affine calibration output is non-negatively truncated according to the following formula:

[0219]

[0220] In the formula, This represents the final instantaneous mass flow rate, expressed in kilograms per second. This represents the calibrated estimated value; This represents the mass flow rate output by the frozen master model. This represents the main model estimate; Indicates installation The dimensionless proportionality parameter is limited to 0.85 to 1.15; Indicates installation The bias parameter, in kilograms per second; Indicates non-negative truncation. Fixed only when using cumulative quality labels. =0; when instantaneous reference mass flow rate is available, The absolute value of the parameter is no greater than 5% of the median mass flow rate of the loaded principal model in the fitted segment. The proportional parameter is dimensionless, and the bias parameter has the same dimensions as the mass flow rate.

[0221] The default initial values ​​for the installation calibration parameters are proportional parameters. Equal to 1, bias parameter It equals 0. It is only fixed when accumulating quality labels. Equal to 0, only estimated When instantaneous reference mass flow rate is available, joint estimation is possible. and . Limited to 0.85 to 1.15; The absolute value does not exceed 5% of the median mass flow rate of the loaded master model in the reference fitted segment; no candidate installation calibration parameters are generated when the total cumulative mass value of the reference or the total cumulative mass value of the candidate prediction is less than 1 kg.

[0222] Each reference measurement segment has a segment reference cumulative mass. For any candidate scale parameter and candidate bias parameter, the candidate calibration output is first calculated by predicting the mass flow rate from the master model, and then the candidate predicted cumulative mass is calculated based on the actual time interval between adjacent valid moments within the segment.

[0223]

[0224] In the formula, Indicates the time within the segment The candidate instantaneous mass flow rate, in kilograms per second. Indicates the candidate estimate; This indicates that the main model output is frozen at the same time. This represents the main model estimate; and These represent the candidate installation ratio and the offset parameter, respectively. This indicates non-negative truncation.

[0225]

[0226] In the formula, Indicates the candidate version in the reference fragment The cumulative mass, in kilograms. This represents the cumulative quality estimate of the candidates; Represents the set of valid sampling times; Represents a set The summation time index is within the range; the summation symbol indicates summation over all... belong Accumulate the effective sampling times; This represents the actual time interval between adjacent valid sampling moments, in seconds.

[0227] The fitting of candidate installation calibration parameters takes the cumulative mass consistency loss as a necessary term:

[0228]

[0229] In the formula, This represents the cumulative relative deviation of the fragment's quality and is a dimensionless quantity. and Represent the candidate cumulative mass and the reference cumulative mass, respectively, in kilograms; stable denominator Take the larger of the reference cumulative mass multiplied by 10 to the power of -6 and 1 kilogram, in kilograms; the denominator is actually the reference cumulative mass and... The larger of the values ​​in the range. Candidate parameter generation stops when the reference cumulative total mass or the predicted cumulative total mass is less than 1 kg.

[0230] The segment reference cumulative mass in the cumulative mass consistency loss is provided by the reference metrology interface, and the numerical stability term is a number with the same dimensions as the segment reference cumulative mass and greater than zero. The first time increment in the summation is based on the starting boundary of the reference metrology segment; if there is no valid time reference for the starting boundary, the summation starts from the second valid sampling time within the segment. Each time increment is converted to units consistent with the mass flow rate time reference. If only batch cumulative mass is available, the candidate predicted cumulative mass of multiple reference metrology segments is directly compared with the segment reference cumulative mass; when instantaneous reference mass flow rate is available, a Smooth L1 point loss is added. The loss weights, parameter boundaries, and stopping conditions are determined by mutually isolated fitted and validation segments.

[0231] (v) Representative Sampling and Reference Label Quality Control

[0232] Before obtaining reference labels, perform label-free predictions on candidate intervals using the current version; divide the load into low, medium, and high load layers based on predicted traffic quantiles, and select segments within each layer that are time-dispersed and whose communication, velocity, idle response benchmarks, and source states are all valid. This can also be done in... The source state, environment, and material coding space can be used to select diverse options, but the predicted load coverage must be checked simultaneously. Continuously truncating the first few points may only cover localized loads and is not a preferred strategy.

[0233] In a preferred embodiment, the calibration budget consists of 5 to 20 reference measurement segments; the specific number is determined by load coverage, reference quantity quality, parameter stability, and independent acceptance. Reference labels need to record the source, traceability information, batch boundaries, time offset, outlier segments, and uncertainty; label noise or time misalignment can directly lead to tail-cumulative bias, which cannot be eliminated simply by increasing the sample size.

[0234] (vi) Risk of unlabeled

[0235] The system can calculate the calibration correction range:

[0236]

[0237] When there are validation segments that did not participate in fitting the candidate installation calibration parameters, the validation residual set is constructed using the difference between the reference mass flow rate and the final predicted mass flow rate:

[0238]

[0239] Construct the following prediction intervals based on the empirical quantiles of the absolute values ​​of the residuals in the verification residual set:

[0240]

[0241] Experience quantile level The empirical quantile is set to 0.95. It is calculated from the absolute value of the validation residuals that did not participate in the fitting of the candidate installation calibration parameters. The risk item of the prediction interval is enabled when the number of valid validation residual samples is not less than 30. If it is less than 30, the prediction interval is not generated and the item is not used as the basis for the candidate version to pass.

[0242] A risk event is formed when any of the following conditions are met: The high quantile threshold of the adapted segment is exceeded; the current prediction is lower than the calibrated minimum or higher than the maximum and exceeds the margin. The following conditions may occur: source state or environmental characteristics exceed training / calibration coverage; communication anomaly rate or retention rate exceeds the upper limit; idle response benchmark has not been reliably updated for a long time or changes abnormally per unit time; prediction interval half-width or residual risk exceeds the threshold.

[0243] Actions for risk events include recording the event, reducing automatic update privileges, switching to conservative physics estimation, requesting reference calibration, selecting candidate segments, or prompting for manual review. Risk events themselves do not have a true quality label, therefore updates should not be made directly based on this. Or residual model.

[0244] (vii) Candidate version verification, runtime pointer switching and rollback

[0245] Candidate installation calibration parameters are written to the candidate version storage area, which does not overlap with the logical address of the effective version. Segments involved in parameter fitting and validation segments are isolated by sample index; each segment contains at least two valid sampling moments with strictly increasing timestamps, and the sum of the actual time intervals between adjacent valid sampling moments is greater than 0; each validation segment lasts for at least 300 seconds; the cumulative mass relative error is calculated for the candidate version and the effective version on the same validation segment, and the stability denominator is taken as the reference cumulative mass. The larger of the values ​​in 1 kg; the error of the candidate version does not exceed 1.0% and is not greater than the error of the effective version.

[0246] Candidate installation ratio parameters Limited to 0.85 to 1.15; bias parameter when only cumulative quality labels are available. When the value is fixed at 0, and instantaneous reference mass flow rate is available. The absolute value of the parameter shall not exceed 5% of the median mass flow rate of the loaded master model; candidate parameters shall only be generated when both the reference cumulative mass total and the candidate predicted cumulative mass total are not less than 1 kg; the low, medium and high load layers shall be divided according to the empirical quantiles of 33.3% and 66.7% of the average mass flow rate of the fitted fragments of the master model, and each layer in the validation set shall contain at least one fragment; the proportion of samples that become 0 after non-negative truncation within a 60-second loaded window shall not exceed 1%; the rate of change of the currently effective unloaded response benchmark relative to the snapshot of the unloaded response benchmark when the candidates are formed shall not exceed 4% at the time of acceptance; the validation fragments shall not contain communication invalidity, saturation, spikes, duplicate frames, time anomalies or previous value hold marks exceeding 5 output cycles.

[0247] After fitting the candidate installation calibration parameters, the processor assembles a candidate version package by combining the main model version identifier, the no-load response baseline update strategy version, the candidate installation calibration parameters, the no-load response baseline snapshot at the time of candidate formation, standardized statistics, data quality rules, version number, verification pass flag, data length, and a 32-bit cyclic redundancy check (CRC) code. The verification pass flag is initially set to 0, and the check code covers all candidate version package fields except for the check code field itself. The processor writes the complete candidate version package with the flag set to 0 to the candidate version storage area, reads it back, and verifies the data length, check code, and flag 0. After the first readback passes, independent acceptance is performed using the verification fragments that were not involved in the fitting. After acceptance passes, the verification pass flag is set to 1, the check code is recalculated based on the complete candidate version package with the flag set to 1, and then written and read back as a whole. Only when the data length, check code, and flag 1 are consistent in the second readback, the processor updates the run pointer after the current complete inference cycle ends and before reading the version snapshot in the next inference cycle. If acceptance fails or any readback fails, the candidate version is marked as invalid, and the original effective version is retained. The sensor acquisition thread continuously writes frames to the circular input buffer during candidate writing, readback, and pointer switching. If the data length or checksum of the effective version is mismatched, the verification pass flag is invalid, or the output contains non-finite values, the program will start to scan historical versions in descending order of version number, and select only historical versions with consistent data length and checksum readback and a verification pass flag of 1; if no valid historical version exists, the metering output will be frozen and an invalid status code will be recorded.

[0248] (viii) Cumulative quality output and measurement traceability

[0249] The final cumulative mass is integrated over the actual time interval between adjacent valid moments:

[0250]

[0251] The first time increment in the summation is based on the initial boundary of the cumulative interval; if there is no valid time reference at the initial boundary, integration begins from the second valid sampling time within the interval. Each time increment is converted to a unit consistent with the mass flow rate time reference; for example, when the mass flow rate is expressed in tons per hour, the time increment is expressed in hours.

[0252] For prolonged periods of packet loss, data interruption, invalid speed, or output freeze, the system must save the status code and employ the project-specified supplementary recording, data stoppage, or conservative estimation strategies; it cannot be silently treated as a normal sample. Each cumulative batch records the device and installation configuration, master model version, and... Sequence, installation calibration parameters, data quality markers, reference calibration events, and risk handling records.

[0253] When only batch cumulative mass is available without instantaneous reference flow, calibration loss omits point loss and compares the predicted cumulative mass with the reference cumulative mass of multiple independent reference metering segments. Segments used for fitting, segments used for selecting hyperparameters, and segments used for acceptance should be isolated according to the amount of available data for the target equipment; acceptance segments must not be used for fitting candidate installation calibration parameters.

[0254] V. Physically Constrained Source Domain Pre-training and On-Site Target Domain Constrained Adaptation

[0255] (i) Reproducible generation of physical constraint source domain data

[0256] 1. Scene, device, and installation mapping

[0257] The generator uses a fixed total random seed of 20260623, setting up scenarios S01 to S04, 8 devices, 16 installation segments, and 57,600 sampling points per second. Each scenario contains 2 devices, each device contains 2 installation segments, and each installation segment continuously generates 3,600 points. The 28,800 points in S01 and S02 are used only for parameter training; the 14,400 points in S03 are used only for source domain verification; and the 14,400 points in S04 are used only for target domain evaluation. The 16 installation segments are traversed by scenario number, device number, and installation number, using 20260623 to 20260638 as installation segment seeds in sequence. Continuous variables and Bernoulli variables use a 64-position transpose congruence generator subflow, and material categories use a Mason rotation algorithm subflow.

[0258] 2. Material and Load Sequence

[0259] The material categories are bituminous coal A, bituminous coal B, anthracite, and lignite, with effective attenuation coefficients of 0.00380, 0.00405, 0.00355, and 0.00435 m³ / kg, respectively; nominal ash content of 12%, 18%, 9%, and 24%, respectively; nominal moisture content of 7.5%, 9%, 5.5%, and 13%, respectively; and standard deviation of particle size hidden disturbance of 0.010, 0.014, 0.008, and 0.018, respectively. One type of material is randomly selected with equal probability for each installation section.

[0260] The load sequence starts with an unloaded segment, the length of which is uniformly sampled from integers between 60 and 239 seconds. The length of the loaded segment is uniformly sampled from integers between 600 and 1799 seconds, and so on, until 3600 points are reached. The load per unit length of the loaded foundation is continuously and uniformly sampled from 45 to 130 kg / m. The sinusoidal amplitude is 18% of the base load, and the termination phase is uniformly sampled from 3π to 8π. The standard deviation of the Gaussian disturbance is 6% of the base load. Each loaded segment is superimposed with 2 to 5 Gaussian local mutations, the time standard deviation of which is uniformly sampled from integers between 20 and 119 seconds, and the peak amplitude is continuously and uniformly sampled from -25% to +35% of the base load. When the superposition result is less than 0, it is taken as 0.

[0261]

[0262] In the formula, Indicates time The load per unit length, expressed in kilograms per meter; This indicates the sampling time index within the currently loaded segment; This indicates the current load on the foundation section, which is continuously and uniformly sampled from 45 to 130 kg per meter; 0.18 represents the amplitude ratio of the sinusoidal ripple to the foundation load. This represents a sinusoidal phase that linearly changes from 0 to the termination phase, with the termination phase continuously and uniformly sampled from 3π to 8π. Represents pi; This represents the number of local mutations, drawn uniformly from integers between 2 and 5. Indicates a local mutation index; Indicates the first A sudden peak was continuously and uniformly extracted from the base load between -25% and +35%. Indicates the first Each mutation center is sampled at the sampling time within the loaded segment, and is discretely and uniformly extracted at all sampling times of the loaded segment. Indicates the first The time standard deviation of each mutation is uniformly sampled within an integer range of 20 to 119 seconds; The standard deviation represents the zero-mean Gaussian perturbation with 6% of the basic load; summation upper and lower limits. =1 and This indicates the summation of all local mutations; Represents the sine function; Represents the natural exponential function; This indicates taking the larger value among the quantities within the parentheses while ensuring the load is not less than 0. For the unloaded section, directly set... Take 0.

[0263] 3. Environment, source status, and sensor response

[0264] The ambient temperature is centered at 18 degrees Celsius, with a sinusoidal amplitude of 8 degrees Celsius and a period of 3600 seconds, superimposed with a random walk of incremental standard deviation of 0.015 degrees Celsius; the ambient relative humidity is centered at 55%, with a sinusoidal amplitude of 18 percentage points and a period of 2880 seconds, superimposed with a random walk of incremental standard deviation of 0.03 percentage points, ultimately cut off between 20% and 95%; the tube voltage is centered at 80 kV, superimposed with white noise of standard deviation of 0.35 kV and a random walk of incremental standard deviation of 0.002 kV; the tube current is centered at 1.5 mA, superimposed with white noise of standard deviation of 0.015 mA and a random walk of incremental standard deviation of 0.0002 mA.

[0265] The transmission speed is centered at 2.4 m / s, superimposed with white noise of 0.025 m / s standard deviation and a sine term with an amplitude of 0.04 m / s and a period of 900 seconds, truncated between 1.8 and 3.0 m / s. The gain of each device is sampled according to a normal distribution with a mean of 1 and a standard deviation of 0.035; the installation factor of each installation section is sampled according to a normal distribution with a mean of 1 and a standard deviation of 0.025; the detector bias is sampled according to a normal distribution with a mean of 0 and a standard deviation of 12; the base no-load digital response is uniformly sampled between 7800 and 9600; the receiver response is superimposed with Gaussian noise of 18 standard deviation, and a result less than 1 is taken as 1.

[0266]

[0267] In the formula, Indicates the sampling time index; Indicates time The generated no-load response is expressed in digital response units. This represents the basic no-load response, which is continuously and uniformly sampled from 7800 to 9600 digital response units. The device gain represents a normal distribution with a mean of 1 and a standard deviation of 0.035. The installation factor is represented by a normal distribution with a mean of 1 and a standard deviation of 0.025. This indicates the tube current, in milliamperes (mA), with the rated current being 1.5 mA in the denominator. The voltage rating is expressed as tube voltage in kilovolts, with 80 kilovolts in the denominator representing the rated tube voltage; the exponent of 1.7 represents the tube voltage response index. express The temperature of the X-ray source casing is in degrees Celsius. 35 degrees Celsius is the temperature reference value, and 0.0018 per degree Celsius is the temperature sensitivity coefficient. This indicates the relative humidity of the environment, expressed in percentage points. 55 percentage points is the humidity reference value, and 0.0004 per percentage point is the humidity sensitivity coefficient. This represents a dimensionless multiplicative random walk with an incremental standard deviation of 0.000018; the constant 1 in each parenthesis is the dimensionless reference factor. The value is 0.80 when the temperature or humidity factor is less than 0.80, and 1.20 when it is greater than 1.20.

[0268]

[0269] In the formula, Indicates the sampling time index; Indicates time The effective attenuation coefficient, expressed in meters per kilogram; This represents the basic attenuation coefficient of the current material, expressed in meters per kilogram. and These represent the current ash content and the nominal ash content of the material, respectively, both expressed as percentage points; the ash sensitivity coefficient is 0.010, and each percentage point represents the change in the multiplicative factor by 0.010 for every 1 percentage point deviation of the ash content from the nominal value. and These represent hidden moisture and nominal moisture content of the material, respectively, both expressed as percentage points; the moisture sensitivity coefficient is 0.006, with each percentage point representing a change of 0.006 in the multiplicative factor for every 1 percentage point deviation of moisture from the nominal value. The tube voltage is represented in kilovolts; 80 kilovolts is the reference value for the tube voltage; the tube voltage sensitivity coefficient is 0.003 per kilovolt, which means that for every 1 kilovolt deviation of the tube voltage from the reference value, the corresponding multiplicative factor changes in the opposite direction by 0.003. Indicates time The dimensionless particle size hidden perturbation is used to generate a zero-mean Gaussian sequence based on the particle size standard deviation of the current material configuration; the constant 1 in each parenthesis is the dimensionless reference factor. Any multiplicative factor less than 0.50 is taken as 0.50, and greater than 1.50 is taken as 1.50.

[0270]

[0271] In the formula, Indicates the sampling time index; Indicates time The digital response generated at the receiving end is expressed in digital response units. This indicates the generation of an unloaded response, in digital response units. This represents the effective attenuation coefficient, expressed in meters per kilogram. This indicates the load per unit length, expressed in kilograms per meter. This indicates the detector bias, expressed in digital response units (DRTs). It follows a normal distribution with a mean of 0 and a standard deviation of 12 DRTs and remains constant within the same installation section. This represents zero-mean Gaussian received noise with a standard deviation of 18 digital response units. Represents the natural exponential function; This indicates taking the larger value among the quantities within the parentheses; the constant 1 represents one digital response unit and constitutes the lower limit of the output.

[0272]

[0273] In the formula, Indicates the index of the valid sampling time; This indicates the sequence number of the last valid sampling time within the current cumulative interval; Indicates time The reference mass flow rate is expressed in kilograms per second. This indicates the conveying speed, measured in meters per second. This indicates the load per unit length, expressed in kilograms per meter. Indicates the first to the second Reference cumulative mass at each valid sampling time, in kilograms; Indicates time The actual time interval between the previous valid sampling time and the previous sampling time, in seconds; the basic generated data is fixed at 1 second; summation upper and lower limits. =1 and This indicates the sum of all valid sampling times within the specified interval.

[0274] 4. Missing and Abnormal Injection

[0275] The ash content missing probability of each installation segment is uniformly sampled between 25% and 55%; when missing, white noise with the nominal ash content of the material plus a standard deviation of 0.7 percentage points is used as filler, and the missing marker is set to 1. Communication anomalies are generated on independent copies of installation segments without anomalies, and the anomaly types do not overlap: packet loss ratios are 1%, 3%, and 5% respectively; spike ratios are 1%, 3%, and 5% respectively, with spike symbols being positive or negative with a probability of 0.5, and amplitudes ranging from 10% to 22% of the current response; duplicate frame ratio is 0.5%; saturation ratio is 0.5%, and the saturation value is 12000. Samples marked with anomalies only enter the anomaly protection stress test.

[0276] 5. Data fields, generation order, and reproduction checks

[0277] Each installation segment is generated in the following fixed order; changing the order will change the location of the pseudo-random number consumption and should be recorded as the new generator version.

[0278] (1) Reset the continuous variable subflow and material category subflow using the installation segment integer seed.

[0279] (2) Extract material type, equipment gain, installation factor, detector bias and foundation no-load response.

[0280] (3) Generate the no-load-load state sequence and the unit length load.

[0281] (4) Generate ambient temperature, relative humidity, tube voltage, tube current and source casing temperature.

[0282] (5) Generate conveying speed, ash content, hidden moisture and particle size disturbance.

[0283] (6) Calculate according to the no-load response formula Calculated according to the effective attenuation formula .

[0284] (7) Calculate according to the received response formula The reference mass flow rate and reference cumulative mass are calculated according to the velocity-load relationship.

[0285] (8) Perform gray missing sampling and generate fill values ​​and missing markers.

[0286] (9) Generate a base record without anomalies and write it into the training, validation or target evaluation identifier.

[0287] (10) Copy the base record and inject packet loss, spike, duplicate frame and saturation into the independent abnormal sub-stream respectively.

[0288] (11) Output field order, generator version, parameter configuration, random seed and record hash value.

[0289] The reproduction check was conducted using the following criteria: each installation segment must contain 3600 strictly incrementing timestamp records; the 16 installation segments totaled 57600 basic records; the number of records for S01, S02, S03, and S04 were 14400, 14400, 14400, and 14400, respectively; the reference quality flow and reference cumulative quality without anomaly records must not contain non-finite values; records marked with anomalies must not be included in the main model weight training; the same generator version, dependency version, parameter configuration, and random seed must produce the same SHA-256 record hash value. If any of these conditions are not met, the release of that batch of source domain data will be stopped.

[0290] (II) Training Objectives of the Main Model

[0291] In the model building phase, considering the high cost of collecting industrial field reference measurement data and the limited number of labels, a parametric simulator consisting of a fixed random seed, a scene-equipment-installation mapping, and numerical ranges for materials, loads, environment, source states, detector responses, and missing and anomaly injections is used to generate data for multiple scenes, multiple devices, and multiple installation segments. Physically constrained source domain data is used to fit the physical fundamental estimator and train the causal temporal residual model; after the S03 source domain validation data, which was not involved in fitting and training, passes verification, the main model weights are fixed as a read-only version and distributed to the field.

[0292] The pre-training objective of the temporal residual model in the physically constrained source domain is:

[0293]

[0294] In the formula, This represents the parameter that minimizes the training objective. This represents the parameters of the causal time-series residual estimator to be optimized. This represents the source domain parameters obtained after optimization. Indicates the training time index; Let S01 and S02 represent the set of training time points after excluding communication anomalies, saturation, spikes, duplicate frames, and previous value preservation samples. The summation only iterates through this set. This represents the smoothed L1 point loss, with its absolute error inflection threshold fixed at 1 kg / s. Indicates the physical basis mass flow rate of the source domain; The parameter is A causal time-series residual estimator; This represents the source domain causal window constructed according to the operational characteristics and standardized using the S01 and S02 statistics; Indicate the window length and set it to 60; and These represent the standard deviation and mean of the source domain residual labels, respectively. This represents the source domain reference quality flow. , , and The unit for all values ​​is kilograms per second.

[0295] After training, the source domain validation metric is calculated only in S03. The main model package is only solidified when the main model output has no non-finite values, the cumulative relative quality error in S03 is no greater than 1.0%, and the mean absolute error is no greater than that of the physical basis estimator; if any of these conditions are not met, the package is not released. S04 never participates in weight updates.

[0296] The temporal residual function in the pre-training objective represents a composite residual mapping that includes the standardized residual output and its destandardization step. This composite residual mapping outputs a residual compensation quantity with mass flow rate dimensions, which, when added to the physical basis mass flow rate, approximates the source domain reference mass flow rate.

[0297] When only 5 to 20 reference calibration segments covering low, medium, and high load ranges are obtained during the target field phase, all master model weights of the physical basis estimator and the causal temporal residual estimator are frozen. Only the input-normalized controlled statistics, the no-load response baseline update gating threshold, and the installation-level scaling parameters are updated. and bias parameters During the target field inference and adaptation phase, the main model weights are not unfrozen. If it is necessary to update the causal time-series residual estimator, the current target field adaptation process is terminated, and a new model building phase is initiated. The model is then retrained and validated using training and validation sets isolated from the field operation data to form a new read-only main model version. Simulation pre-training provides main model initialization, while target installation still performs explicit no-load data acquisition, reference measurement calibration, and independent validation segment verification.

[0298] The input normalized statistics are configuration parameters that are independent of the main network weights, have limited degrees of freedom, and are version-controlled. They are not part of the frozen physical basis estimator and causal temporal residual network weights. Their candidate updates also need to undergo integrity verification, independent acceptance, and versioned release.

[0299] Limited on-site samples necessitate a frozen high-degree-of-freedom main network when samples are scarce, updating only the installation-level scale and bias. Physically constrained source domain data covers a range of operating conditions and initializes the main model; industrial field measured data is incorporated into the target installation parameter fitting, fitting segment and validation segment isolation, and candidate version verification process. Explicit no-load data acquisition and traceable reference measurement are still performed at the target site.

[0300] Figure 2 Three types of isolated data streams are given. Physically constrained source domain samples are only used for training and source domain validation of the physical fundamental estimator and the causal temporal residual main network; industrial field reference fitting segments are only used for calculating installation-level proportions and bias parameters; field validation segments that did not participate in fitting are used to recalculate candidate version and effective version metrics respectively. Candidate version packets are first written to the candidate logical address area and read-back integrity checks are performed. After both integrity and validation metrics meet the thresholds, the processor switches the running pointer between adjacent complete inference cycles.

[0301] VI. Controlled source domain pre-training and cross-site few-sample adaptation

[0302] A controlled dataset was constructed, comprising 4 simulated sites, 8 devices, 16 installation sections, and 57,600 sampling points. Simulated variables included pipe voltage, pipe current, source end temperature, ambient temperature and humidity, coal type, ash content missing, hidden moisture / particle size, installation ratio / bias, no-load segments, and communication anomalies. Scene identifiers S01 to S03 served as physical constraint source domains for pre-training of the main model, while scene identifier S04 served only as a cross-site target domain and did not participate in the main model training.

[0303] Each installation is sampled for 1 second. During the initialization phase, the median of 60 explicit no-load valid responses is taken as... ;calculate and Physical basis estimates are formed using standardized Ridge; the residual objective is... TCN preferably employs causal convolution, kernel size 3, and expanded residual blocks to output a single residual. In the target domain, with the master model frozen, 20 reference calibration segments covering low, medium, and high predicted loads are selected. Only installation-level calibration parameters are fitted, and the cumulative quality deviation is used in the calibration.

[0304] like Figure 6 As shown, in a controlled simulator mismatch test consisting of 20 random seeds and 80 leave-one-out field partitions, the gradient boosting tree lightweight surrogate model PG-TSC-lite-HGB based on physical characteristics and temporal statistics was used. The source domain used a nominal simulator, and the target domain was sequentially incorporating four levels of attenuation coefficients (nominal, mild, moderate, and strong), detector bias, no-load response reference drift, humidity coupling, and noise mismatch. The no-load response reference, dimensionless logarithmic attenuation characteristics, and temporal statistical characteristics were recalculated based on observations in the target domain. Under strong mismatch, the median cumulative quality error of the Physics-Temporal-HGB without installation calibration was 4.863%, and the proportion of partitions below the 1% baseline was 8.750%. After fitting the installation-level parameters using 60 and 180 non-overlapping 1-second reference sampling points, the median cumulative quality errors were 0.350% and 0.198%, respectively, with corresponding proportions of 83.750% and 98.750%. Figure 6 The lightweight proxy experiment shown did not load the weights of the causal temporal convolutional network, nor did it use the cumulative quality loss of continuous segments. Its results only illustrate the controlled boundary of the low-degree-of-freedom installation calibration interface and the change in the number of reference sampling points, and are not used as proof of the accuracy of field measurement or the performance of the causal temporal residual network.

[0305] VII. Small Sample Ratio Anchoring of Target Domain Interfaces in Field-like Scenarios

[0306] The controlled-class on-site target domain data constructed using physical constraints comprises 36,000 sampling records, organized into 5 scene identifiers, 10 devices, and 30 installation sections. Scene identifiers S01 to S03 are used for training, and scene identifiers S04 to S05 serve as the target domain. For each target installation section, the first 30% is the candidate interval, and the remaining 70% is the evaluation interval. The system first runs the current physical-temporal model within the candidate intervals, then selects 5 unique 1-second reference sampling points for each installation section based on predicted flow rate, with fixed bias parameters. If the value is 0, only the reference mass flow rate is used to anchor the proportional parameter. The evaluation interval is not included in the fitting. This result is used to verify the interface of small sample parameters, data isolation and sampling process, and does not represent the conclusions of actual industrial field equipment acceptance, long-term continuous operation or metrological traceability.

[0307] Each reference calibration segment records the segment start time, segment end time, reference cumulative quality, data quality markers, and reference measurement source; the actual time interval between adjacent valid sampling times is used to predict the cumulative quality integral.

[0308] The aforementioned field target domain interface organizes data by installation segment and isolates candidate segments from evaluation segments. Candidate segments calculate proportional parameters using a small number of reference measurement segments and cumulative mass anchoring; evaluation segments do not participate in fitting, only calculating the cumulative mass error and load coverage between candidate and effective versions. When the target device comes online, explicit no-load data acquisition is still performed, and traceable field reference measurement segments are used to complete installation parameter fitting and verification segment validation.

[0309] VIII. No-load gate control and abnormal protection

[0310] The acquisition, synchronization, and serial port protection module first checks the communication status, frame number, timestamp, saturation, and spike markers. Then, the no-load response benchmark gating and update module calculates the write enable flag. In the simulated stress test, the balanced gating achieved a precision of 99.690%, a recall of 64.809%, a loaded sample false write rate of 0.023%, and a cumulative quality error of 0.021%. The loose gating achieved a loaded sample false write rate of 14.104% and a cumulative quality error of 4.742%. When gating conditions are incomplete, communication is abnormal, or the no-load status is not confirmed, the write enable flag is set to 0, and the previous no-load response benchmark remains unchanged. The field threshold is determined by the explicit no-load samples of the target device, the short-term repeatability of the detector, and the loaded exclusion samples.

[0311] In the serial port abnormal stress test, under the 3% spike condition, the replacement of the moving median with the absolute median spike caused the cumulative quality error to change from 0.395% to 0.018%; under the 5% spike condition, it changed from 1.135% to 0.186%. Under the 3% packet loss condition, the cumulative quality errors without protection and with previous value hold were 0.216% and 0.264%, respectively. When packet loss occurs, previous value hold is performed for a maximum of 5 consecutive output cycles, retaining the communication invalidation flag and the previous value hold flag; after 5 output cycles, the increase in cumulative quality stops, the most recent valid cumulative quality is retained, and a communication interruption status code is output. Previous value hold does not recover missing observations, and samples with the hold flag do not participate in the no-load response benchmark update, candidate fitting, or validation fragment recalculation.

[0312] IX. Risk Coverage Sampling and Operational Indicator Switching

[0313] During the label-free operation phase, the system generates risk events based on the installation calibration correction magnitude, load range extrapolation, characteristic extrapolation, anomaly rate, and no-load response baseline update status. In the controlled simulation of the target domain, the cumulative quality error of direct migration of the master model is 1.822%, which is 0.854% after fitting the installation-level parameters using 60 reference segments; the cumulative quality error is 1.945% when using 20 consecutive initial segments, which is 0.623% after selecting 20 segments according to the predicted flow rate stratification. Risk events enter the review or sampling queue, and the installation calibration parameters and residual model remain at their original values ​​until the reference metrology interface provides a valid quality label.

[0314] After obtaining the reference metrology segment, the segment's load coverage, time synchronization, communication quality, reference metrology traceability, and material condition are checked first. Then, with the master model frozen, candidate installation calibration parameters are calculated from the fitted segment, and the integral deviation is limited by cumulative quality consistency constraints. The candidate version package, consisting of the candidate installation calibration parameters and associated configurations, is written to the isolated candidate version storage area. After reading it back, the length, hash value, or cyclic redundancy check code is verified. Then, the cumulative quality error, load coverage, parameter boundaries, data quality, and security status are recalculated using the verification segment. Once the integrity and verification indicators both meet the thresholds, the processor atomically switches the running pointer between adjacent complete inference cycles. The above values ​​are limited to the controlled simulation configuration of the target domain.

[0315] 10. Layout and relocation of ores and other bulk materials

[0316] For materials such as ores that differ from coal, the equipment-side, edge-side, gating, timing residual, installation calibration, version verification, and operation pointer switching structures remain unchanged, but the following configurations are re-executed: Select the X-ray energy spectrum and integration time that ensures no-load unsaturation and maximum material layer resolution; acquire explicit no-load and multiple load reference segments for this material configuration; re-estimate the effective attenuation characteristics and physical model parameters; use dual-energy response and material category coding if necessary; regenerate gating thresholds, installation calibration parameter boundaries, and verification thresholds. If changes in material composition cause the monotonic relationship between single-energy attenuation and areal density to no longer remain usable, the system should switch to a dual-energy or multi-energy scheme, or determine that the material configuration is unsuitable; the coal model should stop extrapolating.

[0317] This configuration migration method is applicable to other bulk materials that meet the calibrable transmission relationship; during deployment, the no-load response benchmark, material-related parameters and calibration parameters should be redefined, and the specific measurement accuracy should be based on the on-site calibration results of the target material.

[0318] The implementation process is described in detail below using a real device as an example:

[0319] 1. New installation section online

[0320] Write device identifier, installation identifier, participating channel mask, full scale of each channel, tube voltage range 75-85 kV, tube current range 1.35-1.65 mA, speed range 0-3.5 m / s, and one-second output cycle; confirm that there is no material on the conveyor belt, obtain 60 positive protection responses that are valid, unsaturated, and without spikes, and initialize the no-load response benchmark with the median; load the read-only master model package, verify the data length and 32-bit cyclic redundancy check code, and begin to form the physical basis mass flow rate and causal timing residuals; collect 5, 10, ... 15 or 20 fitted segments; segments covering low, medium, and high loads, and each segment containing at least two strictly incremental valid sampling times; freeze the master model and search according to the determined grid search ratio and bias parameters; form candidate version packages with a verification flag of 0 and perform the first write and readback; recalculate the five acceptance criteria on the verification segments that did not participate in the fitting; after all criteria are met, set the verification flag to 1, recalculate the checksum, and perform a second overall write and readback; after the second readback passes, switch the running pointer between adjacent complete inference cycles; if it fails, keep the original version.

[0321] 2. In this embodiment, an X-ray transmission measurement device is arranged on a coal conveyor belt with a bandwidth of 1200 mm. The equipment, materials, reference measurement and synchronization configuration are shown in the table below.

[0322]

[0323] 3. Determine the test procedure

[0324] (1) Explicit no-load test for 60 seconds and repeated 3 times; record the mean, median, standard deviation and initial baseline of the no-load response each time.

[0325] (2) Obtain no less than 3 reference segments for each of the low, medium and high loads, with each segment lasting no less than 300 seconds; the three-level boundaries are determined based on the empirical quantiles of 33.3% and 66.7% of the average reference mass flow rate of all candidate loaded segments.

[0326] (3) The material coverage targets are ash content 2%-33%, moisture content 5%-15%, particle size 0-80 mm and unit length load 0-190 kg / m; the uncovered areas are marked as unverified in the results table.

[0327] (4) Continuous operation for no less than 72 hours, during which the cumulative quality register is not reset; shutdown or communication interruption segments are counted separately according to the quality mark.

[0328] (5) For the packet loss test, 1% and 3% of independent random packets were injected respectively; for the spike test, 1%, 3% and 5% of independent random spikes were injected respectively, and the spike amplitude was taken as 10%-22% of the absolute value of the response at that time.

[0329] (6) The temperature drift test is maintained at 18, 25 and 35 degrees Celsius for no less than 30 minutes respectively; when the ambient temperature cannot be controlled, use a real running segment that includes the range of plus or minus 1 degree Celsius around these three temperatures.

[0330] (7) The source aging playback is multiplied by a coefficient that linearly decreases from 1.00 to 0.96 on the original received response. The playback sequence must not be used for main model training or parameter fitting.

[0331] (8) For each test, the same time segment is used to calculate the results before and after calibration, and the records of the first write, independent acceptance, second write, pointer switching or rejection of switching of the candidate version are saved.

[0332] 4. Equipment performance indicators

[0333]

[0334] In the formula, It represents the mean absolute error, expressed in kilograms per second. This represents the root mean square error, expressed in kilograms per second. This indicates the number of sampling points that participated in the evaluation and whose communication was valid and whose reference labels were valid. Indicates from 1 to Evaluation sampling point index; Indicates time The predicted mass flow rate is expressed in kilograms per second. This represents the reference mass flow rate after time synchronization at the same moment, in kilograms per second; the vertical line indicates the absolute value; the upper and lower limits of the two summation symbols. =1 and Indicates the sum of all. Each evaluation sampling point is used. No load, low load, medium load, and high load are calculated separately, without using the full average value to replace the stratified results.

[0335]

[0336] In the formula, This represents the cumulative relative mass error and is a dimensionless quantity. This indicates the predicted cumulative mass of the evaluation segment, in kilograms. This indicates the cumulative reference mass, expressed in kilograms. This indicates that the larger value among the quantities in parentheses is taken; 1 kg is the lower limit of the denominator. The values ​​before and after calibration are reported for each load segment, the entire 72-hour period, the temperature drift segment, the source aging playback segment, and each abnormal injection segment.

[0337] 5. Equipment test results

[0338] The median receiver responses for the three 60-second explicit no-load tests were 9235, 9228, and 9241, with standard deviations of 18.4, 19.1, and 17.6 digital response units, respectively. The maximum rate of change of the no-load response baseline relative to the start of each maintenance window over 72 hours was 2.9%, and the false write rate for loaded samples was 0.07%. Detailed test results are shown in the table below.

[0339] Table 1:

[0340]

[0341] Table 2:

[0342]

[0343] During the 72-hour operation, two candidate installation parameter versions were generated. One of them was switched after independent verification and secondary integrity readback, while the other remained the original effective version due to the cumulative quality relative error of the high-load verification segment being 1.17%. The number of running pointer switches was 1, and there was no state of no valid historical version.

[0344] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A self-calibration method for X-ray weighing based on physical constraints and time-series compensation, applied to the weighing of bulk materials, wherein the bulk materials are materials with a calibrable transmission attenuation relationship with the material mass surface density under at least one X-ray energy response channel, characterized in that... This includes the model building phase and the on-site reasoning and adaptation phase; The model construction phase includes: generating physically constrained source domain data based on the Beer-Lambert attenuation relationship and the ranges of the X-ray source, receiver detector, installation dimensions, and material parameters; calculating the receiver digital response at each source domain sampling time according to the relationship that the receiver digital response equals the product of the no-load response reference and the natural exponential attenuation term, where the exponent of the natural exponential attenuation term is the negative of the product of the effective attenuation coefficient and the unit length load, and calculating the source domain reference mass flow rate according to the product of the unit length load and the delivery speed; and changing the source output, detector gain and bias, effective attenuation coefficient, material state, installation ratio and bias, environmental disturbances, missing auxiliary variables, and communication. At least one of the anomalies is used to form source domain samples across equipment, installation sections, and operating conditions; using dimensionless logarithmic decay characteristics, the product of the dimensionless logarithmic decay characteristics and the conveying speed, and equipment operating status as physical inputs, and using the source domain reference mass flow rate as the fitting target, the physical basis estimator parameters are obtained and the source domain physical basis mass flow rate is obtained; using the difference between the source domain reference mass flow rate and the source domain physical basis mass flow rate as the residual label, a causal temporal residual estimator is trained; using source domain validation fragments that are not involved in fitting and training, the main model composed of the physical basis estimator and the causal temporal residual estimator is validated, and the weights of the validated main model are set to read-only and the model building stage is ended; After entering the field inference and adaptation phase, the edge processing device loads and freezes the read-only master model, updates the installation calibration parameters only based on the field reference metrology segment, and performs the following steps: S1. Obtain the digital response of the receiver after X-rays pass through the conveying mechanism and the bulk material it conveys, the conveying speed, and the operating status of at least one device, synchronize them according to the timestamp, and generate data quality markers based on the communication status, range status, and time continuity. S2. Identify packet loss, duplication, saturation, spikes, or time anomalies in the digital response of the receiving end to obtain an effective response after anomaly protection; generate an idle gating value based on communication effectiveness, the closeness of the effective response to the idle response reference, attenuation, short window stability, and transmission speed stability; update the idle response reference only when the idle gating value meets the update conditions, freeze the idle response reference when the update conditions are not met, and prohibit abnormal data from participating in the update of the idle response reference and the update of installation calibration parameters; S3. Calculate the dimensionless logarithmic decay characteristic based on the no-load response benchmark and the current effective response. Input the dimensionless logarithmic decay characteristic, the product of the dimensionless logarithmic decay characteristic and the conveying speed, and the equipment operating status into the physical basis estimator to obtain the physical basis mass flow rate. S4. The multi-source features at the current time and before the current time are combined into a causal time window, which is then input into the causal time series residual estimator to obtain the residual compensation amount relative to the physical foundation mass flow rate. The physical foundation mass flow rate is added to the residual compensation amount to obtain the main model mass flow rate. S5. Using at least one of the dimensionless proportional parameter corresponding to the target equipment or target installation section and the bias parameter with the dimension of mass flow rate, construct the installation calibration parameters, and perform affine calibration and non-negative truncation on the mass flow rate of the master model according to the following formula: Wherein, the subscript d represents the target device or target installation section. This represents the mass flow rate of the master model input to the calibration layer at the t-th valid sampling time. It is a dimensionless proportionality parameter. A bias parameter with the same dimensions as mass flow rate. The final instantaneous mass flow rate after affine calibration and non-negative truncation is represented; the proportional parameter not involved in the field estimation is set to 1, and the bias parameter not involved in the field estimation is set to 0; the final instantaneous mass flow rate is integrated according to the actual time interval between adjacent valid moments to obtain the cumulative mass.

2. The method according to claim 1, characterized in that, The reference fitting segment and the verification segment are obtained through the reference metrology interface; with the master model consisting of the physical basis estimator and the causal time series residual estimator frozen, the candidate installation calibration parameters are fitted with the cumulative quality consistency loss as a necessary constraint. Candidate installation calibration parameters are generated only when both the reference cumulative mass total and the candidate predicted cumulative mass total meet the preset minimum mass threshold, and the dimensionless proportional parameters are limited to a preset range. When only cumulative quality labels are available, the bias parameter is fixed at 0. When instantaneous reference quality flow is available, the absolute value of the bias parameter does not exceed a preset proportion of the median of the loaded master model's quality flow. Candidate installation calibration parameters and their associated configurations are grouped into a candidate version package, and integrity verification is performed on the candidate version package; the candidate version is independently accepted using the verification fragments that were not involved in the fitting; if the acceptance is successful, the running pointer is switched at the boundary of the complete inference cycle to make the candidate version effective; if the acceptance or integrity verification fails, the original effective version is retained. It also includes risk management steps: compiling risk indicators according to a preset statistical period; A risk event is generated when the proportion of abnormal frames exceeds a preset threshold, the installation calibration correction magnitude exceeds a preset empirical quantile, the main model mass flow rate or dimensionless logarithmic decay feature exceeds the distribution range of the reference fitting segment or source domain training set, the relative change rate of the idle response benchmark exceeds a preset threshold, or the effective version integrity verification information is inconsistent with the non-volatile version storage area record. Risk events trigger the review of reference segments or the selection of new calibration segments, and freeze the write enable of installation calibration parameters; communication anomalies cause the previous value to remain for more than a preset number of cycles, suspend the cumulative quality increment and output an abnormal status code; risk events themselves do not directly modify the installation calibration parameters.

3. The method according to claim 1, characterized in that, S2 includes: providing positive value protection for the current effective response after anomaly protection. In the formula, This indicates the digital response of the aggregated receiver after communication protection. This represents the response after positive value protection; the constant 1 is the minimum calculated value of the digital code at the receiving end; Calculate the non-negative conservative attenuation and normalized relative deviation based on the effective response after positive protection and the previous no-load response benchmark: In the formula, Indicates the index of the current output cycle; Indicates time The non-negative conservative decay; This indicates the relative deviation of the current protection response from the previous no-load response baseline; This represents the baseline of the no-load response at the previous moment; This indicates that the current state is in the protection response phase; The logarithmic molecule protection constant; Define the basic online update gate as: In the formula, This indicates that the basic online update gate value is being implemented. Indicates the output cycle index; Indicates the number of samples initiated; Indicates a non-negative conservative decay; Indicates the attenuation threshold; Indicates relative deviation; Indicates the relative deviation threshold; When gating is successful, update the no-load response baseline according to the following formula; when gating is unsuccessful, maintain the no-load response baseline from the previous moment: In the formula, and These represent the current and previous time-independent no-load response baselines, respectively. This indicates that the current state is in the protection response phase; This represents the baseline update coefficient; when the gate value is 0, the current baseline retains the previous value; when the gate value is 1, it is updated exponentially.

4. The method according to claim 3, characterized in that, When the output cycle is 1 second and the conveying mechanism is a belt conveyor, the no-load response benchmark is initialized using the median of the explicit no-load valid samples, and the initial sample number, gating threshold, and update coefficient are: In the formula, Indicates the initial no-load response baseline; to This represents the continuously acquired positive protection response during the explicit no-load phase; when the output period is 1 second, the number of initial samples is... The value is 60, corresponding to a 60-second explicit no-load stabilization time; the output cycle is... seconds, Divide 60 by The rounded-up value of the obtained numerical value, where This indicates the time interval between adjacent output moments, in seconds.

5. The method according to claim 3, characterized in that, Under the premise that the basic online update gating is established, the product of the communication validity condition, response proximity condition, low attenuation condition, short window normalized fluctuation condition, transmission speed stability condition, and external no-load confirmation condition is set as the multi-source extended gating factor: In the formula, Indicates the multi-source extended gating factor; , , , , and These represent, in order, communication, near-response, low attenuation, short-window fluctuation, stable transmission speed, and external confirmation conditions; The product of the basic online update gate value and the multi-source extended gate factor is used as the final gate value: When no external no-load confirmation signal is configured, the external no-load confirmation condition is set to 1; the update state is entered only after the final gate value is 1 for 5 consecutive output cycles, and the update state is exited when any gate condition is not met for 2 consecutive output cycles, thus forming an entry and exit hysteresis; the previous value caused by communication abnormality is allowed to be retained for a maximum of 5 consecutive output cycles, and the final gate value is forcibly set to 0 when it exceeds 5 output cycles.

6. The method according to claim 1, characterized in that, The S3 constructs a dimensionless logarithmic attenuation characteristic based on the following X-ray attenuation relationship: In the formula, Indicates time The digital response from the receiving end; This indicates the no-load response reference used at the same time. Represents the natural exponential function; Indicates load per unit length; Indicates the sampling time index; Indicates time Dimensionless logarithmic decay characteristic; Indicates the effective attenuation coefficient; The physical input consists of the dimensionless logarithmic decay characteristic, the product of the dimensionless logarithmic decay characteristic and the conveying speed, and the equipment operating status. A regularized physical estimator is then used to obtain the physical baseline mass flow rate. In the formula, Indicates time The physical input vector; the constant 1 is the intercept term; and The product represents the attenuation-velocity coupling amount; Indicates the conveying speed; Indicates the tube voltage; Indicates tube current; Indicates the source casing temperature; Represents environmental observations; Indicates the granularity of the proxy quantity; Indicates missing auxiliary quantity markers. This represents the parameter that minimizes the objective function; Represents the optimization variable. This represents the parameters of the physical basis estimator obtained through the objective function; Represents the reference mass flow vector of the source domain training samples; Represents the source domain physical input matrix. This represents the matrix after standardization using the source domain training statistics; This represents the square of the Euclidean second norm; Indicates time The physical input vector; Indicates time The physical basis of mass flow rate.

7. The method according to claim 6, characterized in that, The auxiliary variable is ash content; when ash content cannot be continuously obtained online, it is filled using the ash content statistics of historically confirmed data from the training domain or the current material configuration, and the filled value and the corresponding missing value are jointly input into the physical basis estimator and the causal time-series residual estimator. The relationship between the missing value and the filling is as follows: In the formula, Indicates time Gray missing marker; This represents the ash content value fed into the model; Indicates time The effective ash content measurement value; This represents the median gray value calculated solely from historical values ​​confirmed in the source domain training set or the current material configuration.

8. The method according to claim 1, characterized in that, The causal temporal residual estimator is a temporal convolutional network. Each residual block includes two layers of causal one-dimensional convolution, nonlinear activation, Dropout, and residual connections. The dilation rate of each residual block increases. When the number of channels is mismatched, a 1×1 convolutional projection is used. The operation relationship of the residual blocks is as follows: In the formula, and They represent the first Input and output characteristics of each residual block; Take 0 to -1, number of residual blocks Take 4; Indicates a causal convolution branch; This indicates a shortcut branch for dimension matching; This represents the linear rectification activation function.

9. The method according to claim 1, characterized in that, The installation calibration parameters include at least the proportional parameters and bias parameters corresponding to the target equipment or target installation section, and the final instantaneous mass flow rate is obtained according to the following formula: In the formula, Indicates the first The final instantaneous mass flow rate estimate for each valid sampling time; This indicates that the mass flow rate estimate output by the main model is frozen at the same time. Indicates installation The corresponding dimensionless proportionality parameter; Indicates installation The corresponding bias parameters have the same dimensions as the mass flow rate; This indicates taking the maximum value among the quantities within the parentheses, with non-negative truncation using 0; subscript Indicates the valid sampling time, subscript Indicates the target device or target installation section. Using the predicted mass flow rate of the master model within each reference measurement segment as input to the candidate installation calibration parameters, the candidate calibration output and the candidate predicted cumulative mass are calculated according to the following formulas: In the formula, Indicates the time within the reference segment The candidate instantaneous mass flow rate estimates; This indicates that the mass flow rate estimate output by the main model is frozen at the same time. and These represent the candidate installation ratio parameter and the offset parameter, respectively; Indicates the candidate version in the reference fragment The predicted cumulative quality; Represents the set of valid sampling times; Represents a set The summation time index within; The actual time interval between adjacent valid sampling moments is represented by the following cumulative quality consistency loss as a necessary term for fitting the candidate installation calibration parameters: In the formula, Indicates the cumulative reference mass of the same reference segment; This indicates the cumulative relative deviation in quality of the segments; To stabilize the denominator.

10. An X-ray weighing self-calibration system based on physical constraints and time-series compensation, characterized in that, include: A conveying mechanism that carries and continuously conveys bulk materials, an X-ray source and a receiving detector arranged on opposite sides of the conveying mechanism, a conveying speed sensor that measures the conveying speed of the conveying mechanism, a reference measurement interface connected to a reference measurement device, and an edge processing device that is connected to the receiving detector, the conveying speed sensor and the reference measurement interface respectively. The edge processing device includes a processor, a memory, and a communication interface. The memory includes a volatile working memory area and a non-volatile version memory area. The volatile working memory area is configured with an input circular buffer, and the non-volatile version memory area is configured with non-overlapping logical address areas for the effective version, candidate versions, historical versions, and a running pointer record area. The processor executes program instructions in the memory to form: The acquisition synchronization and anomaly protection unit receives the timestamped digital response, transmission speed and equipment auxiliary status of the receiving end through the communication interface and writes them into the input circular buffer. It marks packet loss, duplicate frames, spikes, saturation and time anomalies, and prevents frames with anomaly marks from being written into the idle response benchmark and candidate parameter fitting data. The gating and no-load response reference update unit generates a write enable flag based on the communication status, the closeness of the received response to the no-load response reference at the previous moment, the non-negative conservative attenuation, the normalized relative deviation, and the stability of the transmission speed. It only modifies the stored value of the no-load response reference when the write enable flag is valid. The physical basis estimation unit constructs a dimensionless logarithmic decay characteristic and its product with the conveying speed based on the no-load response benchmark, the current effective response and the conveying speed, and calculates the physical basis mass flow rate. The causal temporal residual unit loads read-only master model weights trained with physically constrained source domain data and validated by independent source domain verification fragments. It reads multi-source feature windows from the input circular buffer for the current time and previous times and calculates residual compensation quantities with mass flow rate dimensions. The training label of the residual compensation quantity is the difference between the source domain reference mass flow rate and the source domain physical basis mass flow rate. The read-only master model weights are not modified during the on-site inference and adaptation stages. The fusion and installation calibration unit adds the physical basis mass flow rate and the residual compensation amount, multiplies the resulting master model mass flow rate by the proportional parameter of the target equipment or target installation section and adds the bias parameter, compares it with 0 and takes the larger value as the final instantaneous mass flow rate, and integrates it according to the actual time interval between adjacent effective moments to form the cumulative mass. The candidate parameter calculation and verification unit calculates candidate proportional parameters and candidate bias parameters using field reference fitting segments while the main model parameters are frozen. The candidate version package is composed of the main model version identifier, the no-load response baseline update strategy version, the candidate installation calibration parameters, the no-load response baseline snapshot when the candidate is formed, the standardized statistics, the data quality rules, the version number, the verification pass flag, the data length, and the 32-bit cyclic redundancy check code; the verification pass flag is initially set to 0, and the check code covers all candidate version package fields except for the check code field itself. After the candidate version package is written into the candidate version logical address area for the first time, it is read back and the data length, check code and verification pass flag 0 are verified. Then, the cumulative quality error, load coverage and parameter boundary of the candidate version and the effective version are recalculated using the verification fragments that did not participate in the fitting. The version switching and rollback unit sets the verification pass flag to 1 when the verification segment meets the following conditions: the cumulative relative quality error does not exceed 1.0% and is not greater than the effective version; the non-negative truncation ratio of the 60-second loaded window does not exceed 1%; the change rate of the current effective unloaded response benchmark relative to the snapshot when the candidate was formed does not exceed 4% during acceptance; the verification segment has no abnormal markers; and each of the low, medium, and high load layers contains at least one verification segment. The 32-bit cyclic redundancy check code is recalculated based on the complete candidate version package after the flag is set to 1, and then written and read back as a whole. The running pointer is switched only after a complete inference cycle ends and before the version snapshot is read in the next inference cycle, provided that the data length, check code, and verification pass flag 1 are consistent in the second readback. If any verification fails, the original effective version is maintained. If the data length or check code of the effective version is mismatched, the verification pass flag is invalid, or the output contains a non-finite value, historical versions are scanned from largest to smallest version number, and only the most recent historical version with consistent data length and check code readback and a verification pass flag of 1 is switched to. If no valid historical version exists, the metering output is frozen and an invalid status code is recorded. At the beginning of each inference cycle, the processor reads the running pointer and latches the corresponding version identifier. During this cycle, the idle gating, physical basis estimation, timing residual calculation, installation calibration, and cumulative mass increment all use the same version snapshot. During the candidate version writing, the communication interface continuously receives sensor frames, and the candidate version writing does not overwrite the effective version logical address area or clear the input ring buffer.

Images