Metal element detection method for metal material

Abstract

The invention relates to the technical field of metal material component detection, and discloses a metal element detection method for a metal material. The method comprises the following steps: establishing a dynamic detection path for adjusting detection point space distribution in real time according to a material form, and driving a probe to scan and collect an original spectral signal flow; and performing real-time quality evaluation on the signal flow, selectively activating a signal enhancement mechanism according to an evaluation result, and generating a high-quality signal sequence. And decoupling the feature component of each element from the sequence by using an element feature decoupling network, calculating a quantitative index, and matching the quantitative index with a standard reference to identify an abnormal element. And for the abnormal elements, starting a depth verification process adopting the differentiated excitation parameters to perform re-detection. And finally, fusing the initial and verification results, generating an element content distribution map, and updating a material database. According to the method, adaptive optimization of the detection process and accurate re-check of abnormal results are realized, and the accuracy and reliability of detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal material composition detection, in particular to a metal element detection method for metal materials. BACKGROUND

[0002] In the process of element detection and analysis of metal materials, spectral analysis technology is widely used due to its rapidity and non-destructive characteristics. However, the existing detection methods usually rely on preset fixed paths or simple grid scanning strategies for data acquisition. This static sampling method is difficult to adapt to the complex situations that may exist on the surface of actual metal materials, such as morphological fluctuations, uneven thickness of oxidation layer, etc., resulting in insufficient sampling in key areas or poor signal quality in areas being treated equally, and the overall quality of the original spectral signals collected is unstable and inconsistent.

[0003] The existing technology generally uses a unified and fixed preprocessing process for the collected spectral signals, such as using the same filtering parameters and enhancement algorithms to process all signals. This processing method cannot distinguish the quality of the signals, and for signals with good quality, unnecessary calculations or even distortion may be introduced, while for signals with very low signal-to-noise ratio or severe interference, the processing may not be sufficient, and effective feature information cannot be effectively extracted, affecting the accuracy of subsequent quantitative analysis. When preliminary detection finds that the element content deviates from the standard range, the conventional recheck method usually repeats the measurement several times at the original detection point or adjacent positions using the same instrument parameters to take the average value. This simple repeated verification cannot rule out the systematic errors caused by matrix effect, spectral interference between elements or transient instrument state fluctuations in the initial detection, so that there is a risk of misjudgment or omission in the judgment of abnormal elements. SUMMARY

[0004] The purpose of the present application is to provide a metal element detection method for metal materials to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides a metal element detection method for metal materials, which comprises: establishing a dynamic detection path for metal materials, the dynamic detection path adjusts the spatial distribution of detection points in real time according to the material morphology; driving the detection probe to scan according to the dynamic detection path, and synchronously collecting the original spectral signal stream returned by the detection probe; performing signal quality evaluation on the original spectral signal stream, and activating a signal enhancement mechanism to generate an enhanced signal sequence according to the evaluation result; inputting the enhanced signal sequence into an element feature decoupling network to decouple the feature components corresponding to different metal elements; calculating the quantitative indicators of each metal element based on the decoupled feature components. Matching the quantification index with the benchmark range in the material standard database identifies abnormal elements with deviations; Starting a deep verification process for abnormal elements, which uses differentiated excitation parameters for re-detection; Fusing the initial detection results and the output of the deep verification process to generate a final element content distribution map; Updating the historical records in the material standard database according to the final element content distribution map.

[0006] Preferably, the establishment of the dynamic detection path for the metal material includes: Obtaining three-dimensional surface topological data of the metal material, and performing curvature feature extraction on the three-dimensional surface topological data; Determining a high-probability defect area according to the curvature features, and arranging a dense detection point array in the high-probability defect area; Generating a sparse detection point array in a non-high-probability defect area using an adaptive grid division algorithm; Optimizing and splicing the dense detection point array and the sparse detection point array to form a complete dynamic detection path.

[0007] Preferably, the signal quality assessment on the original spectrum signal stream includes: Segmenting the original spectrum signal stream, and calculating a signal-to-noise ratio index of each signal segment; When the signal-to-noise ratio index is lower than a preset threshold, triggering a signal enhancement mechanism; The signal enhancement mechanism uses a multi-scale filtering algorithm to perform noise reduction processing on the signal; Performing baseline correction on the noise-reduced signal to eliminate background interference components; Splicing the processed signal segments into an enhanced signal sequence through a signal reconstruction algorithm.

[0008] Preferably, the operation of inputting the enhanced signal sequence into the element feature decoupling network includes: Constructing a feature template library containing standard spectra of multiple metal elements; Using a sparse coding algorithm to decompose the enhanced signal sequence into a linear combination of feature templates; Extracting the weight coefficients corresponding to each feature template through an orthogonal matching pursuit algorithm; Determining the feature templates with weight coefficients greater than an activation threshold as effective element features; Generating a feature component set according to the effective element features.

[0009] Preferably, the calculation of the quantification index of each metal element includes: Performing integral intensity calculation on each feature component to obtain an original intensity value; The original intensity value is corrected by using an internal standard method to eliminate the influence of instrument fluctuation. The corrected intensity value is converted into an element concentration value through a standard curve. The concentration values of multiple detections of the same element are subjected to statistical consistency test. The arithmetic mean of the concentration values passing the consistency test is taken as the final quantification index.

[0010] Preferably, the depth verification process comprises: Adjusting the energy parameters of the excitation source and performing excitation in a stepwise energy scanning mode; Changing the collection angle of the detection probe to obtain multi-angle spectral signals; Independently analyzing the multi-angle spectral signals to obtain a set of verification concentration values; Calculating the deviation coefficient of the set of verification concentration values and the initial quantification index, and confirming the validity of the abnormal element detection result when the deviation coefficient is within the allowable range.

[0011] Preferably, the generation of the final element content distribution map comprises: Mapping the confirmed valid abnormal element concentration values back to the corresponding dynamic detection path coordinates; Using the Kriging interpolation algorithm to construct an element concentration distribution surface of the entire material surface; Extracting the contour lines of the element concentration distribution surface to identify the concentration gradient change area; Integrating the distribution surfaces of multiple metal elements into a multi-level element content distribution map.

[0012] Preferably, the updating of the material standard database according to the final element content distribution map comprises: Extracting statistical characteristic parameters from the final element content distribution map and performing similarity matching between the statistical characteristic parameters and historical data in the material standard database; When the similarity is lower than the update threshold, starting the database adaptive update mechanism and using the sliding window algorithm to retain the latest batch of detection data; Recalculating the upper and lower limits of the reference range in the material standard database.

[0013] Preferably, the signal enhancement mechanism uses a multi-scale filtering algorithm for noise reduction processing of the signal, comprising: Performing wavelet transform on the original spectral signal stream to decompose the signal into multiple scale wavelet coefficient sequences; For each scale wavelet coefficient sequence, a scale-dependent noise threshold is calculated, and soft threshold processing is performed on the wavelet coefficients to eliminate noise components; The threshold-processed wavelet coefficient sequence is reconstructed into a noise-reduced signal segment through inverse wavelet transform.

[0014] Preferably, the extracting the weight coefficient corresponding to each feature template by the orthogonal matching pursuit algorithm comprises the following steps: initializing the residual as the enhanced signal sequence, initializing the activated atom set as an empty set, and initializing the weight coefficient vector as a zero vector; iteratively performing the following steps until the norm of the residual is lower than a preset threshold or the maximum number of iterations is reached: calculating the inner product of the current residual and each feature template in the feature template library, and selecting the feature template with the maximum absolute value of the inner product as the activated atom of the current iteration; adding the current activated atom to the activated atom set; solving the linear approximation of the activated atom set to the enhanced signal sequence by the least square method, and calculating the weight coefficient corresponding to the current activated atom set; updating the residual as the difference between the enhanced signal sequence and the weighted sum of the activated atom set; outputting the finally obtained weight coefficient vector as the weight coefficient corresponding to the feature template.

[0015] Compared with the prior art, the present application has the following advantages: The original spectral signal stream is evaluated in real time and online, and the signal enhancement mechanism is dynamically determined and activated according to the evaluation results. This mechanism realizes the transition of the data processing strategy from passive fixation to active adaptation. The signal quality evaluation module continuously monitors the signal-to-noise ratio and intensity stability of the signal stream. Only when the indicators are lower than the preset threshold, the corresponding enhancement algorithm such as adaptive filtering or cumulative averaging is triggered. This selective activation mode based on real-time feedback avoids the waste of computing resources and potential interference to high-quality signals caused by indiscriminate processing of all data. For signals of qualified quality, the system directly releases them, ensuring the data processing efficiency. It optimizes the allocation of computing resources, concentrates the processing capacity on signals of poor quality, and thus improves the effectiveness and reliability of the signal sequence, laying a solid foundation for high-quality data for subsequent accurate decoupling.

[0016] For the abnormal elements identified by preliminary detection, a differentiated excitation parameter is used to start the deep verification process, rather than simply repeating under the same conditions. The core of this process is to customize the excitation conditions according to the physical and chemical properties of the abnormal elements. This parameter differentiation based on the characteristics of the target object changes the mode of interaction between the excitation source and the sample, aiming to overcome the matrix effect or spectral line overlap interference that may exist in the initial detection. It can obtain the response signal of the abnormal element under optimized conditions, which is complementary or contrasted with the initial detection result, thereby improving the confidence of the abnormal element content determination and reducing the possibility of misjudgment due to the limitations of a single detection condition, making the recheck result more convincing and accurate. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The working principle diagram of the metal element detection method for metal materials; Figure 2 The flow chart for constructing a dynamic detection path; Figure 3 The contrast diagram of the metal material spectrum signal enhancement effect; Figure 4 The flow chart of the element characteristic decoupling network processing; Figure 5 The metal element detection depth verification bias coefficient statistical diagram. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] Please refer to Figure 1 The present application provides a metal element detection method for metal materials, which comprises: a dynamic detection path needs to be established according to the actual physical form of the detected metal material. The path is not fixed and can be adjusted in real time according to the three-dimensional topographic features of the material surface to adjust the spatial distribution density and position of the detection points on the material surface. After the dynamic detection path is established, the detection probe is controlled to move automatically according to the path, and the original spectrum signal stream returned by the detection probe is collected synchronously and continuously during the movement of the probe. The signal stream contains the spectrum information of each point on the material surface. The collected original spectrum signal stream is evaluated in real time or near real time for signal quality, and the main evaluation index is signal-to-noise ratio. Once the evaluation result indicates that the signal quality does not meet the preset requirement, the built-in signal enhancement mechanism is activated immediately to process the signal, and finally an enhanced signal sequence with improved quality is output.

[0020] The enhanced signal sequence is input into a pre-trained element feature decoupling network, which can separate and identify the characteristic spectral components corresponding to different metal elements from the complex spectral signal. Based on the decoupled characteristic components of each element, a quantitative analysis algorithm is used to calculate the content quantitative indicators of each metal element, such as concentration values. These quantitative indicators are compared with the reference content range of the corresponding elements pre-stored in the material standard database, thereby identifying abnormal elements whose content exceeds the allowable deviation range. For these abnormal elements, in order to ensure the reliability of the detection results, a deep verification process needs to be started, which re-detects by adjusting the excitation parameters of the detection to exclude accidental errors. The initial detection results and the deep verification results are data fused to generate a final element content distribution map that can intuitively reflect the distribution of elements on the material surface, and the results of this detection are used to update the historical records in the material standard database, realizing the self-learning and optimization of the database.

[0021] Embodiment 1: see Figure 2 In specific implementation, the establishment of the dynamic detection path for metal materials starts from obtaining the accurate three-dimensional topographic information of the metal material surface. Through a high-precision non-contact three-dimensional scanning device such as a laser line scanner or a structured light three-dimensional scanner, the metal material sample placed on the detection platform is scanned in all directions. The laser line scanner projects a laser line onto the metal material surface. Due to the ups and downs of the metal material surface, the laser line will be deformed. By capturing the deformed laser line with an industrial camera, combined with the relative position parameters of the camera and the laser, the three-dimensional coordinates of each point on the laser line can be calculated. The structured light three-dimensional scanner projects a series of coded light spots or grating patterns onto the metal material surface. By analyzing the distortion of the pattern, three-dimensional point cloud data is calculated. The scanning process needs to cover the entire area to be detected of the metal material. The final output is a three-dimensional surface topological data point cloud composed of a large number of three-dimensional space points.

[0022] In a specific implementation, curvature feature extraction on the obtained three-dimensional surface topology data point cloud is a key step for determining the distribution of detection points. The curvature feature extraction algorithm is based on the principle of local surface fitting. For each data point in the three-dimensional surface topology data point cloud, a number of points in its neighborhood are selected, and a local quadratic surface, such as a biquadratic surface, is fitted using the least squares method. By calculating the first and second fundamental forms of the fitted surface, the principal curvature, Gaussian curvature, and mean curvature of the data point can be derived. High curvature regions usually correspond to places where the curvature of the surface is large, such as edges, corners, concave or convex parts, and these regions are often stress concentration areas or areas prone to defects during processing. The extracted curvature values need to be normalized to facilitate the setting of thresholds to distinguish between high curvature regions and low curvature regions. It can be understood that the accuracy of curvature feature extraction directly depends on the quality of the three-dimensional surface topology data point cloud and the robustness of the local surface fitting algorithm.

[0023] In a specific implementation, the high-probability defect region and the non-high-probability defect region are divided according to the calculated curvature feature map. A curvature threshold is set, and regions with a curvature absolute value greater than the threshold are marked as high-probability defect regions, which require more dense detection points to capture possible small defects or composition segregation. In the high-probability defect region, a regular grid subdivision method is used to arrange a dense detection point array, and the spacing of the dense detection point array can be dynamically adjusted according to the size of the curvature. The larger the curvature of a sub-region, the smaller the point spacing of the dense detection point array. For non-high-probability defect regions with a curvature absolute value less than the threshold, they are identified as relatively flat or uniform regions, and an adaptive grid division algorithm is used to generate a sparse detection point array. The adaptive grid division algorithm uses the curvature change gradient as the basis for grid subdivision, generating larger grid cells in regions with a gentle curvature change, and the vertices of the grid cells are the detection points of the sparse detection point array. In transition regions with a slightly larger curvature change, the grid is moderately subdivided to increase the density of the sparse detection point array. It can be understood that by dividing the regions and assigning different densities of detection point arrays in a curvature-driven manner, the overall scanning efficiency can be significantly improved while ensuring detection accuracy.

[0024] In a specific implementation, the generated dense and sparse point arrays are path-optimized and stitched in the 3D space to form a continuous and efficient dynamic scanning path. The path-optimized stitching needs to solve the optimal order problem of the detection probe accessing all the detection points, with the goal of minimizing the total moving distance and time of the detection probe while avoiding collisions with the metal material sample or fixture. The path-optimized algorithm first regards all the detection points in the dense and sparse point arrays as a whole point set. Then a variant algorithm of the Traveling Salesman Problem model is applied for global path planning, for example, the nearest neighbor algorithm is used to quickly generate an initial path, which always selects the unvisited detection point closest to the current detection point as the next moving target. Further, local search algorithms such as 2-opt or 3-opt can be used to optimize the initial path by exchanging edges in the path to gradually shorten the total path length. Optionally, for large-scale detection point sets, more advanced meta-heuristic algorithms such as genetic algorithms or ant colony algorithms can be used to search for approximately optimal detection paths. The collision detection module runs continuously during the entire path planning process to ensure the safety of the generated dynamic scanning path in the physical space. The final output dynamic scanning path contains a series of ordered 3D coordinate point sequences and the moving trajectory of the detection probe between points.

[0025] In a specific implementation, the generation process of the dynamic scanning path needs to be deeply integrated with the detection control software. The 3D surface topology data point cloud obtained by the 3D scanning device is directly imported into the path planning module. A series of steps such as curvature feature extraction, region division, point array generation, and path optimization stitching are automatically completed by software algorithms without human intervention. Users can preset key parameters in the software interface, such as the curvature threshold, the minimum point spacing of the dense detection point array, the initial grid size of the sparse detection point array, and the objective function of path optimization. Optionally, the software can provide a path simulation preview function to allow users to check the rationality and safety of the dynamic scanning path in a virtual environment. The final generated dynamic scanning path file is transmitted to the motion control system, which will accurately drive the detection probe to execute the scanning task according to the dynamic scanning path file.

[0026] In a specific implementation, the signal quality assessment on the original spectral signal stream, which is a time-intensity sequence continuously acquired by the detection probe during the scanning along the dynamic detection path, starts with a signal segmentation process. The segmentation process divides the continuous original spectral signal stream into multiple signal segments according to the time stamp or detection point trigger signal, each signal segment corresponding to a detection point or a very short scanning interval. For each signal segment obtained by the segmentation process, the signal-to-noise ratio index is calculated as the core parameter for assessing the signal quality. The signal-to-noise ratio index is calculated by extracting the average intensity of the preset feature peak region in the signal segment, and at the same time selecting the intensity standard deviation of the non-feature background region near the feature peak. The ratio of the feature peak average intensity to the background standard deviation is taken as the signal-to-noise ratio index value of the signal segment. The calculated signal-to-noise ratio index value is compared with the preset signal-to-noise ratio threshold value, which is the minimum required value based on a large amount of experimental data statistics to ensure the accuracy of subsequent quantitative analysis. When the signal-to-noise ratio index of a certain signal segment is lower than the preset signal-to-noise ratio threshold value, it is determined that the quality of the signal segment is unqualified, and the signal enhancement mechanism is triggered immediately to process the signal segment individually.

[0027] In a specific implementation, the signal enhancement mechanism first uses a multi-scale filtering algorithm to denoise the signal segment with an unqualified signal-to-noise ratio index. The multi-scale filtering algorithm is based on wavelet transform theory, and selects an appropriate wavelet basis function to perform multi-level decomposition on the input signal segment. The signal segment is decomposed into wavelet coefficient sequences at different frequency scales, including high-frequency detail coefficients and low-frequency approximation coefficients. For each layer of scale wavelet coefficient sequence obtained by decomposition, a scale-dependent noise threshold needs to be calculated, and the noise threshold is usually calculated considering the median value of the layer scale wavelet coefficient and the scale factor. Based on the calculated scale-dependent noise threshold, the soft threshold processing is performed on the wavelet coefficient sequence of each layer scale. The soft threshold processing function sets the wavelet coefficients with absolute values less than the noise threshold to zero, and shrinks the wavelet coefficients with absolute values greater than the noise threshold. The wavelet coefficient sequences of all scales after soft threshold processing are reconstructed using the inverse wavelet transform algorithm to obtain the denoised signal segment. Although the noise of the denoised signal segment is suppressed, there may still be baseline drift.

[0028] In a specific implementation, the baseline correction of the noise-reduced signal segment is a key step to eliminate background interference. The baseline correction algorithm aims to estimate and subtract the slowly varying background component superimposed on the useful signal. An iterative polynomial fitting algorithm can be used, which first performs a global polynomial fitting on the noise-reduced signal segment to obtain an initial baseline estimate; then subtracts the initial baseline from the original noise-reduced signal and identifies the peak regions in the remaining signal that are significantly higher in intensity than the noise level; after excluding the data points in these peak regions, the remaining data points are again fitted with a polynomial to obtain a more accurate baseline estimate; the above process is iteratively performed until the baseline estimate converges. It can be understood that the baseline correction effectively separates the characteristic peak signal of the target element from the background interference component caused by continuous radiation, scattered light, etc., significantly improving the signal-to-background ratio of the characteristic peak.

[0029] In a specific implementation, the processed signal segments are spliced into a continuous enhanced signal sequence by a signal reconstruction algorithm. Since each signal segment is independently processed by noise reduction and baseline correction, direct splicing may cause discontinuity in amplitude or phase at the connection of the signal segments. The signal reconstruction algorithm uses the overlap-add method to ensure the smoothness of splicing, and sets an overlap region for adjacent signal segments. In the overlap region, the signal values belonging to two signal segments are weighted and averaged, and the weight function is usually selected as a window function such as the Hanning window. In the center of the overlap region, the weight of the current signal segment is the largest, and as it moves towards the edge of the region, the weight gradually decreases, and the weight of the adjacent signal segment increases accordingly. Finally, all signal segments and their processed overlap regions are smoothly connected to form a complete and consistent enhanced signal sequence, providing high-quality input for the subsequent element feature decoupling network. The signal-to-noise ratio and baseline stability of the enhanced signal sequence are significantly improved compared to the original spectral signal stream.

[0030] Referring to Figure 3 The figure shows the difference between the original spectral signal and the enhanced signal in metal element detection: the horizontal axis is the wavelength, and the vertical axis is the signal intensity. The light gray curve is the original spectral signal, which shows obvious random fluctuations and noise interference, which is the initial data collected by the detection probe; the black curve is the enhanced signal, after processing such as multi-scale filtering and baseline correction, the noise is greatly suppressed, and the signal profile is smoother and the characteristic peak is clearer. The figure corresponds to the technical steps of signal quality evaluation and activation of the enhancement mechanism for the original spectral signal stream: the effect of multi-scale filtering and baseline correction is intuitively reflected by comparison, which shows that the enhancement mechanism can preserve the effective spectral features while improving the signal quality, providing a more reliable data basis for subsequent element feature decoupling. This processing avoids redundant calculation of high-quality signals and solves the distortion problem of low signal-to-noise ratio signals, which is a key preprocessing step to ensure the accuracy of element detection.

[0031] Example 3: Referring toFigure 4 In a specific implementation, constructing a feature template library containing standard spectra of multiple metal elements is the basis for the operation of the element feature decoupling network, and the establishment of the feature template library needs to be completed by measuring a series of high-purity metal or alloy standard samples under standardized excitation and acquisition conditions. For each metal element to be detected, such as iron, nickel, chromium, molybdenum, etc., a calibrated spectrometer is used to collect its characteristic spectrum in a specific wavelength range. The collected characteristic spectrum needs to be preprocessed, including wavelength calibration to eliminate instrument drift, intensity normalization to eliminate the influence of absolute intensity fluctuations, and smoothing and denoising to improve signal quality. The preprocessed single-element characteristic spectrum is stored as an atomic vector in the feature template library, and the feature template library can be mathematically represented as a matrix, with each column of the matrix corresponding to a feature template vector of a specific element. The completeness of the feature template library directly affects the accuracy of subsequent decoupling, so the feature template library should cover as many elements as possible that may exist in the metal materials being detected. After the construction of the feature template library is completed, it is usually stored in the form of a database file for the element feature decoupling network to call.

[0032] In a specific implementation, the above sparse coding problem is numerically solved by the orthogonal matching pursuit algorithm to extract the weight coefficients corresponding to each feature template. The orthogonal matching pursuit algorithm is a greedy iterative algorithm. When the algorithm is initialized, the current residual vector is set to the enhanced signal sequence vector, the active atom set is set to an empty set, and the weight coefficient vector is initialized to a zero vector. The iteration process starts. In each iteration, the inner product of the current residual vector and each atomic vector (i.e., each column) in the feature template library matrix is calculated. The absolute value of the inner product reflects the similarity between the current residual and the atomic vector. The atomic vector with the largest absolute value of the inner product is selected as the active atom selected in this iteration, and the index of the selected active atom is added to the active atom set. Then, based on all atomic vectors contained in the current active atom set, a linear least squares problem is solved by least squares method, i.e., finding a set of weight coefficients such that the L2 norm error between the linear combination of the active atom set and the original enhanced signal sequence vector is minimized. Update the weight coefficient vector, assign the newly calculated weight coefficients to the positions corresponding to the active atoms, and keep the weight coefficients corresponding to the non-active atoms as zero. Subsequently, according to the current active atom set and the updated weight coefficient vector, a new residual vector is calculated, which is equal to the original enhanced signal sequence vector minus the vector obtained by linearly combining the current active atom set according to its weight coefficients. Check whether the L2 norm of the new residual vector is lower than the preset tolerance threshold, or whether the number of iterations has reached the preset maximum number of iterations. If either condition is met, terminate the iteration; otherwise, continue with the next iteration. After the iteration is terminated, the final weight coefficient vector is output.

[0033] In a specific implementation, the effective element features are determined and the feature component set is generated according to the output of the orthogonal matching pursuit algorithm. An activation threshold is set, which is a positive number, to distinguish the small weight coefficients caused by noise from the significant weight coefficients corresponding to the real element features. Each component in the weight coefficient vector is traversed, and the feature template corresponding to the component with an absolute weight coefficient value greater than the activation threshold is determined as an effective element feature. Each effective element feature corresponds to a metal element identified in the enhanced signal sequence. The size of the weight coefficient value corresponding to the effective element feature reflects the contribution intensity of the element feature spectrum in the mixed signal to some extent. Finally, all the effective element features and their corresponding weight coefficients and feature template information are combined to form the feature component set corresponding to the enhanced signal sequence. The feature component set is the output of the element feature decoupling network and provides decoupled spectral feature data directly related to a specific element for subsequent quantitative calculation.

[0034] It can be understood that the orthogonal matching pursuit algorithm can effectively approximate the original signal from the over-complete feature template library by iteratively selecting the atom most relevant to the current residual. The advantage of the sparse coding framework combined with the orthogonal matching pursuit algorithm is that it can handle the case where there is overlap or interference between the feature spectra, and force the decoupling result to focus on a few significant elements through sparse constraint, which is consistent with the physical fact that the actual spectrum usually contains only a limited number of major and trace elements. The performance of the element feature decoupling network depends on the quality of the feature template library, the selection of the regularization parameter, and the setting of the activation threshold.

[0035] In a specific implementation, the calculation of the quantification index of each metal element starts with processing the feature component set output by the element feature decoupling network, which contains the feature spectral components corresponding to each metal element identified. The integral intensity calculation is performed on each feature component, which is a numerical integral of the spectral intensity value within a specific wavelength range of the element feature spectrum, for example, using the trapezoidal rule or Simpson's rule to calculate the integral area to obtain the original intensity value representing the strength of the element signal. In order to eliminate the system error caused by instrument power fluctuation, environmental temperature change or sample surface state difference, the original intensity value is corrected by the internal standard method, which needs to select a metal element with stable content in the material as the internal standard element, and calculate the ratio of the original intensity value of the element to be measured to the original intensity value of the internal standard element, i.e. the intensity ratio, which is used as the corrected intensity index. Then the corrected intensity index is converted into element concentration value through the standard curve, which is obtained by measuring a series of standard samples with known concentration, drawing the relationship curve between intensity ratio and concentration and regressing fitting, and the fitting function can be linear function or quadratic function. Since multiple measurements may be performed at the same or adjacent detection points, a set of multiple concentration values for the same element will be obtained, which need to be statistically consistent, for example, using Grubbs test method to identify and eliminate outliers.

[0036] In a specific implementation, the calculated quantification index of each metal element is matched with the pre-stored reference range in the material standard database, which stores the upper and lower limits of the content of each element in various standard metal materials. The matching process compares the quantification index concentration value of each element with the corresponding reference range upper and lower limits, and the metal element whose quantification index concentration value falls outside the reference range is identified as an abnormal element. The identified abnormal element needs to start a deep verification process to confirm the reliability of its detection result, which first adjusts the energy parameters of the excitation source, which can be a laser or an X-ray tube, including laser energy, pulse width, repetition frequency or tube voltage, tube current, uses a stepwise energy scanning mode to excite at multiple energy levels from low to high or vice versa, and reacquires the spectral signal at each energy level. In a specific implementation, the deep verification process also involves changing the collection angle of the detection probe, which is accurately controlled by the mechanism to change the incident angle and receiving angle of the detection probe relative to the sample surface, so as to collect the spectral signal of the same detection point from different geometric angles to obtain a multi-angle spectral signal set.

[0037] In practice, the multi-angle spectral signals acquired during the depth verification process are analyzed independently. For each angle, signal quality assessment, signal enhancement, elemental feature decoupling, and quantitative calculation are performed independently, resulting in a set of verification concentration values ​​for the same anomalous element. This set of verification concentration values ​​may contain multiple data points, each corresponding to a combination of excitation parameters and acquisition angles. To evaluate the consistency between the depth verification results and the initial detection results, a deviation coefficient needs to be calculated. The formula for calculating the deviation coefficient is as follows:

[0038] Where: symbol Represents the deviation coefficient, symbol Represents the number of data points in the set of validation concentration values, with the sign... The first value in the set of verified concentration values Concentration values ​​for each data point, symbol This represents the concentration value of the anomalous element obtained from the initial detection. The calculated deviation coefficient. Compared with a preset allowable deviation range (e.g., 10%), if the deviation coefficient... If the deviation coefficient is less than or equal to the upper limit of the allowable range, then the initial identification of the anomalous element is confirmed to be valid; if the deviation coefficient is... If the result exceeds the allowable range, it indicates that there may be random errors in the initial detection, requiring further verification or the result should be discarded as appropriate. The deep verification process significantly improves the reliability of abnormal element detection results through multi-parameter and multi-angle cross-validation.

[0039] Understandably, the calculation process of the quantitative indicators effectively eliminates systematic errors through the internal standard method and standard curve, while the statistical consistency test ensures the reliability of the data. The in-depth validation process effectively identifies measurement deviations caused by accidental factors by retesting under different detection conditions. Table 1 shows a set of validation concentration values ​​for a hypothetical multi-angle spectral signal, illustrating the calculation process of the deviation coefficient.

[0040] Table 1: Concentration Values ​​for Verification of Multi-Angle Spectral Signals

[0041] Assuming the initial detection yields the quantitative index concentration value of this anomalous element The deviation coefficient is calculated based on the set of verification concentration values ​​in the table above. First, calculate the standard deviation of the validation concentration value relative to the initial concentration. Then, divide by the initial concentration and multiply by 100% to obtain the deviation coefficient in percentage form. This calculation allows for an objective assessment of the consistency between the validation results and the initial results.

[0042] Optionally, the statistical consistency test can use the Dixon test method instead of the Grubbs test method. The Dixon test method determines outliers by calculating the range ratio, which is suitable for scenarios with small sample sizes. Optionally, in the depth verification process, the stepwise energy scanning mode can use a high-to-low order or a random order to eliminate the time sequence influence of possible energy-dependent effects. In some embodiments, the fitting of the standard curve can use weighted least squares to consider the heteroscedasticity of measurement errors at different concentration points. In some embodiments, the collection angle of the multi-angle spectral signal can be selected to be more than two angles, for example, adding a 60-degree angle collection to obtain more rich angle-dependent information. It can be understood that the selection of the internal standard element is crucial. The ideal internal standard element should be uniformly distributed in the material, have stable content, and not be significantly affected by heat treatment or processing technology.

[0043] Referring to Figure 5 The figure presents the bias coefficients of 5 groups of depth verification in a bar chart: the horizontal axis is the verification number, the vertical axis is the bias coefficient, and the black dotted line represents the allowed bias coefficient. The figure corresponds to the technical link of the depth verification process: after re-detecting by adjusting the excitation energy, collection angle, and other differential parameters, the bias coefficient of the verification concentration and the initial quantification index is calculated to confirm the effectiveness of the abnormal element detection result. All data in the figure are within the allowed range, directly reflecting the reliability of the depth verification process, which not only eliminates systematic errors such as matrix effect and instrument fluctuation, but also improves the confidence of abnormal element determination. It is a key link to avoid misjudgment / omission and provides a guarantee for the accuracy of the final element content distribution map.

[0044] In a specific implementation, the generation of the final element content distribution map starts with the accurate association of the element concentration data validated by the deep verification process with their spatial positions on the material surface. The three-dimensional spatial coordinates of each detection point come from the position information recorded during the dynamic detection path planning, while the element concentration data of the point come from the results after quantitative calculation and deep verification. For each metal element that needs to be analyzed for its distribution, such as chromium or nickel, the element concentration values of all detection points are combined with their three-dimensional coordinates to form a set of discrete spatial data points. The data point set contains X, Y, Z coordinates and concentration value C. In order to generate a continuous element concentration distribution surface from the discrete data point set to visually display the distribution trend of the element on the entire material surface, a spatial interpolation algorithm is needed. The Kriging interpolation algorithm is a best linear unbiased estimation method based on geostatistics, which is used to complete this task. The Kriging interpolation algorithm first analyzes the spatial autocorrelation of the concentration values in the spatial data point set. By calculating the experimental variogram, the characteristics of the concentration value change with the increase of the spatial distance are quantified. The experimental variogram describes the relationship between the concentration difference and the distance between the points. Based on the theoretical variogram model obtained by fitting, the Kriging interpolation algorithm calculates the optimal weight coefficient for each grid node position to be interpolated, so that the variance of the interpolation estimate is minimized and the estimate is unbiased. Finally, a continuous element concentration distribution surface that covers the entire material surface, smooth and consistent with the spatial statistical law is constructed.

[0045] In a specific implementation, the element concentration distribution surface generated by the Kriging interpolation algorithm is subjected to contour extraction. Contour lines are lines connecting points on the surface with equal concentration values. The contour extraction algorithm traverses the entire interpolated concentration grid to find grid edges with concentration values equal to a series of preset contour levels, and accurately determines the position of the contour line through linear interpolation. The extracted contour lines are superimposed on the two-dimensional projection or three-dimensional model of the material surface. Areas with dense contour lines indicate that the element concentration gradient is large at that location, which may correspond to the segregation band, diffusion interface or microcomponent inhomogeneous region of the element. Areas with sparse contour lines indicate that the concentration changes smoothly and the element distribution is relatively uniform. Integrating the concentration distribution surfaces of multiple metal elements forms a multi-level element content distribution map. In software implementation, the layer management method is usually used, with each layer corresponding to the distribution information of a metal element. Users can display or hide the distribution map of a specific element by checking or unchecking, and can also adjust the opacity of different layers to observe the superposition relationship between the distributions of multiple elements. The final element content distribution map is output in the format of an image file or raster data containing geographic reference information.

[0046] In a specific implementation, the process of updating the material standard database according to the final element content distribution map is a data-driven self-learning link. Statistical characteristic parameters that can represent the element distribution characteristics are extracted from the final element content distribution map. These statistical characteristic parameters include but are not limited to: the average concentration value of each element in the batch material, the standard deviation of the concentration value, the maximum value of the concentration value, the minimum value of the concentration value, the skewness coefficient of the concentration distribution, the kurtosis coefficient of the concentration distribution, and the spatial uniformity index calculated based on the spatial interpolation grid. The spatial uniformity index can be calculated by calculating the coefficient of variation of the concentration grid value or using the information entropy theory to evaluate the degree of disorder of the distribution. Similarity matching calculation is performed between these extracted statistical characteristic parameters and the statistical characteristic parameters of the historical detection data of the same type of material stored in the material standard database. Similarity matching can use distance measurement methods in multi-dimensional space, such as calculating the Euclidean distance between the current statistical characteristic parameter vector and each vector in the set of historical statistical characteristic parameter vectors, and taking the minimum distance as the similarity index. The similarity index calculated is compared with a preset update threshold. If the similarity index is lower than the update threshold, it means that the element distribution characteristics of the current detection material are significantly different from the historical data, and at this time the adaptive updating mechanism of the database is started.

[0047] In a specific implementation, the database adaptive updating mechanism uses a sliding window algorithm to manage the historical record data in the material standard database. The sliding window algorithm sets a fixed window size, for example, only the data records of the last 100 qualified detection batches and their corresponding statistical characteristic parameters are retained. When new detection batch data needs to be added and the update condition is triggered, the new data record is added to the database window, and at the same time, if the window is full, the historical data record that has been in the window the longest is removed from the window, so as to keep the size of the database window constant. After updating the data records in the window, the upper and lower limits of the reference content range of each metal element in this type of material need to be recalculated based on all the current data records in the window. The recalculation is usually based on statistical methods, for example, the new upper limit of the reference range can take the average of the average concentration values of all elements in the window plus three times the standard deviation, and the new lower limit of the reference range takes the average minus three times the standard deviation. The reference range update formula can be represented as:

[0048] wherein the symbol represents the newly calculated lower limit of the reference range, the symbol represents the newly calculated upper limit of the reference range, the symbol represents the arithmetic mean of the average concentration values of the element in the sliding window, and the symbol represents the standard deviation of the average concentration values of the element in the sliding window. is a constant factor, usually taking the value of 2 or 3, used to control the width of the range. In this way, the material standard database can dynamically adapt to the improvement of production process, the introduction of new material grades or the change of detection standards, so that the benchmark range always reflects the normal level under the current production status.

[0049] It can be understood that the Kriging interpolation algorithm can make full use of the correlation of spatial data, and the generated spatial distribution surface is not only smooth, but also can reflect the spatial structure characteristics of the data. The contour extraction converts the numerical information into intuitive graphical expression, which is convenient for quickly identifying abnormal areas. The multi-level element content distribution map provides comprehensive spatial distribution information of the components. The sliding window updating mechanism ensures the timeliness and adaptability of the material standard database, and avoids misjudgment caused by using outdated standards. It can be understood that the complete closed loop from the distribution map generation to the database update enables the entire detection system to have the ability of continuous learning and self-optimization. Optionally, the contour extraction can use a more complex contour filling algorithm to fill the area between adjacent contours with different colors or patterns to enhance the visual contrast effect. Optionally, the similarity matching can also use other measurement methods such as cosine similarity or Mahalanobis distance to adapt to different data distribution characteristics. In some embodiments, the statistical characteristic parameters can also include quantile information such as median and quartile range to enhance the robustness of the distribution shape description. In some embodiments, for very important elements, a smaller sliding window or more frequent update trigger condition can be set for them separately to achieve more refined monitoring. Optionally, the material standard database can also include a historical database to record the evolution of the material quality over time, which can provide more comprehensive information for the production process analysis and quality control. It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for detecting metallic elements in metallic materials, characterized in that, The method includes: A dynamic detection path is established for metallic materials, wherein the spatial distribution of detection points is adjusted in real time according to the material morphology. The detection probe is driven to scan according to the dynamic detection path, and the raw spectral signal stream returned by the detection probe is acquired simultaneously. The original spectral signal stream is evaluated for signal quality, and the signal enhancement mechanism is activated based on the evaluation results to generate an enhanced signal sequence. The enhanced signal sequence is input into the element feature decoupling network to decouple the feature components corresponding to different metal elements. The quantitative index of each metal element is calculated based on the decoupled characteristic components. The quantitative indicators are matched with the benchmark range in the material standard database to identify abnormal elements with deviations. A deep verification process is initiated for abnormal elements, and the deep verification process uses differentiated excitation parameters for re-detection; By integrating the initial detection results with the output of the deep verification process, a final elemental content distribution map is generated. The historical records in the material standards database are updated based on the final elemental content distribution map.

2. The method for detecting metallic elements in metallic materials according to claim 1, characterized in that, The establishment of a dynamic detection path for metallic materials includes: Obtain three-dimensional surface topology data of metallic materials and extract curvature features from the three-dimensional surface topology data; High-probability defect areas are identified based on curvature characteristics, and dense detection arrays are arranged in these areas. An adaptive mesh generation algorithm is used to generate sparse detection points in non-high probability defect areas; The dense detection array and the sparse detection array are combined through path optimization to form a complete dynamic detection path.

3. The method for detecting metallic elements in metallic materials according to claim 2, characterized in that, The signal quality assessment of the original spectral signal stream includes: The original spectral signal stream is segmented, and the signal-to-noise ratio of each segment is calculated. When the signal-to-noise ratio falls below a preset threshold, the signal enhancement mechanism is triggered. The signal enhancement mechanism uses a multi-scale filtering algorithm to reduce noise in the signal; Baseline correction is performed on the denoised signal to eliminate background interference components; The processed signal segments are spliced ​​together into an enhanced signal sequence using a signal reconstruction algorithm.

4. The method for detecting metallic elements in metallic materials according to claim 1, characterized in that, The operation of inputting the enhanced signal sequence into the element feature decoupling network includes: Construct a feature template library containing standard spectra of multiple metal elements; The enhanced signal sequence is decomposed into a linear combination of feature templates using a sparse coding algorithm; The weight coefficients corresponding to each feature template are extracted using the orthogonal matching pursuit algorithm; Feature templates with weight coefficients greater than the activation threshold are identified as valid element features; Generate a set of feature components based on the characteristics of valid elements.

5. The method for detecting metallic elements in metallic materials according to claim 4, characterized in that, The quantitative indicators for calculating each metallic element include: Integral intensity calculation is performed on each feature component to obtain the original intensity value; The original strength value was corrected using the internal standard method to eliminate the influence of instrument fluctuations; The corrected intensity values ​​are converted into elemental concentration values ​​using a standard curve. Perform a statistical consistency test on the concentration values ​​of the same element from multiple detections; The arithmetic mean of the concentration values ​​that pass the consistency test is taken as the final quantitative indicator.

6. The method for detecting metallic elements in metallic materials according to claim 1, characterized in that, The deep verification process includes: Adjust the energy parameters of the excitation source and use a stepped energy scanning mode for excitation; By changing the acquisition angle of the detection probe, multi-angle spectral signals can be obtained; Independent analysis of multi-angle spectral signals yields a set of verification concentration values; Calculate the deviation coefficient between the set of verified concentration values ​​and the initial quantification index. When the deviation coefficient is within the allowable range, the validity of the abnormal element detection results is confirmed.

7. The method for detecting metallic elements in metallic materials according to claim 6, characterized in that, The generation of the final elemental content distribution map includes: Map the confirmed valid abnormal element concentration values ​​back to the corresponding dynamic detection path coordinates; The element concentration distribution surface of the entire material surface is constructed using the Kriging interpolation algorithm; Contour lines are extracted from the element concentration distribution surface to identify regions of concentration gradient change. The distribution surfaces of various metal elements are integrated into a multi-level element content distribution map.

8. The method for detecting metallic elements in metallic materials according to claim 1, characterized in that, The process of updating the material standard database based on the final elemental content distribution map includes: Statistical characteristic parameters were extracted from the final elemental content distribution map, and the statistical characteristic parameters were matched with historical data in the material standard database. When the similarity is lower than the update threshold, the database adaptive update mechanism is activated, and the sliding window algorithm is used to retain the latest batch of detection data. Recalculate the upper and lower limits of the reference range in the material standard database.

9. A method for detecting metallic elements in metallic materials according to claim 3, characterized in that, The signal enhancement mechanism employs a multi-scale filtering algorithm for signal noise reduction, including: Wavelet transform is applied to the original spectral signal stream to decompose the signal into a sequence of wavelet coefficients at multiple scales; For each scale of the wavelet coefficient sequence, a scale-dependent noise threshold is calculated, and the wavelet coefficients are subjected to soft thresholding to eliminate noise components. The wavelet coefficient sequence after thresholding is reconstructed into a denoised signal segment through inverse wavelet transform.

10. A method for detecting metallic elements in metallic materials according to claim 4, characterized in that, The step of extracting the weight coefficients corresponding to each feature template using the orthogonal matching pursuit algorithm includes: Initialize the residual to the enhanced signal sequence, initialize the set of activation atoms to an empty set, and initialize the weight coefficient vector to a zero vector; The following steps are performed iteratively until the norm of the residual is below a preset threshold or the maximum number of iterations is reached: Calculate the inner product between the current residual and each feature template in the feature template library, and select the feature template with the largest absolute value of the inner product as the activation atom of the current iteration; Add the currently active atom to the set of active atoms; The least squares method is used to solve the linear approximation of the enhanced signal sequence by the activation atom set, and the weight coefficients corresponding to the current activation atom set are calculated. The updated residual is the difference between the enhanced signal sequence and the weighted sum of the activated atom set; The final weight coefficient vector is output as the weight coefficients corresponding to the feature template.

Images