A method and system for online thickness measurement of alloy hot-rolled plates

Abstract

This invention relates to the field of thickness measurement technology, and discloses an online thickness measurement method and system for hot-rolled alloy plates. The method includes: acquiring measurement data of the hot-rolled alloy plate to be measured; preprocessing the measurement data to obtain standardized data; performing calibration calculations on the standardized data using a calibration algorithm based on a dynamic benchmark and two-segment mapping to determine the basic thickness data of the hot-rolled alloy plate to be measured; extracting features from the standardized data using deep learning to generate global fusion features; performing error compensation calculations based on the global fusion features and standardized data to determine the error compensation factor; and correcting the basic thickness data based on the error compensation factor to determine the accurate thickness measurement data of the hot-rolled alloy plate to be measured. This invention integrates a physical mechanism model with a data-driven model, overcoming the limitations of a single model and improving the accuracy, stability, and adaptability of thickness measurement in complex and dynamic hot-rolling environments.

Description

Technical Field

[0001] This invention relates to the field of thickness measurement technology, and in particular to an online thickness measurement method and system for hot-rolled alloy plates. Background Technology

[0002] In the production process of alloy hot-rolled plates, online thickness accuracy is one of the core indicators determining product quality and yield. Due to the harsh environment of the hot rolling site, including high temperature, moisture, vibration, and the variable properties of alloy materials, the interference to the measuring sensors is strong and unstable. This makes achieving high-precision and high-stability online thickness measurement a technical challenge in this field.

[0003] In existing technologies, the most commonly used online thickness measurement techniques in industry mainly include radiographic thickness gauges and laser thickness gauges. Radiographic thickness gauges, which calculate thickness by measuring the attenuation of rays after passing through the strip, are currently the most widely used technology. Laser thickness gauges, on the other hand, utilize the principle of laser triangulation for non-contact measurement. These technologies form the basis of current thickness measurement methods.

[0004] However, existing technologies heavily rely on initial calibration, and the calibration models are static and fixed. In actual production, rolling conditions (such as temperature and alloy composition fluctuations) are dynamic, and static calibration models cannot adaptively adjust, leading to a "drift" phenomenon and a decrease in measurement accuracy over time. Secondly, existing technologies mostly involve isolated single-point measurements or simple linear compensation, making it difficult to extract deep features reflecting the true thickness and overall error from massive amounts of noisy measurement data. This results in unstable measurement results and insufficient accuracy under complex working conditions. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention provides an online thickness measurement method for hot-rolled alloy plates, comprising:

[0006] S1: Obtain measurement data of the hot-rolled alloy plate to be tested;

[0007] S2: Preprocess the measurement data to obtain standardized data;

[0008] S3: Through the calibration algorithm of dynamic benchmark and two-segment mapping, the standardized data are calibrated and the basic thickness data of the hot-rolled alloy plate to be tested are determined.

[0009] S4: Through deep learning, feature extraction is performed on standardized data to generate globally fused features;

[0010] S5: Based on the global fusion characteristics and standardized data, perform error compensation calculations and determine the error compensation factor;

[0011] S6: Based on the error compensation factor, the basic thickness data is corrected to determine the accurate thickness measurement data of the hot-rolled alloy plate to be tested.

[0012] A second aspect of the present invention provides an online thickness measurement system for hot-rolled alloy plates, comprising: a processor and a memory;

[0013] The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the online thickness measurement method for hot-rolled alloy plates as described in the first aspect.

[0014] The beneficial effects of this invention are reflected in the fact that by adopting a calibration algorithm based on dynamic benchmarks and mapping, it effectively overcomes the measurement drift problem caused by dynamic changes in rolling conditions in existing static calibration models, and significantly improves the long-term accuracy and stability of online thickness measurement. At the same time, by using deep learning technology to perform deep feature extraction and global fusion on massive and noisy measurement data, it breaks through the limitations of traditional single-point measurement and simple linear compensation, and can accurately mine and compensate for comprehensive errors under complex working conditions, thereby improving the stability and accuracy of thickness measurement results in real industrial environments. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of an online thickness measurement method for hot-rolled alloy plates provided by the present invention.

[0016] Figure 2 This is a flowchart illustrating a calibration method based on a dynamic benchmark and two-segment mapping provided by the present invention.

[0017] Figure 3 This is a flowchart illustrating a method for establishing a fixed mapping relationship between current and metered values ​​provided by the present invention.

[0018] Figure 4 This is a flowchart illustrating a standardized current timing data conversion method provided by the present invention.

[0019] Figure 5 This is a flowchart illustrating a feature extraction method provided by the present invention.

[0020] Figure 6 This is a flowchart illustrating an error compensation method provided by the present invention.

[0021] Figure 7 This is a flowchart illustrating a correction calculation method provided by the present invention.

[0022] Figure 8 This is a schematic diagram of the structure of an online thickness measurement system for hot-rolled alloy plates provided by the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1:

[0025] Reference Figures 1-7 The present invention provides an online thickness measurement method for hot-rolled alloy plates, comprising:

[0026] S1: Obtain measurement data of the hot-rolled alloy plate to be tested.

[0027] Optionally, the measurement data includes γ The current timing data, geometric parameters, alloy content data, and temperature timing data of the X-ray passing through the hot-rolled alloy plate under test.

[0028] Specifically, outliers are removed from the collected measurement data, and the parameters and temperature data of the hot-rolled alloy plate under test are labeled and correlated with the X-ray intensity signal, while standardization processing is performed.

[0029] S2: Preprocess the measurement data to obtain standardized data.

[0030] Preprocessing refers to filtering, denoising, and normalizing the raw measurement data to eliminate obvious outliers and dimensional effects, providing clean, uniform, and standardized data for subsequent calculations.

[0031] In this embodiment of the invention, data preprocessing effectively eliminates abnormal interference and dimensional differences in the original measurement data, providing a high-quality and standardized data foundation for subsequent calibration calculations and feature extraction.

[0032] S3: Through the calibration algorithm of dynamic benchmark and two-segment mapping, the standardized data is calibrated and the basic thickness data of the hot-rolled alloy plate to be tested is determined.

[0033] Among them, dynamic benchmark refers to the core mechanism that overcomes the limitations of traditional static calibration. It monitors the working parameters of the rolling process in real time (such as slab temperature, rolling speed, and alloy composition identification) and uses an adaptive algorithm to dynamically update the reference benchmark point or benchmark curve for thickness calculation. This enables the measurement system to automatically compensate for systematic errors caused by equipment thermal expansion, roll gap drift, or material changes, thereby ensuring the consistency between the calibration benchmark and the current production status.

[0034] The two-stage mapping refers to the computational architecture for achieving high-precision thickness inversion. This model decomposes thickness calculation into two consecutive mathematical transformation stages: the first stage mapping converts the preprocessed standardized data into a base thickness value under standard operating conditions based on a dynamic benchmark. The second stage mapping further introduces compensation coefficients for specific nonlinear factors (such as temperature gradient and stress distribution) to refine the base thickness value, ultimately outputting a thickness value that accurately reflects the true geometric dimensions of the plate.

[0035] In this embodiment of the invention, a calibration algorithm based on a dynamic benchmark and two-segment mapping is used to achieve real-time adaptation to changes in working conditions, effectively overcoming the measurement drift problem of traditional static calibration models and significantly improving the long-term stability and reliability of the basic thickness data.

[0036] In one possible implementation, S3 specifically includes sub-steps S301 to S304:

[0037] S301: Collect current data from various calibration plates of known thickness.

[0038] The calibration plate refers to a physical standard that is known, precise, and has a specific pattern. Its core function is to act as a "reference ruler," providing a dimensional reference for the measurement system, thereby accurately converting the raw signals collected by the sensor (such as pixel grayscale values ​​and voltage values) into dimensional units with actual physical meaning (such as millimeters).

[0039] S302: Based on the current data, establish a fixed mapping relationship between the current and the metered value.

[0040] Among them, the fixed mapping relationship refers to the establishment and maintenance of a mathematical transformation relationship between the system's input and output that, once set, remains unchanged and does not change with external conditions.

[0041] In one possible implementation, S302 specifically includes sub-steps S3021 and S3022:

[0042] S3021: Calculate based on the physical characteristics of the radioactive source. γ The theoretical measured value of the remaining energy after the rays pass through calibration plates of different thicknesses.

[0043] Among them, the physical characteristics of a radioactive source refer to the inherent physical parameters of a radioactive material during its decay process, such as its half-life, decay type, and the energy and intensity of the emitted rays. These characteristics determine the stability and applicability of a radioactive source as a measurement medium.

[0044] in, γX-rays are high-frequency electromagnetic waves generated by the transition of atomic nuclei energy levels. They have extremely strong penetrating power. In thickness gauges, the degree of attenuation of their intensity after passing through the material is a function of the thickness and density of the object being measured.

[0045] Among them, the measured value specifically refers to the value obtained through the measurement system. γ The final quantitative result obtained after collecting, converting and calculating physical signals such as X-ray intensity attenuation is used to characterize the physical properties (such as thickness) of the object under test.

[0046] S3022: Establish a fixed mapping relationship based on current data and theoretical measurement values.

[0047] In this embodiment of the invention, by combining the theoretical measurement values ​​established by the physical characteristics of the radioactive source with the measured current data, a fixed mapping relationship unaffected by fluctuations in operating conditions is constructed, providing a stable and reliable physical benchmark for thickness measurement and effectively ensuring the long-term accuracy consistency of the measurement system.

[0048] S303: Dynamically acquire real-time baseline values.

[0049] In this context, the reference value refers to a fundamental parameter in a measurement system that serves as a constant or known standard for comparison, calculation, or calibration with actual measured values.

[0050] Specifically, before each rolling operation begins, current data is collected as a segment of hot-rolled steel plate without the alloy being tested passes through the detector. The average value of this current data segment is used as the real-time reference value for the current rolling operation. The real-time reference value is used to eliminate reference drift caused by radioactive source decay, on-site temperature fluctuations, and mechanical structure offsets, replacing the traditional fixed reference value.

[0051] S304: Based on the real-time reference value and the fixed mapping relationship, the standardized current time series data in the standardized data is converted to obtain the basic thickness data.

[0052] Among them, current time series data refers to the sequence of current measurement values ​​that are continuously collected and recorded in chronological order. It reflects the dynamic characteristics of the current in the circuit or device under test as a function of time.

[0053] In one possible implementation, S304 specifically includes sub-steps S3041 and S3042:

[0054] S3041: Converts standardized current time-series data into intermediate metering values ​​based on a fixed mapping relationship.

[0055] S3042: Based on real-time reference values, intermediate measurement values ​​are converted into basic thickness data according to the physical correspondence between the radiation source energy and the hot-rolled alloy plate under test.

[0056] Among them, radioactive source energy refers to the radiation released by radioactive materials during decay (such as...). α , β , γ The energy carried by rays is usually measured in kiloelectron volts or megaelectron volts, and its magnitude determines the penetrating power and application scenarios of rays.

[0057] In this embodiment of the invention, by converting standardized current data into intermediate measurement values ​​and performing a secondary conversion based on a real-time reference, the effects of radioactive source decay and environmental interference are effectively isolated, achieving high-precision and stable measurement of basic thickness data.

[0058] It should be noted that, considering the characteristics of continuous steel rolling production and the presence of radioactive source decay, temperature fluctuations, and mechanical deviations, the system adapts to dynamic interference by "dynamically acquiring real-time reference values." Before each steel rolling operation begins, the system automatically collects the average current value of the hot-rolled plate without the alloy being measured as a real-time reference. Unlike traditional solutions, it does not rely on a fixed reference, avoiding the need for machine shutdown and equipment recalibration after reference drift. Simultaneously, through a fixed mapping of current to measurement value and a two-stage conversion to isolate interference, a fixed mapping reflecting only the inherent characteristics of the detector is first constructed based on a calibration plate of known thickness. This converts the real-time current during the steel rolling process into a measurement value unaffected by radioactive source decay and temperature fluctuations. The thickness is then calculated by combining this with the real-time reference value, avoiding the instability in accuracy caused by traditional direct current-thickness mapping. Furthermore, the fixed mapping and dynamic reference work together to adapt to changes in operating conditions at different times, ensuring measurement consistency. Existing technologies typically set a fixed reference at the initial stage of equipment installation, directly establishing a current-thickness mapping. This requires machine shutdown and recalibration every 8 hours or 2 days, which is complex and impacts production. In contrast, this invention eliminates the need for machine downtime for recalibration, is fully automated without human intervention, and significantly improves production efficiency and measurement stability.

[0059] In this embodiment of the invention, by establishing a collaborative mechanism of fixed mapping relationship and dynamically acquiring real-time reference values, the influence of dynamic interference such as radioactive source decay and temperature fluctuation on the measurement system is effectively isolated, realizing online adaptive thickness measurement without downtime calibration, and significantly improving the accuracy of basic thickness data and production continuity.

[0060] S4: Through deep learning, feature extraction is performed on standardized data to generate globally fused features.

[0061] Deep learning refers to a machine learning method based on multi-layer neural networks. It automatically learns high-level abstract representations of data through hierarchical feature extraction and nonlinear transformation, thereby realizing the recognition and modeling of complex patterns.

[0062] Feature extraction refers to the process of filtering and calculating a subset of information from the original data that can effectively characterize the key attributes of the object being tested.

[0063] Among them, global fusion features refer to the integration of features from different sources and dimensions through a specific algorithm to form a high-order representation that can comprehensively reflect the overall state of the object being tested.

[0064] In this embodiment of the invention, deep learning technology is used to extract and fuse physically meaningful features from multi-source heterogeneous data to construct a global fusion feature that can comprehensively characterize complex working conditions, providing high-dimensional feature support for subsequent error compensation.

[0065] In one possible implementation, S4 specifically includes sub-steps S401 to S403:

[0066] S401: Extract X-ray-physicochemical fusion characteristics based on standardized current time-series data, standardized temperature time-series data, and standardized alloy content data from the standardized data:

[0067] .

[0068] .

[0069] .

[0070] in, F R Indicates the original ray characteristics, Represents Long Short-Term Memory Network, R Represents standardized current timing data. M This represents the temperature-alloy cross-modulation factor. Represents the hyperbolic tangent function. Indicates a gated loop unit. T This represents standardized temperature time-series data. Indicates a fully connected layer. A This represents standardized alloy content data. F RT This indicates the ray-physicochemical fusion characteristics after cross-modulation. Indicates the scale parameter. Indicates a convolutional layer. It represents the Hadamardi (or Hadama) stack.

[0071] Temperature time series data refers to a sequence of temperature measurements collected continuously in chronological order, reflecting the dynamic process of temperature change of the measured object or environment over time.

[0072] Among them, alloy content data refers to quantitative information obtained through chemical analysis or spectroscopic detection, which represents the proportion or concentration of a specific chemical element in the alloy.

[0073] Among them, X-ray-physicochemical fusion characteristics refer to the high-order composite characteristics that can more comprehensively characterize the material state after cross-domain correlation and integration of X-ray measurement data and the physicochemical property data of alloys (such as composition and temperature).

[0074] Specifically, γ Both X-ray signals and temperature data exhibit time-series dynamic changes. Although alloy content data is relatively static, it has a significant coupling effect with temperature. The difficulty lies in the fact that the three do not affect X-ray characteristics independently, but are interconnected and interfere with each other. For example, the time-series drift of temperature will amplify the nonlinear effect of alloy content on X-ray attenuation. The dynamic fluctuation of time-series data also makes it difficult for feature extraction at a single time point to reflect the overall working condition. As a result, processing only one type of data alone cannot accurately obtain effective features that can support subsequent error compensation.

[0075] To address the dynamic fluctuation characteristics of current time-series data, this invention employs... LSTM The data is processed to obtain the original ray characteristics, because LSTM It has the ability to capture the long-term dependence and fluctuation patterns of time-series data, and can effectively retain key information about the changes of gamma-ray signals with the rolling process, laying the foundation for subsequent feature fusion.

[0076] Considering the temporal variation properties of temperature data, we selected... GRU The system processes time-series temperature data, capturing the temporal drift trend of temperature. Simultaneously, for alloy content percentage data, a fully connected layer maps it to a form matching the temperature feature dimension, ensuring effective computation between the two. tanh After function normalization, the temperature-alloy cross-modulation factor is obtained. The generation logic of this factor is based on the coupling effect of temperature and alloy on X-ray attenuation—temperature changes alter the activity level of alloy atoms, thus affecting the alloy's ability to absorb X-rays, while the dimensional adaptation of the fully connected layer and... tanh Normalization of the function transforms this coupling relationship into a quantization factor that can interact with the ray features.

[0077] The original ray features are subjected to a Hadamard product operation with the temperature-alloy cross-modulation factor to realize the correlation between the ray features and the temperature-alloy coupling effect. Then, the local correlation of features at adjacent time points is strengthened by a convolutional layer. The modulation intensity is controlled by the scale parameter. Finally, it is added to the original ray features. This not only preserves the basic information of the ray features, but also incorporates the cross-modulation effect of temperature-alloy, thereby solving the problem of inaccurate feature extraction caused by the coupling of the three factors. This ensures that the extracted ray-physicochemical fusion features can fully reflect the comprehensive influence of multiple factors under actual working conditions.

[0078] S402: Calculate the geometric path convolution kernel and geometric path features based on the standardized geometric parameters in the standardized data.

[0079] .

[0080] .

[0081] .

[0082] in, Represents the geometric path convolution kernel. express Softmax function, G Represents standardized geometric parameters, Represents the original geometric path convolution kernel. Represents the spatial feature map of geometric parameters. This indicates a dimension expansion operation. Indicates geometric path characteristics.

[0083] Among them, the geometric path convolution kernel refers to the key information that can reflect the structural morphology of the measured object, extracted from the spatial positional relationship between the measurement sensor and the measured object and the geometric properties of the ray penetration path.

[0084] Among them, geometric path features are a specially designed convolutional neural network operator whose parameter weights are associated with the specific geometric path through which the ray penetrates the measured object, and are used to adaptively extract depth features related to spatial structure from the raw measurement data.

[0085] Optionally, the standardized geometric parameters include convexity value, wedge angle, plate width, edge thinning amount, and peak position of transverse thickness distribution.

[0086] The convexity value refers to the positive difference between the thickness at the center point of the cross-section of the strip and the average thickness at two designated points on both sides, and is used to characterize the degree of central bulge in the cross-section of the strip.

[0087] The wedge angle refers to the ratio of the thickness difference between the two edges of the cross-section of the strip to the width of the strip, which is used to quantify the asymmetric thinning trend of the strip along the width direction.

[0088] Among them, the strip width refers to the overall width dimension of the strip perpendicular to the rolling direction, and it is one of the key process control parameters in the rolling process.

[0089] Among them, edge thinning refers to the negative difference between the thickness of a specific area at the edge of the strip and the thickness of the central area, which is used to characterize the edge thickness reduction phenomenon caused by stress concentration during the rolling process.

[0090] Among them, the peak position of the transverse thickness distribution refers to the offset distance of the point with the maximum thickness on the cross-section of the strip relative to the width centerline, which is used to determine the symmetry and skewness characteristics of the thickness distribution curve.

[0091] It should be noted that this scenario is a geometric path feature extraction scenario for hot-rolled alloy plates, characterized by its geometric parameters. G The study incorporates multi-dimensional parameters such as convexity, wedge angle, plate width, edge thinning, and peak position of transverse thickness distribution. Each parameter is directly related to the actual path length of gamma rays penetrating the hot-rolled alloy plate under test. For example, the convexity determines the degree of convexity in the transverse thickness of the hot-rolled alloy plate, the wedge angle affects the tilt distribution of the thickness, and the plate width and edge thinning together determine the ray penetration path in the edge region. The combined effect of these parameters directly leads to the spatial heterogeneity of the ray path. The influence of each parameter on the path is interdependent; processing a single parameter alone will miss the interrelationship between parameters, resulting in the extracted geometric features failing to accurately reflect the actual ray path, thus affecting the effectiveness of subsequent global feature fusion.

[0092] This invention addresses geometric parameters G The multidimensional characteristics and scalar / low-dimensional vector form are achieved through fully connected layers. G The process involves integrating parameters of different dimensions, such as convexity and wedge angle, into a feature vector of a unified dimension, thereby achieving the initial fusion of multi-dimensional geometric information and laying the foundation for subsequent transformation.

[0093] When calculating the geometric path convolution kernel, the output of the fully connected layer is passed through... Softmax The function transforms the integrated geometric feature vector into a weight distribution, highlighting parameters that have a more significant impact on the ray path (such as the effect of convexity on the path to the center of the plate, and the effect of wedge angle on the path in the direction of plate inclination). This weight distribution is then multiplied by the original convolution kernel, allowing the geometric path convolution kernel to carry path-sensitive information assigned by the geometric parameters. That is, different combinations of geometric parameters will correspond to different weight distributions in the geometric path convolution kernel; for example, as the wedge angle increases... Softmax The corresponding weights of the output are increased, and the convolution weights of the geometric path convolution kernel in the tilt direction are also enhanced, so that the convolution kernel can adapt to the path tilt characteristics caused by the wedge.

[0094] To address the problem of geometric parameters not matching spatial feature maps, by... Expand The dimension expansion operation processes the output of the fully connected layer, transforming the low-dimensional geometric feature vector into a geometric parameter space feature map, thus realizing the transformation of geometric information from non-spatial to spatial.

[0095] Finally, the spatialized geometric parameter spatial feature map is convolved by a convolutional layer using a geometric path convolution kernel carrying path information. At this point, the convolution process is not only a conventional feature extraction, but also a deep integration of geometric parameters and spatial path features. The geometric parameter spatial feature map provides the spatial distribution of geometric parameters, and the geometric path convolution kernel provides path-sensitive convolution weights. The geometric path features obtained after the interaction of the two can accurately characterize the ray path features under different combinations of geometric parameters. For example, the region of the geometric parameter spatial feature map corresponding to the peak position of the lateral thickness distribution will be enhanced under the action of the geometric path convolution kernel, thereby accurately reflecting the ray path length at that position. This effectively solves the technical problem of geometric parameter dimension mismatch and its impact on coupling.

[0096] S403: By using feature intermodulation and attention mechanisms, ray-physicochemical fusion features and geometric path features are fused to generate global fused features.

[0097] .

[0098] .

[0099] .

[0100] in, Indicates the first intermodulation characteristic. This indicates the second intermodulation characteristic. Represents a hyperbolic sine function. Indicates global fusion features, This represents the attention mechanism.

[0101] Feature intermodulation refers to a computational mechanism that enables bidirectional information exchange between different feature vectors through specific operations, thereby generating a new feature representation that includes cross-sensing capabilities.

[0102] Among them, the attention mechanism is a resource allocation strategy that mimics human cognition. By dynamically calculating the importance weights of features, the model can adaptively focus on the key parts of the input information.

[0103] In this embodiment of the invention, a multimodal deep feature fusion architecture is used to achieve deep interaction and adaptive weighting of ray physical properties, material physicochemical properties and geometric path features, and to construct a global fusion feature that can comprehensively characterize complex rolling conditions, providing a high-dimensional feature representation with cross-domain correlation for thickness measurement.

[0104] S5: Based on the global fusion characteristics and standardized data, perform error compensation calculations and determine the error compensation factor.

[0105] Error compensation calculation refers to the data processing process of quantitatively estimating and correcting the deviation between the system's measured value and the true value of the measured quantity by establishing an error model and algorithm.

[0106] Among them, the error compensation factor refers to the correction coefficient generated in the error compensation calculation, which is used to quantitatively characterize the magnitude and direction of a specific deviation. It eliminates systematic errors by performing mathematical operations with the original measurement value.

[0107] In this embodiment of the invention, global fusion features are associated with multi-source standardized data through error compensation calculation, and a deep learning-based error tracing and compensation mechanism is established, which significantly improves the accuracy and anti-interference capability of thickness measurement.

[0108] In one possible implementation, S5 specifically includes sub-steps S501 to S503:

[0109] S501: Based on standardized temperature time-series data, standardized alloy content data, and standardized geometric parameters from the standardized data, error basis features are extracted:

[0110] .

[0111] .

[0112] .

[0113] in, This represents the basic characteristic of temperature error. T This represents standardized temperature time-series data. Represents an exponential function. Indicates the temperature decay coefficient. Indicates the basic characteristics of alloy error. Indicates a fully connected layer. It represents the Hadamah accumulation. This represents the basic characteristic of thickness error. Indicates vector normalization, This represents the gradient of the standardized geometric parameters.

[0114] Among them, error basis features refer to characteristic parameters extracted from raw data or intermediate calculation results that can directly characterize or indirectly reflect the source and variation law of system measurement error.

[0115] S502: Calculate the error-sensitive features based on the error basis features and the global fusion features:

[0116] .

[0117] .

[0118] .

[0119] in, Indicates temperature error sensitivity. This represents a one-dimensional convolutional layer. Indicates global fusion features, Indicates the alloy's error-sensitive characteristics. Indicates a thickness error sensitive characteristic. Indicates feature splicing, Indicates geometric path characteristics.

[0120] Among them, error-sensitive characteristics refer to those key parameters that have significant response characteristics to measurement system errors or environmental disturbances, and their small changes can directly expose or amplify the potential errors of the system.

[0121] S503: Calculate the error compensation factor based on error sensitivity characteristics:

[0122] .

[0123] in, C Indicates the error compensation factor. This represents the activation function.

[0124] Specifically, the error feature extraction and compensation factor calculation scenario is characterized by the fact that temperature time series data, alloy content data, and geometric parameters correspond to three types of core error sources, and these three types of error sources do not act independently.

[0125] Firstly, considering the time-series nonlinear characteristics of temperature error, a method combining temperature time-series data and an exponential function is used to calculate the basic characteristics of temperature error, since temperature affects... γThe effect of X-ray attenuation follows an exponential law; the X-ray attenuation coefficient changes exponentially with increasing temperature, leading to a non-linear increase in thickness measurement error with temperature. An exponential function can be used to capture this non-linear correlation. By multiplying the temperature time-series data with the exponential term, the temperature error basis feature can directly reflect the fundamental error attributes under different temperature time-series states, providing a physically sound basis for the subsequent generation of temperature error sensitive features. For the non-linear saturation characteristics of alloy errors, element-level Hadamard product operations are first performed on the alloy content data, and then the result is processed through a fully connected layer. This integrates multi-dimensional alloy content data into a single-dimensional alloy error basis feature, effectively condensing the fundamental information of alloy errors. Regarding the spatial variation characteristics of geometric parameters, gradient operations are performed on the geometric parameters, followed by normalization to obtain the thickness error basis feature. This highlights the spatial variation rate of geometric parameters (such as lateral thickness variation caused by convexity and tilt thickness variation caused by wedge angle). Normalization ensures that the gradient values ​​of different geometric parameters are within a uniform range, avoiding imbalance in the error basis feature weights due to differences in parameter magnitudes, allowing the thickness error basis feature to accurately reflect the fundamental error information corresponding to geometric spatial variations.

[0126] Secondly, different global fusion methods are designed to generate error-sensitive features based on the attributes of different error basis features: For temperature error basis features, the one-dimensional convolution result of the global fusion feature is used to perform a Hadamard product operation with it before being superimposed on the temperature error basis feature, so that the temperature error-sensitive feature carries both basic temperature error information and incorporates the global temporal interaction effect. For alloy error basis features, the squared result of the fully connected component of the global fusion feature is multiplied by the alloy error basis feature, and then superimposed on the alloy error basis feature. The squaring operation can enhance the nonlinear modulation effect of global operating conditions on alloy errors. When the temperature rises or the radiation intensity fluctuates greatly in the global operating conditions, the squared term of the fully connected result will amplify this modulation effect, which conforms to the actual law that "changes in global operating conditions exacerbate the nonlinearity of alloy errors," so that the alloy error-sensitive feature can accurately reflect the changes in alloy errors under global information. For thickness error basis features, the convolution result of the thickness error feature with the global fusion feature is first concatenated to integrate the basic information of geometric spatial errors and the global spatial correlation information, and then a Hadamard product operation is performed with the geometric path feature to ensure that the thickness error-sensitive feature can reflect the global error correlation in the spatial dimension and avoid the geometric error feature from being disconnected from the global information.

[0127] Finally, by splicing together the three error-sensitive features, sensitive information from three error sources—temperature, alloy, and geometry—can be comprehensively aggregated. This multi-source error information is then integrated through a fully connected layer. SigmoidThe function normalizes the integrated results to a reasonable range, and the resulting error correction factor is used to quantify the coupling interaction strength of the three types of error sources. When the interaction between two or three types of error sources is significant (such as high temperature + high alloy content + large convexity), the spliced ​​features will carry stronger interaction information, solving the technical problems of inaccurate feature extraction and shallow fusion of multiple error sources.

[0128] In this embodiment of the invention, by constructing a multi-level calculation framework of error base features and error sensitive features, the accurate quantification and coupling analysis of multi-source errors such as temperature, alloy composition and geometry are realized, thereby generating an adaptive error compensation factor, which significantly improves the anti-interference ability and accuracy stability of the thickness measurement system under complex working conditions.

[0129] S6: Based on the error compensation factor, the basic thickness data is corrected to determine the accurate thickness measurement data of the hot-rolled alloy plate to be tested.

[0130] Specifically, the basic thickness calculation formula derived from Lambert-Beer's law is called, the pre-processed gamma-ray intensity data is input, and the error compensation parameters corresponding to different thicknesses of the hot-rolled plate of the alloy to be tested are incorporated. Based on the error correction factor, the basic calculation results are corrected, and finally the accurate thickness measurement value of the thick hot-rolled plate is obtained.

[0131] In this embodiment of the invention, the basic thickness data is adaptively corrected by an error compensation factor, which effectively eliminates the coupled effects of multiple sources of error such as temperature fluctuations, alloy composition changes and geometric deformation, and finally realizes high-precision online measurement of the thickness of alloy hot-rolled plates.

[0132] In one possible implementation, S6 specifically includes sub-steps S601 and S602:

[0133] S601: The basic calculation result of the basic thickness data is calculated using the basic thickness calculation formula.

[0134] S602: Based on the error compensation factor, the basic calculation results are corrected to obtain accurate thickness measurement data.

[0135] In this embodiment of the invention, the basic thickness calculation results are adaptively corrected by an error compensation factor, which effectively eliminates multi-source system errors and ultimately achieves high-precision online measurement of the thickness of alloy hot-rolled plates.

[0136] In practical applications, the traditional static calibration is first upgraded to a dynamic system that adapts to changing working conditions through a dynamic benchmark and a two-stage mapping mechanism. Real-time benchmark values ​​and a fixed mapping relationship are used to isolate equipment drift interference and establish stable basic thickness data. Then, features are extracted from X-ray signals, physicochemical parameters, and geometric paths using a multimodal deep learning network. Feature intermodulation and attention mechanisms are employed to achieve cross-domain feature fusion, forming a globally fused feature that comprehensively reflects the working conditions. Finally, feature basis vectors for three error sources—temperature, alloy, and geometry—are constructed. Error-sensitive features are generated based on these global features, and compensation factors are synthesized to achieve precise correction of the basic thickness. This approach overcomes the limitations of traditional static calibration, achieving continuous self-calibration through a dynamic benchmark mechanism, solving the problem of measurement accuracy degradation caused by radioactive source decay and temperature drift. Multimodal feature fusion overcomes the insufficient characterization capability of a single data source, significantly improving feature representation capabilities under complex working conditions. The established mathematical model of the error propagation link realizes closed-loop control from error tracing to compensation, improving the accuracy of thickness measurement.

[0137] The beneficial effects of this invention are reflected in the fact that by adopting a calibration algorithm based on dynamic benchmarks and mapping, it effectively overcomes the measurement drift problem caused by dynamic changes in rolling conditions in existing static calibration models, and significantly improves the long-term accuracy and stability of online thickness measurement. At the same time, by using deep learning technology to perform deep feature extraction and global fusion on massive and noisy measurement data, it breaks through the limitations of traditional single-point measurement and simple linear compensation, and can accurately mine and compensate for comprehensive errors under complex working conditions, thereby improving the stability and accuracy of thickness measurement results in real industrial environments.

[0138] Example 2:

[0139] Reference Figure 8 The present invention provides an online thickness measurement system 20 for hot-rolled alloy plates, comprising: a processor 201 and a memory 202.

[0140] The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the online thickness measurement method for hot-rolled alloy plates as described in Example 1 and achieve the same technical effect. To avoid repetition, the present invention will not repeat the above.

[0141] In the description of the embodiments of the present invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "center," "top," "bottom," "top," "bottom," "inner," "outer," "inner side," and "outer side," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. "Inner side" refers to the interior or enclosed area or space. "Outer perimeter" refers to the area surrounding a specific component or specific area.

[0142] In the description of embodiments of the present invention, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature. In the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0143] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "joining," and "assembly" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0144] In the description of embodiments of the present invention, specific features, structures, materials or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0145] In the description of the embodiments of the present invention, it should be understood that "-" and "~" represent a range of two numerical values, and this range includes the endpoints. For example, "AB" represents a range greater than or equal to A and less than or equal to B. "A~B" represents a range greater than or equal to A and less than or equal to B.

[0146] In the description of embodiments of the present invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online thickness measurement of an alloy hot-rolled sheet, characterized by, The method comprises the following steps: S1: obtaining measurement data of a hot-rolled plate of an alloy to be measured; S2: preprocessing the measurement data to obtain standardized data; S3: performing calibration calculation on the standardized data through a dynamic reference and a two-segment mapping calibration algorithm to determine the basic thickness data of the hot-rolled plate of the alloy to be measured; S4: performing feature extraction on the standardized data through deep learning to generate global fusion features; S5: performing error compensation calculation according to the global fusion features and the standardized data to determine an error compensation factor; S6: correcting the basic thickness data according to the error compensation factor to determine the accurate thickness measurement data of the hot-rolled plate of the alloy to be measured; The S3 specifically comprises: S301: collecting current data of multiple different thickness calibration plates with known thicknesses; S302: establishing a fixed mapping relationship between current and measurement value according to the current data; S303: dynamically obtaining a real-time reference value; S304: converting standardized current time series data in the standardized data according to the real-time reference value and the fixed mapping relationship to obtain the basic thickness data.

2. The alloy hot-rolled plate on-line thickness measurement method according to claim 1, characterized by, The measurement data includes The S302 specifically comprises: The current time series data, the geometric parameter, the alloy content data and the temperature time series data of the ray passing through the alloy hot-rolled plate to be measured.

3. The alloy hot-rolled plate on-line thickness measurement method according to claim 1, characterized by, The S3022: based on the current data and the theoretical measurement value, the fixed mapping relationship is established. S3021: calculating the theoretical measurement value corresponding to the theoretical residual energy of the rays after passing through the calibration plate of different thickness according to the physical characteristics of the radioactive source The S304 specifically comprises: theoretical measurement value corresponding to the theoretical residual energy of the rays after passing through the calibration plate of different thickness S3041: converting the standardized current time series data into intermediate measurement values according to the fixed mapping relationship; 4. The alloy hot-rolled plate on-line thickness measurement method according to claim 1, characterized by, S3042: based on the real-time reference value, converting the intermediate measurement values into the basic thickness data according to the physical correspondence between the radiation source energy and the hot-rolled plate of the alloy to be measured. The S4 specifically comprises: S401: extracting ray-physicochemical fusion features according to standardized current time series data, standardized temperature time series data and standardized alloy content data in the standardized data:

5. The alloy hot-rolled plate on-line thickness measurement method according to claim 1, characterized by, S402: calculating geometric path convolution kernels and geometric path features according to standardized geometric parameters in the standardized data; Softmax ； ； ； wherein, F R denotes original ray features, denotes long short-term memory network, R denotes normalized current time series data, M denotes temperature-alloy cross modulation factor, denotes hyperbolic tangent function, denotes gated recurrent unit, T denotes normalized temperature time series data, denotes fully connected layer, A denotes normalized alloy content data, F RT denotes ray-physicochemical fusion features after cross modulation, denotes scale parameter, denotes convolution layer, denotes Hadamard product; S403: fusing the ray-physicochemical fusion features and the geometric path features through feature intermodulation and attention mechanism to generate the global fusion features: ； ； ； wherein, denotes a geometric path kernel, denotes The standardized geometric parameters include convexity value, wedge angle, plate width, edge thinning amount and transverse thickness distribution peak position. a function, G denotes a normalized geometric parameter, denotes an original geometric path kernel, denotes a geometric parameter space feature map, denotes a dimension expansion operation, denotes a geometric path feature; The S5 specifically comprises: ； ； ； wherein, denotes a first intermodulation feature, denotes a second intermodulation feature, denotes a hyperbolic sine function, denotes a global fusion feature, denotes an attention mechanism.

6. The alloy hot-rolled plate on-line thickness measurement method according to claim 5, characterized by, S501: extracting error base features based on standardized temperature time series data, standardized alloy content data and standardized geometric parameters in the standardized data:

7. The alloy hot-rolled plate on-line thickness measurement method according to claim 1, characterized by, S502: calculating error sensitive features according to the error base features and the global fusion features: S503: calculating the error compensation factor based on the error sensitive features: ； ； ； wherein, represents a temperature error base feature, T represents normalized temperature time series data, represents an exponential function, represents a temperature decay coefficient, represents an alloy error base feature, represents a fully connected layer, represents a Hadamard product, represents a thickness error base feature, represents vector normalization, represents a gradient of normalized geometry parameters; The S6 specifically comprises: ； ； ； wherein, represents a temperature error sensitive feature, represents a one-dimensional convolutional layer, represents a global fusion feature, represents an alloy error sensitive feature, represents a thickness error sensitive feature, represents feature concatenation, represents a geometric path feature; S601: calculating a basic calculation result of the basic thickness data through a basic thickness calculation formula; ； wherein, C denotes an error compensation factor, denotes an activation function.

8. The alloy hot rolled plate on-line thickness measurement method according to claim 1, characterized by, S602: correcting the basic calculation result based on the error compensation factor to obtain the accurate thickness measurement data. The method comprises the following steps: processor and memory; 9. An alloy hot rolled plate on-line thickness measurement system characterized by, ​ ​ The memory stores a program or instructions executable on the processor, which, when executed by the processor, implements the steps of the alloy hot-rolled plate on-line thickness measurement method according to any one of claims 1 to 8.

Images