Online thickness measurement method of converter lining based on laser 3D scanning and edge computing

Abstract

The present invention discloses an online thickness measurement method for a converter lining based on laser three-dimensional scanning and edge computing, and relates to the technical field of converter lining thickness detection. The method comprises the following steps: collecting multi-source data; matching a control mode based on the working state of the converter, and based on the control mode, performing spatiotemporal registration and feature fusion of the multi-source data to generate a wear feature matrix, including generating a control coefficient; at the same time, obtaining a lining vibration signal and infrared thermal imaging temperature of a known wear form, extracting a first characteristic parameter, including a temperature characteristic and a vibration characteristic, and mapping and associating them based on the control coefficient to establish a corresponding correlation relationship model; based on the correlation relationship, obtaining a characteristic parameter classification range for each type of wear, and calculating a corresponding wear value; based on the characteristic parameter classification range, obtaining laser point cloud data, wear form, and wear value under abnormal wear, and importing them as input information into a pre-set residual thickness prediction model to output a residual thickness value; the present invention improves the monitoring accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of converter lining thickness detection, and specifically to an online converter lining thickness measurement method based on laser three-dimensional scanning and edge computing. Background Art

[0002] The thickness of the converter lining has a very important impact on the safety of steelmaking production and the confirmation of the converter life. In steelmaking production, the erosion of molten steel and the accumulation of slag will cause the shape and thickness of the lining to change. The lining will become thinner and thinner. In industrial production, it is necessary to ensure that the thickness of the lining is not less than the safe thickness. Therefore, it is necessary to monitor the thickness of the lining in real time to extend the service life of the converter and reduce steelmaking costs.

[0003] In traditional converter lining thickness detection technology, the detection must be carried out after the converter is shut down, and online real-time monitoring cannot be achieved. As a continuous production equipment, the converter shutdown detection will lead to production capacity loss and increase the production cycle cost. With the development of technology, neural networks have also been introduced to predict residual thickness. During the prediction process, only single-dimensional data (such as thickness or temperature) can be obtained, and it is impossible to correlate with the characteristics of the converter lining wear process (such as vibration, geometric wear), resulting in incomplete analysis of the wear mechanism. Traditional detection data needs to be transmitted offline to the central server for processing, with a delay of up to several hours, which cannot meet the real-time control needs of the smelting process. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides an online thickness measurement method for converter lining based on laser three-dimensional scanning and edge computing. Based on the working state of the converter, the first deviation, the second deviation and the third deviation are obtained, and the working environment of the converter lining is used to obtain the control coefficient to switch different control modes. Based on the control mode, the corresponding space-time coordinate system is established, and the timestamps of multi-source data are synchronized and spatially aligned to obtain laser point cloud data, wear forms and wear values ​​under abnormal wear. The data are imported as input information into a pre-set residual thickness prediction model to output the residual thickness value, thereby solving the problems raised in the background technology.

[0006] (2) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] In a first aspect, the present application provides a method for online thickness measurement of a converter lining based on laser three-dimensional scanning and edge computing, the method comprising:

[0009] Collect multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals;

[0010] Based on the converter's operating status, the control mode is matched. Based on the control mode, multi-source data is temporally and spatially aligned and feature-fused to generate a wear feature matrix, including control coefficients. Simultaneously, the lining vibration signal and infrared thermal imaging temperature of known wear patterns are acquired, and the first characteristic parameters, including temperature and vibration characteristics, are extracted. These are mapped and associated based on the control coefficients to establish a corresponding association relationship model. Based on the association relationship, the classification range of characteristic parameters for each type of wear is obtained.

[0011] Based on the classification range of characteristic parameters, the laser point cloud data, wear form and wear value under abnormal wear are obtained and imported as input information into the pre-set residual thickness prediction model to output the residual thickness value.

[0012] Furthermore, a wear feature matrix is ​​generated, including:

[0013] Based on the converter's operating status, operating factors are extracted, including tilting height, tilting angle, and angle dwell time, to form a first operating vector set. The first operating vector set is compared with a corresponding preset baseline operating vector set to obtain different deviations. Different weights are assigned to the different deviations, and the operating environment of the converter lining is retrieved to obtain an environmental compensation coefficient. The control coefficient is obtained by combining the deviation, weight, and environmental compensation coefficient to switch to the corresponding control mode.

[0014] Under different control modes, a corresponding space-time coordinate system is established, the timestamps of multi-source data are synchronized and spatially aligned, and the second parameter features are extracted, including at least geometric features, temperature features, and vibration features. The extracted parameter features are fused to generate a wear feature matrix.

[0015] Furthermore, the control mode includes a regular monitoring mode, an enhanced monitoring mode and an emergency alarm mode.

[0016] Furthermore, based on the control coefficient mapping association, a corresponding association relationship model is established; based on the association relationship, the classification range of characteristic parameters for each type of wear is obtained, including:

[0017] Data preparation: Based on the known wear form and the first characteristic parameter, a multidimensional feature vector is constructed, including the first characteristic parameter, the wear form, and a timestamp. The temperature feature includes at least the temperature gradient and hot spot entropy, and the vibration feature includes at least the peak value, the main frequency offset, and the wavelet packet energy entropy.

[0018] Vector construction: Obtain the average and fluctuation values ​​of the parameters corresponding to the vibration or temperature characteristics. For any time period Ta, mark the wear form of this time period Ta as Sa, and the average and fluctuation values ​​as p_a and p_b respectively. Then, according to different combinations, obtain the first judgment vector [p_a+p_b, Sa] and the second judgment vector [p_a-p_b, Sa];

[0019] Correlation analysis: retrieve the control index, perform correlation analysis, and obtain the corresponding correlation relationship model; wherein, correlation analysis includes primary correlation and secondary correlation.

[0020] Furthermore, a primary association is performed: based on the first judgment vector, the control index is retrieved, a first time tag is created for the control index according to the time series, and a correlation relationship model between the first judgment vector and the control index is formed;

[0021] Secondary association: Based on the second judgment vector, the control index is retrieved, and a second time tag is created for the second judgment vector according to the time series to form an association relationship model between the second judgment vector and the control index.

[0022] Furthermore, based on the correlation, the classification range of characteristic parameters for each type of wear is obtained, including:

[0023] Based on the association relationship, the matching values ​​between the first judgment vector and the second judgment vector and the control index are extracted, and the first matching value and the second matching value are obtained accordingly. If the first matching value or the second matching value exceeds a preset matching threshold, the data set corresponding to the matching value is marked as a high matching data set;

[0024] Clustering operations are performed on high-matching data sets to generate wear type classification clusters. The rectangular indicators of each feature point are preset, and the matrix indicators include monitoring boundaries, boundary intervals, and boundary windows. Abnormal feature points that exceed the monitoring boundaries are screened out. The rectangular lengths of the screened abnormal feature points fall into the corresponding boundary intervals for classification, and several equidistant boundary intervals are labeled and weighted, and the labels and weights are kept synchronized. Otherwise, normal type classification clusters are marked.

[0025] Furthermore, the step of obtaining the classification range of characteristic parameters for each type of wear includes: judging whether the data parameters of the real-time scan fall into the classification range corresponding to the normal operating condition fluctuation. If they fall into the normal type classification cluster, it is determined to be normal deformation; otherwise, it is marked as abnormal wear and the wear value is obtained.

[0026] Furthermore, the wear value includes:

[0027] The frequency of abnormal feature points of each feature parameter is extracted; the feature vectors corresponding to the classified feature points are weighted and concatenated to form a feature matrix, and each row and column of the feature matrix corresponds to a fused feature value; the product of the fused feature value and the frequency of abnormal feature points is marked as the wear value.

[0028] Furthermore, the residual thickness prediction model has a built-in SVM classifier and neural network architecture, including:

[0029] Data preparation and processing: Collect laser point cloud data under known abnormal wear conditions, extract the second feature information, and combine it with the corresponding wear form and wear value to obtain the target data set; the second feature parameters include at least geometric features;

[0030] Model training and evaluation: The target dataset is input into the SVM classifier, which is trained to output the probability distribution of each wear form. The classification probability output by the SVM is marked as a classification feature, combined with the second feature information, and input into the neural network. The attention mechanism layer in the neural network automatically learns the influence weights of different wear forms on the residual thickness value. The classification features are weighted and fused based on the influence weights, and the residual thickness prediction value is output through the regression layer.

[0031] Model prediction: Use the trained model to predict the laser point cloud data under new abnormal wear and output the residual thickness value under the corresponding wear form.

[0032] In a second aspect, the present application provides an online converter lining thickness measurement system based on laser three-dimensional scanning and edge computing, the system comprising:

[0033] Acquisition module, which collects multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals;

[0034] The processing module matches the control mode based on the converter's operating status. Based on the control mode, it performs spatiotemporal registration and feature fusion of multi-source data to generate a wear feature matrix, including control coefficients. Simultaneously, it obtains the lining vibration signal and infrared thermal imaging temperature of known wear patterns, extracts the first characteristic parameters, including temperature and vibration characteristics, and maps and associates them based on the control coefficients to establish a corresponding association relationship model. Based on the association relationship, it obtains the classification range of the characteristic parameters for each type of wear and calculates the corresponding wear value.

[0035] The thickness measurement module obtains laser point cloud data, wear form and wear value under abnormal wear based on the classification range of characteristic parameters, and imports them as input information into the pre-set residual thickness prediction model to output the residual thickness value.

[0036] (3) Beneficial effects

[0037] The present invention provides an online converter lining thickness measurement method based on laser three-dimensional scanning and edge computing, which has the following beneficial effects:

[0038] 1. The present invention obtains the first deviation, second deviation, and third deviation based on the working state of the converter, avoiding the problem that a single-dimensional deviation may not be sufficient to cause risks, but the superposition of multi-dimensional deviations will significantly aggravate lining erosion (such as height abnormalities, angle abnormalities, and long-term dwelling). In combination with the working conditions of the converter lining, the control coefficient is obtained to switch between different control modes. Based on different control modes, different coordinate systems are designed, and multi-source data is spatiotemporally aligned and feature-fused to improve monitoring accuracy. The purpose is to provide an explainable, traceable, and predictable digital foundation for equipment lifecycle management.

[0039] 2. The present invention quantifies temperature and vibration characteristics into average values ​​and fluctuation values ​​through control coefficient mapping association, reducing the complexity of analysis. At the same time, the first judgment vector and the second judgment vector are designed to amplify and reduce the comprehensive value after the characteristic parameter fluctuation, comprehensively summarize the extreme and stable operating conditions of the converter, reduce the misjudgment rate of a single indicator, and improve the robustness of the model;

[0040] 3. The present invention uses a residual thickness prediction model, and the residual thickness prediction model has a built-in SVM classifier and neural network architecture. After fusing classification and regression labels, the target data set contains both wear form semantics and residual thickness values, avoiding the loss of feature semantics caused by a single label; to a certain extent, it improves the generalization ability of the model to achieve high-precision predictions, provide a decision-making basis for industrial maintenance, shorten equipment downtime detection time, and improve production line efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The figure is a schematic diagram showing the steps of a method for online thickness measurement of a converter lining according to an exemplary embodiment. DETAILED DESCRIPTION

[0042] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] Example 1:

[0044] The embodiment of the present invention provides an online thickness measurement method for converter lining based on laser three-dimensional scanning and edge computing; Figure 1 is a schematic diagram showing the steps of a method for online thickness measurement of a converter lining according to an exemplary embodiment; Figure 1 , the method comprises the following steps;

[0045] Collect multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals;

[0046] Among them, laser point cloud data is used to reconstruct the three-dimensional geometric morphology of the furnace lining, infrared thermal imaging temperature distribution is used to reflect the surface temperature field characteristics of the furnace lining, and the furnace lining vibration signal is used to monitor the impact of mechanical shock on the furnace lining;

[0047] Specifically, the collection steps include:

[0048] Laser point cloud data collection: When the converter enters the inspection station, laser scanning is triggered. A pulsed 1550nm laser radar array scans the furnace lining surface, with each radar group emitting laser beams alternately at a 20Hz frequency. (For example, the laser scanning configuration uses eight laser radars arranged in a circular array, covering the entire circumference of the furnace lining. Each radar unit is equipped with a Class 1M laser with a transmission power of 50mW, a 5-megapixel high-speed CMOS receiver, and a high-temperature filter.) A dynamic voxel filtering algorithm is used to process the raw point cloud, eliminating interference from furnace mouth splashes and generating a millimeter-level accurate 3D point cloud model for detecting geometric deformation defects such as lining erosion and spalling.

[0049] By performing point cloud data registration, splicing, and surface reconstruction on the 3D laser point cloud data of the converter obtained at different angles, the subsequent detection of the converter lining thickness is facilitated, providing a guarantee for the safe production of converter steelmaking.

[0050] Infrared thermal imaging temperature acquisition: An infrared thermal imager installed above the furnace mouth synchronously captures the temperature field of the furnace lining surface; a non-uniformity correction algorithm is used to process the original thermal image to eliminate smoke interference; and thermal imaging data with a temperature resolution of 0.5°C is output for identifying abnormal areas such as thermal shock cracks and local overheating.

[0051] Acquisition of furnace lining vibration signals: Triaxial acceleration sensors are arranged in a 1.5m x 1.5m grid on the outer wall of the furnace shell. Vibration signals are transmitted through high-temperature resistant waveguide rods (operating temperature ≤ 150°C). Vibration waveforms are collected in real time at a 10kHz sampling rate to detect mechanical defects such as loose anchors and structural fatigue.

[0052] By configuring edge computing nodes and connecting them to edge controllers, a unified trigger signal is generated using the built-in IEEE 1588v2 precision clock protocol. This hardware-level synchronization ensures that the laser scanner, infrared thermal imager, and triaxial accelerometer all start data acquisition within the same time window (for example, a ±100μs time window). Time consistency is achieved for multi-source data through timestamp alignment services.

[0053] Based on the converter's operating status, the control mode is matched. Based on the control mode, multi-source data is temporally and spatially aligned and feature-fused to generate a wear feature matrix. The lining vibration signal and infrared thermal imaging temperature under known wear patterns are obtained, and the first characteristic parameters, including vibration characteristics and temperature characteristics, are extracted. These characteristics are mapped and associated to establish an association relationship. Based on the association relationship, the classification range of characteristic parameters for each type of wear is obtained. In calculating the classification range of characteristic parameters for each type of wear, the following steps are included: determining whether the data parameters of the real-time scan fall within the classification range corresponding to normal operating conditions. If they match, the deformation is determined to be normal; otherwise, it is marked as abnormal wear.

[0054] The steps for generating the wear characteristic matrix include:

[0055] Based on the working state of the converter, working factors are extracted, including the tilting height, tilting angle, and angle dwell time, and a first working vector is formed. The first working vector is compared with a corresponding preset reference working vector to obtain a first deviation, a second deviation, and a third deviation. The first deviation, the second deviation, and the third deviation are assigned a first weight, a second weight, and a third weight, respectively.

[0056] The first weight is associated with the tilting height, the second weight is associated with the tilting angle deviation, and the third weight is related to the angle dwell time. Under the constraint that the sum of these three weights is 1, the respective proportions are dynamically adjusted according to the real-time converter process, thereby achieving accurate monitoring and control of the converter working status.

[0057] The first weight is derived from the degree of influence of the tilting furnace height on the mechanical erosion of the furnace lining. Its significance lies in quantifying the contribution of the tilting furnace height to the mechanical wear of local areas of the furnace lining (such as the furnace body and furnace bottom) when the tilting furnace height deviates from the standard height. The tilting furnace height directly determines the flow path of molten steel and slag. When the first deviation value is larger, it indicates that the height is abnormal (too high or too low), which will cause the molten steel impact position to shift, exacerbating the erosion of specific areas, and the first weight will automatically increase. Conversely, the smaller the first deviation value, the first weight will decrease accordingly. The function is to dynamically adjust the proportion of the tilting furnace height in the comprehensive decision-making based on its credibility. During the calculation, the corresponding weight is obtained by searching a preset nonlinear mapping table (such as a piecewise function), realizing adaptive weighting of the tilting furnace height.

[0058] The second weight is derived from the ability of the tilting angle to influence the heat load distribution of the furnace lining. Its significance is to quantify the risk of lining temperature field imbalance caused by angle deviation. The tilting angle determines the contact area between the high-temperature areas (such as the melt pool and slag line) in the furnace and the furnace lining. An abnormal angle can cause local overheating (such as a temperature increase of more than 50°C in the slope area of ​​the furnace cap), accelerating the softening and spalling of the refractory material. During the calculation process, an exponential function is used to adjust the weight. For example, when the second deviation is low, the second weight is automatically reduced; when the second deviation is high, the second weight is correspondingly increased. The purpose is to highlight the indicative role of the tilting angle through a nonlinear weight distribution strategy, and enhance sensitivity to temperature field imbalance scenarios.

[0059] The third weight is derived from the continuous effect of angular dwell time on chemical erosion of the furnace lining. Its significance lies in distinguishing the impact of normal process dwell time and abnormally long dwell time on slag erosion. Excessively long angular dwell time can cause slag to remain in specific areas (such as the middle and lower part of the furnace body) for a long time, exacerbating the chemical penetration of CaO and FeO into magnesia carbon bricks. A third deviation threshold is preset. When the third deviation is greater than or equal to the third deviation threshold, the third weight takes a higher value, indicating that the abnormal dwell time contributes significantly to chemical erosion. When the third deviation is less than the third deviation threshold, the weight coefficient takes a lower value, indicating that short-term fluctuations generally do not cause significant erosion. The function is to filter out invalid interference in infrared data through graded weight setting. During calculation, the weight assignment after threshold comparison is realized through conditional judgment statements, ensuring that the infrared dimension plays an important role only in effective occlusion scenarios.

[0060] Retrieve the working environment of the converter lining and obtain the environmental compensation coefficient;

[0061] Environmental compensation coefficient: This term is derived from the impact of the operating environment on the converter's operating status. Its purpose is to eliminate interference from environmental noise. The real-time acquisition of the operating environment, including smelting temperature, scrap impurity content, and molten steel flow rate, uses infrared thermal imagers to monitor the furnace lining surface temperature field, extracting the average slag line temperature as the smelting temperature. Before entering the furnace, a spectrometer (XRF) is used to quickly detect the content of elements such as P, S, and Cu in the scrap to determine the scrap impurity content. A laser rangefinder monitors the molten steel flow in real time. The aforementioned sensor components or equipment are not shown in the figure and are installed adaptively within the operating environment.

[0062] Input the working environment into the pre-established compensation model, first make dimensionless treatment of smelting temperature, scrap steel impurity content and molten steel flow rate, weight and multiply them, and output the environmental compensation coefficient;

[0063] For example, at high smelting temperatures, the fluidity of molten steel increases, and the flow rate increases during tilting, which can easily lead to splashing. The risk of softening the furnace lining increases, and the tilting angle control requires more precise control. Therefore, the weight of the tilting height is reduced to slow the flow rate, the weight of the tilting angle is increased to accurately control the angle to avoid splashing, and the weight of the residence time is increased to extend the residence time to balance the flow of molten steel.

[0064] For example, when the impurity content of scrap steel is high, the oxidation heat of impurities (such as Si, P, and S) increases, and the smelting temperature fluctuates greatly. The viscosity of the slag increases, and the difficulty of slag separation increases. Therefore, the weight of the tilting furnace height is reduced to increase the height to enhance the slag separation space, the weight of the tilting furnace angle is reduced to accurately control the angle to prevent slag from mixing with molten steel, and the weight of the residence time is reduced to extend the residence time to ensure slag stratification.

[0065] For example, at high molten steel flow rates, the molten steel impacts the furnace lining and tapping port with great force, which can easily lead to equipment wear. Excessive flow rates may cause splashing or overoxidation of the molten steel. Therefore, the weight of the tilting furnace height can be reduced to reduce the height and impact potential energy, the weight of the tilting furnace angle can be increased to fine-tune the angle and slow down the flow rate, and the weight of the residence time can be increased to control the flow rate in sections to avoid continuous high speed.

[0066] The control coefficient is obtained by combining the deviation, weight and environmental compensation coefficient to switch the corresponding control mode;

[0067] Among them, the preset reference vectors are standard height, standard angle, and standard residence time. The first deviation is the absolute value of the difference between the tilting furnace height and the standard height, the second deviation is the absolute value of the difference between the tilting furnace angle and the standard angle, and the third deviation is the absolute value of the difference between the angle residence time and the standard residence time.

[0068] The steps to obtain the control mode include:

[0069] Multiply the first deviation, second deviation, and third deviation by the corresponding first to third weights respectively to obtain the weighted values ​​of each dimension, sum the weighted values, and then divide the sum by the environmental compensation coefficient to obtain the control index:

[0070] Wherein, kz represents the control index, pc1, pc2 and pc3 correspond to the first deviation, second deviation and third deviation respectively, β1, β2 and β3 correspond to the first weight, second weight and third weight respectively, and L represents the environmental compensation coefficient;

[0071] The significance of the control index lies in the fact that a single-dimensional deviation may not be sufficient to trigger a risk, but the accumulation of multiple-dimensional deviations (such as height anomalies, angle anomalies, and long dwell times) can significantly exacerbate lining erosion. The control index, through weighted summation, converts this coupling effect into a numerical increase, avoiding missed judgments where a single parameter is normal but the overall risk is out of control.

[0072] The control mode is switched based on the control index, and the control mode includes conventional monitoring mode, enhanced monitoring mode and emergency alarm mode. Specifically, the control index is compared with the preset risk range:

[0073] If the control index is less than the preset risk range, it means that the deviation of the converter operation is within the normal range and there is no significant risk. The control index at this time is marked as a and assigned a level 1 character. The combination of a and the level 1 character generates a level 1 risk state, which matches the conventional monitoring mode. For example, the laser scanning is maintained at a certain frequency (e.g., 10 Hz), the scanning trajectory is a standard spiral, and each scanning time is about 30 seconds.

[0074] If the control index is within the preset risk range, it indicates that the converter's operating deviation exceeds the normal range and requires close monitoring. The control index at this time is marked as b and assigned a secondary character. The b is combined with the secondary character to generate a secondary risk state, matching the enhanced monitoring mode. For example, the laser scanning frequency is increased (for example, to 20Hz), and the corresponding deviation area is scanned in a focused manner (for example, when the furnace tilt angle is abnormal, the furnace cap slope area is scanned more frequently). The infrared thermal imager activates the ROI (region of interest) tracking mode and performs 50Hz high-frequency sampling on the middle and lower parts of the furnace lining (a high-risk area for erosion). The vibration sensor starts synchronous acquisition to capture abnormal vibration signals of the furnace body (for example, the sampling rate is 10kHz).

[0075] If the control index exceeds the preset risk range, it means that the deviation of the converter operation endangers the safety of the lining and requires emergency intervention. The control index at this time is marked as c and assigned a level 3 character. The combination of c and the level 3 character generates a level 3 risk state, matches the emergency alarm mode, and issues a high-frequency alarm.

[0076] Under different control modes, a corresponding spatiotemporal coordinate system is established to synchronize the timestamps and spatially register the multi-source data. Parameter features are extracted, including at least geometric features, temperature features, and vibration features. The extracted parameter features are fused to generate a wear feature matrix.

[0077] Preprocessing: denoising and filtering of multi-source data;

[0078] Establishing a spatiotemporal coordinate system: In conventional monitoring mode, a three-dimensional coordinate system (X: furnace axis, Y: tilting horizontal direction, Z: vertical height) is established with the geometric center of the converter steel shell as the origin. In enhanced monitoring mode, a right-handed coordinate system is constructed with the center or centroid of the deviation corresponding area as the midpoint (normal vector axis: the normal vector of the lining surface at the center point (obtained by fitting the laser point cloud plane), pointing to the outside of the lining (away from the molten steel side), normal axis 1: a unit vector parallel to the converter axis in the tangent plane of the deviation corresponding area, normal axis 2: determined by the right-hand rule).

[0079] Under different control modes, timestamp synchronization and spatial registration (for example, rigid body transformation, fiducial calibration, and ICP fine registration) are performed to map the laser point cloud, infrared pixel, and vibration sensor coordinates to a unified spatial reference, ensuring spatial position consistency of multi-source data. Timestamp synchronization uses a linear interpolation algorithm to resample data at different sampling rates (such as 10Hz for laser scanning and 10kHz for vibration monitoring) to a unified time grid (100Hz), eliminating timing deviations and ensuring high-precision alignment of multi-source data under different control modes. This provides a reliable time and space reference for intelligent monitoring of converter linings.

[0080] Feature extraction: Extract the geometric features of laser point cloud data, including at least surface roughness, volume wear, surface curvature, and normal vector deviation, reflecting the microscopic fluctuations, wear volume, and local structural abnormalities of the furnace lining; extract the temperature features of infrared thermal imaging, including at least temperature gradient and hot spot entropy area, to identify overheating areas; extract the vibration features of the furnace lining vibration signal, including at least peak value, main frequency offset, and wavelet packet energy entropy, to capture impact vibration, resonant frequency changes, and vibration mode complexity, and use multi-source features to jointly quantify the wear status of the furnace lining;

[0081] Feature fusion: Standardize the extracted features and concatenate the standardized geometric, temperature, and vibration features into high-dimensional vectors by dimension, preserving the original feature details. Compress the feature dimensions through principal component analysis or autoencoders (such as AE), retain the principal components with the largest cumulative variance contribution, and eliminate inter-feature correlations (such as the collinearity between temperature gradient and surface curvature). Finally, generate a wear feature matrix.

[0082] Based on the control coefficient mapping association, a corresponding association relationship model is established, including:

[0083] Data preparation: Based on the known wear form and the first characteristic parameter, a multidimensional feature vector is constructed, including the first characteristic parameter, the wear form, and a timestamp. The temperature feature includes at least the temperature gradient and hot spot entropy, and the vibration feature includes at least the peak value, the main frequency offset, and the wavelet packet energy entropy.

[0084] Obtain the average value and fluctuation value of the parameters corresponding to the vibration feature or temperature feature. For any time period Ta, mark the wear form of the time period Ta as Sa, and the wear form includes at least mechanical wear, thermal wear, and corrosive wear. The average value and fluctuation value are p_a and p_b, respectively. Obtain a first judgment vector and a second judgment vector according to different combinations; wherein the average value represents the average value of a characteristic parameter calculated within the time period Ta under a certain feature, and the fluctuation value represents the average value of the difference between the maximum and minimum values ​​of a characteristic parameter calculated within the time period Ta under a certain feature;

[0085] The first judgment vector includes: [p_a+k*p_b, Sa], which represents the comprehensive value of the characteristic parameters after fluctuation amplification and reflects the characteristic deviation under extreme working conditions;

[0086] The second judgment vector includes: [p_a-k*p_b, Sa], which represents the comprehensive value of the characteristic parameters after the fluctuation is reduced and reflects the stability of the characteristic;

[0087] Here, k represents the conditional proportionality index, and k is usually set in the range of 1.5-2 to balance reliability and sensitivity. The specific value is based on the preset classification model for de-peaking. The essence of mean ± k * fluctuation value is to quantify the uncertainty range of the data by superimposing the central trend and the fluctuation amplitude. It is used in engineering for threshold setting, error analysis and system control. When k increases, the range of the judgment vector [p_a±k*p_b, Sa] expands, and the sensitivity to fluctuation characteristics increases. Conversely, when k decreases, the judgment vector is closer to the mean and more robust to small fluctuations. This allows the judgment vector to accurately distinguish different wear forms (Sa), that is, the judgment vectors of the same type of wear are highly clustered, and the judgment vectors of different types of wear are significantly separated.

[0088] Build a classification model: Use [p_a, ±p_b, Sa] in the historical data as the training set, and input the judgment vector constructed with k as the parameter into the classifier (such as SVM, random forest) as the input information. The goal is to make the model have the highest classification accuracy for Sa;

[0089] Cross-validation optimization: Use k-fold cross-validation to automatically search for the optimal k value during training to avoid overfitting. For example, historical data is divided into a training set and a validation set. For each candidate k, a judgment vector is generated using the training set and the classifier is trained. The accuracy is calculated on the validation set, and the k with the highest accuracy is selected.

[0090] Retrieve the control index and conduct correlation analysis, including primary and secondary correlations;

[0091] Primary association: Based on the first judgment vector, the control index is retrieved, a first time tag is created for the control index according to the time series, and an association relationship model between the first judgment vector and the control index is formed;

[0092] Secondary association: Based on the second judgment vector, the control index is retrieved, and a second time tag is created for the second judgment vector according to the time series to form an association relationship model between the second judgment vector and the control index;

[0093] The significance of the above analysis lies in the following: by amplifying the extreme fluctuations of the characteristic parameters through the judgment vector = average value + k × fluctuation value, a strong correlation can be established with the peak operating conditions of the control index (such as full-load equipment operation), thus capturing the wear startup characteristics under high-risk operating conditions; when the control index is too high, the corresponding control mode is also more urgent, indicating that extreme operating conditions may occur in the converter lining. Improving the correlation model of the primary correlation can provide early warning of wear caused by extreme operating conditions; by narrowing the fluctuation range of the characteristic parameters through the judgment vector = average value - k × fluctuation value, a correlation can be established with the steady-state operating conditions of the control index, identifying hidden wear under low fluctuations; when the control index is stable, the corresponding control mode also tends to be regular, indicating that only minor wear may exist in the converter lining or that the wear is slowly developing. The correlation model of the secondary correlation can detect local thermal damage risks of the lining through minor anomalies of the characteristic parameters (such as a decrease in the mean temperature gradient but an increase in the fluctuation value); through the secondary correlation analysis, the misjudgment rate of a single indicator can be reduced to a certain extent, improving the robustness of the model.

[0094] Based on the correlation, the classification range of characteristic parameters for each type of wear is obtained, including:

[0095] Based on the association relationship, the matching values ​​between the first judgment vector and the second judgment vector and the control index are extracted, and the first matching value and the second matching value are obtained accordingly. If the first matching value or the second matching value exceeds a preset matching threshold, the data set corresponding to the matching value is marked as a high matching data set;

[0096] Clustering operations are performed on highly matched data sets to generate wear type classification clusters. Rectangular indicators for each feature point are preset, and matrix indicators include monitoring boundaries, boundary intervals, and boundary windows. Abnormal feature points that exceed the monitoring boundaries are screened out. The rectangular lengths of the screened abnormal feature points fall into the corresponding boundary intervals for classification, and several equally spaced boundary intervals are labeled and weighted, with the labels and weights kept synchronized. Otherwise, normal type classification clusters are marked.

[0097] Monitoring Boundary: Collect a large amount of historical wear data and perform statistics on various features (such as surface curvature and normal vector deviation in geometric features, gradient fluctuation in temperature features, and amplitude in vibration features). Calculate their mean and standard deviation. Typically, the monitoring boundary can be set within the range of the mean plus or minus a certain number of standard deviations. For example, for vibration features, the historical statistical mean of the data is 50 Hz and the standard deviation is 5 Hz. If the monitoring boundary is set to the mean plus 3 times the standard deviation, the lower limit of the monitoring boundary is 50 - 3 × 5 = 35 Hz, and the upper limit is 50 + 3 × 5 = 65 Hz. This setting means that when the vibration feature value exceeds the range of 35 Hz to 65 Hz, the data point may be abnormal and require further attention.

[0098] Boundary intervals: Through in-depth analysis of historical data, such as drawing histograms, we can observe the distribution of data within the monitoring boundary. If we find that data is concentrated in certain areas, we can set narrower boundary intervals in these areas to capture data changes more precisely. In areas with sparse data distribution, we can set wider intervals.

[0099] The divided boundary intervals are changed in sequence. For the boundary interval numbered 1, the system assigns a weight of 1. For the boundary interval numbered 2, the system assigns a weight of 2. And so on. The weights 1 and 2 are just an example. They can also be in the form of a ratio. The specific value depends on the actual situation. It can be expressed that the larger the number, the greater the weight.

[0100] Boundary window: Count the number of feature points in the cluster, select 10% of the points as the window, and move it smoothly to ensure that the window covers enough data and avoids computational redundancy;

[0101] Through range definition, it is determined whether the real-time scanned data parameters fall into the classification range corresponding to normal operating condition fluctuations. If they fall into the normal type classification cluster, it is determined to be normal deformation; conversely, if they fall into the classification cluster for each type of wear, it is marked as abnormal wear and the wear value is obtained. The specific steps include:

[0102] Extract the frequency of abnormal feature points of each feature parameter;

[0103] The eigenvectors corresponding to the classified feature points are weighted and concatenated to form a feature matrix, where each row and column of the feature matrix corresponds to a fused eigenvalue.

[0104] The product of the fused feature value and the frequency of abnormal feature points is marked as the wear value;

[0105] Based on the characteristic parameter classification range, the laser point cloud data, wear form, and wear value under abnormal wear are obtained and imported as input into a pre-set residual thickness prediction model, which outputs the residual thickness value. The residual thickness prediction model has a built-in SVM classifier and neural network architecture.

[0106] Data preparation and processing: Collect laser point cloud data under known abnormal wear conditions, extract the second feature information, and combine it with the corresponding wear form and wear value to obtain the target data set; the second feature parameters include at least geometric features;

[0107] Model training and evaluation: The target dataset is input into the SVM classifier and divided into a training set, a test set, and a validation set. The training set is used to train the SVM classifier, outputting the probability distribution of each wear form. The classification probability output by the SVM is marked as a classification feature, combined with secondary feature information (such as geometric features), and input into the neural network. The attention mechanism layer in the neural network automatically learns the influence weights of different wear forms (represented by classification probabilities) on the residual thickness value, and the classification features are weighted and fused based on the influence weights. The residual thickness prediction value is output through the regression layer. The validation set is used to evaluate the regression performance using the MAE and RMSE indicators, and the validation set is used to evaluate the classification performance using the accuracy, recall, precision, and F1 score indicators.

[0108] Model prediction: Use the trained model to predict the laser point cloud data under new abnormal wear and output the residual thickness value under the corresponding wear form;

[0109] After the above model integrates classification and regression labels, the target dataset contains both the semantics of wear form (type) and residual thickness values, avoiding the loss of feature semantics caused by a single label (residual thickness only). For example, two sets of point cloud data may have similar residual thickness but different wear types (such as uniform wear and localized wear). The classification labels can help the model distinguish their potential differences and reduce prediction ambiguity. To a certain extent, the generalization ability of the model is improved to achieve high-precision predictions, providing a decision-making basis for industrial maintenance (for example, giving priority to equipment with high-weight wear types), thereby shortening equipment downtime inspection time and improving production line efficiency.

[0110] Example 2:

[0111] An embodiment of the present invention provides an online converter lining thickness measurement system based on laser three-dimensional scanning and edge computing; the system includes: an acquisition module, a processing module, and a thickness measurement module, and the acquisition module, the processing module, and the thickness measurement module are communicatively connected;

[0112] Acquisition module, which collects multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals;

[0113] The processing module matches the control mode based on the converter's operating status. Based on the control mode, it performs spatiotemporal registration and feature fusion of multi-source data to generate a wear feature matrix, including control coefficients. Simultaneously, it obtains the lining vibration signal and infrared thermal imaging temperature of known wear patterns, extracts the first characteristic parameters, including temperature and vibration characteristics, and maps and associates them based on the control coefficients to establish a corresponding association relationship model. Based on the association relationship, it obtains the classification range of the characteristic parameters for each type of wear and calculates the corresponding wear value.

[0114] The thickness measurement module obtains laser point cloud data, wear form and wear value under abnormal wear based on the classification range of characteristic parameters, and imports them as input information into the pre-set residual thickness prediction model to output the residual thickness value.

[0115] Weights are determined using the coefficient of variation method, which assigns weights to each indicator based on the degree of variation between its current value and the target value. If the numerical differences between an indicator are large, they can clearly distinguish between the evaluated objects, indicating that the indicator has rich discriminative information and should be given a larger weight. Conversely, if the numerical differences between the evaluated objects on an indicator are small, then the indicator's ability to distinguish between the evaluated objects is weak and should be given a smaller weight. This method directly utilizes the information contained in each indicator to calculate the indicator weight, making it objective.

[0116] In the application, the several formulas involved are all calculated by taking their numerical values ​​after removing the dimensions, and the formula is a formula of the most recent real situation obtained by collecting a large amount of data and performing software simulation. The formula is set by technical personnel in this field according to actual conditions.

[0117] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0118] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0119] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. The online thickness measurement method of converter lining based on laser 3D scanning and edge computing is characterized by: The method comprises: Collect multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals; Based on the converter's operating status, the control mode is matched. Based on the control mode, multi-source data is temporally and spatially aligned and feature-fused to generate a wear feature matrix, including control coefficients. Simultaneously, the lining vibration signal and infrared thermal imaging temperature of known wear patterns are acquired, and the first characteristic parameters, including temperature and vibration characteristics, are extracted. These are mapped and associated based on the control coefficients to establish a corresponding association relationship model. Based on the association relationship, the characteristic parameter classification range for each type of wear is obtained, and the corresponding wear value is calculated. Based on the characteristic parameter classification range, the laser point cloud data, wear form and wear value under abnormal wear are obtained and imported into the pre-set residual thickness prediction model as input information to output the residual thickness value; Among them, based on the control mode, the multi-source data is temporally and spatially aligned and the features are fused to generate the wear feature matrix, including the generation of control coefficients, including: Based on the converter's operating status, operating factors are extracted, including tilting height, tilting angle, and angle dwell time, to form a first operating vector set. The first operating vector set is compared with a corresponding preset baseline operating vector set to obtain different deviations. Different weights are assigned to the different deviations, and the operating environment of the converter lining is retrieved to obtain an environmental compensation coefficient. The control coefficient is obtained by combining the deviation, weight, and environmental compensation coefficient to switch to the corresponding control mode. Under different control modes, a corresponding space-time coordinate system is established, the timestamps of multi-source data are synchronized and spatially aligned, and the second parameter features are extracted, including at least geometric features, temperature features, and vibration features. The extracted parameter features are fused to generate a wear feature matrix.

2. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 1 is characterized in that: The control modes include a normal monitoring mode, an enhanced monitoring mode and an emergency alarm mode.

3. The online converter lining thickness measurement method based on laser three-dimensional scanning and edge computing according to claim 1 is characterized in that: The mapping association based on the control coefficient establishes a corresponding association relationship model; Based on the correlation, the classification range of characteristic parameters for each type of wear is obtained, including: Data preparation: Based on the known wear form and the first characteristic parameter, a multidimensional feature vector is constructed, including the first characteristic parameter, the wear form, and a timestamp. The temperature feature includes at least the temperature gradient and hot spot entropy, and the vibration feature includes at least the peak value, the main frequency offset, and the wavelet packet energy entropy. Vector construction: Obtain the average and fluctuation values ​​of the parameters corresponding to the vibration or temperature characteristics. For any time period Ta, mark the wear form of this time period Ta as Sa, and the average and fluctuation values ​​as p_a and p_b respectively. Then, according to different combinations, obtain the first judgment vector [p_a+p_b, Sa] and the second judgment vector [p_a-p_b, Sa]; Correlation analysis: retrieve the control index, perform correlation analysis, and obtain the corresponding correlation relationship model; wherein, correlation analysis includes primary correlation and secondary correlation.

4. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 3 is characterized in that: The primary association: based on the first judgment vector, calling the control index, creating a first time tag for the control index according to the time series, and forming an association relationship model between the first judgment vector and the control index; The secondary association: based on the second judgment vector, the control index is retrieved, a second time tag is created for the second judgment vector according to the time series, and a correlation relationship model between the second judgment vector and the control index is formed.

5. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 1 is characterized in that: The characteristic parameter classification range of each type of wear is obtained based on the association relationship, including: Based on the association relationship, the matching values ​​between the first judgment vector and the second judgment vector and the control index are extracted, and the first matching value and the second matching value are obtained accordingly. If the first matching value or the second matching value exceeds a preset matching threshold, the data set corresponding to the matching value is marked as a high matching data set; Clustering operations are performed on high-matching data sets to generate wear type classification clusters. The rectangular indicators of each feature point are preset, and the matrix indicators include monitoring boundaries, boundary intervals, and boundary windows. Abnormal feature points that exceed the monitoring boundaries are screened out. The rectangular lengths of the screened abnormal feature points fall into the corresponding boundary intervals for classification, and several equidistant boundary intervals are labeled and weighted, and the labels and weights are kept synchronized. Otherwise, normal type classification clusters are marked.

6. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 1 is characterized in that: The step of obtaining the classification range of characteristic parameters for each type of wear includes: determining whether the data parameters of the real-time scan fall into the classification range corresponding to the normal operating condition fluctuation; if they fall into the normal type classification cluster, it is determined to be normal deformation; otherwise, it is marked as abnormal wear and the wear value is obtained.

7. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 6 is characterized in that: The wear value includes: The frequency of abnormal feature points of each feature parameter is extracted; the feature vectors corresponding to the classified feature points are weighted and concatenated to form a feature matrix, and each row and column of the feature matrix corresponds to a fused feature value; the product of the fused feature value and the frequency of abnormal feature points is marked as the wear value.

8. The method for online thickness measurement of converter lining based on laser three-dimensional scanning and edge computing according to claim 1 is characterized in that: The residual thickness prediction model has a built-in SVM classifier and neural network architecture, including: Data preparation and processing: Collect laser point cloud data under known abnormal wear conditions, extract the second feature information, and combine it with the corresponding wear form and wear value to obtain the target data set; the second feature parameters include at least geometric features; Model training and evaluation: The target dataset is input into the SVM classifier, which is trained to output the probability distribution of each wear form. The classification probability output by the SVM is marked as a classification feature, combined with the second feature information, and input into the neural network. The attention mechanism layer in the neural network automatically learns the influence weights of different wear forms on the residual thickness value. The classification features are weighted and fused based on the influence weights, and the residual thickness prediction value is output through the regression layer. Model prediction: Use the trained model to predict the laser point cloud data under new abnormal wear and output the residual thickness value under the corresponding wear form.

9. The converter lining online thickness measurement system based on laser 3D scanning and edge computing is characterized by: include: Acquisition module, which collects multi-source data, including laser point cloud data, infrared thermal imaging temperature and furnace lining vibration signals; The processing module matches the control mode based on the converter's operating status. Based on the control mode, it performs spatiotemporal registration and feature fusion of multi-source data to generate a wear feature matrix, including control coefficients. Simultaneously, it obtains the lining vibration signal and infrared thermal imaging temperature of known wear patterns, extracts the first characteristic parameters, including temperature and vibration characteristics, and maps and associates them based on the control coefficients to establish a corresponding association model. Based on the correlation, the classification range of characteristic parameters of each type of wear is obtained and the corresponding wear value is calculated; The thickness measurement module obtains laser point cloud data, wear form, and wear value under abnormal wear based on the characteristic parameter classification range, and imports them as input information into the pre-set residual thickness prediction model to output the residual thickness value; Among them, based on the control mode, the multi-source data is temporally and spatially aligned and the features are fused to generate the wear feature matrix, including the generation of control coefficients, including: Based on the converter's operating status, operating factors are extracted, including tilting height, tilting angle, and angle dwell time, to form a first operating vector set. The first operating vector set is compared with a corresponding preset baseline operating vector set to obtain different deviations. Different weights are assigned to the different deviations, and the operating environment of the converter lining is retrieved to obtain an environmental compensation coefficient. The control coefficient is obtained by combining the deviation, weight, and environmental compensation coefficient to switch to the corresponding control mode. Under different control modes, a corresponding space-time coordinate system is established, the timestamps of multi-source data are synchronized and spatially aligned, and the second parameter features are extracted, including at least geometric features, temperature features, and vibration features. The extracted parameter features are fused to generate a wear feature matrix.

Images