A carburized layer growth rate prediction model training method based on historical processes

Abstract

The application discloses a carburized layer growth rate prediction model training method based on historical processes, and relates to the technical field of industrial artificial intelligence.The steps of the method include the following steps: first, the original time sequence data in the pit furnace is obtained by executing probe drift compensation, and the sensor distortion characteristics under high carbon potential are corrected in reverse to generate a corrected carbon potential sequence;second, a spatial environment quantization matrix of the loaded workpiece is constructed, the position parameters and the shielding parameters of the target workpiece in the furnace are obtained, the retarding effect of the local environment in the furnace on the carbon flux is quantized, and a spatial topological feature vector of the target workpiece is generated;finally, a prediction model including dynamic branches and static branches is constructed, the corrected carbon potential sequence and the spatial topological feature vector are respectively input into the corresponding branches for feature fusion, and the carburized layer prediction result of the target workpiece is output.The application solves the feature distortion problem caused by sensor drift under high carbon potential working conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial artificial intelligence technology, specifically to a training method for a carburized layer growth rate prediction model based on historical processes. Background Technology

[0002] With the increasing demands for reliability in transmission components in high-end manufacturing, carburizing and quenching, as a key heat treatment process that imparts high surface hardness and good core toughness to workpieces, requires precise control of its process. However, due to the black-box nature of the furnace environment, the real-time growth depth of the carburized layer cannot be directly and non-destructively measured. Core quality indicators are often highly delayed and can only be obtained through destructive cross-section physicochemical testing after the process is completed.

[0003] To overcome the aforementioned prediction blind spots, existing technologies typically employ two types of solutions: one is to construct a pure numerical simulation model based on traditional diffusion kinetics, and to perform theoretical deductions by inputting constant process parameters; the other is to introduce conventional data-driven methods, using machine learning or deep neural networks to directly extract continuous time-series data recorded by the main control thermocouple and oxygen probe in historical process logs, and to construct an end-to-end prediction model from operating parameters to the final diffusion depth, attempting to map the internal diffusion law of carbon atoms through historical data fitting.

[0004] However, conventional data-driven models, due to their heavy reliance on dial monitoring data representation, suffer from severe distortion and missing dimensions in complex and extreme physical environments. The problems manifest as follows: 1) Sensor drift and input feature distortion at extremely high carbon potentials. Historical carbon potential curves are typically calculated using a zirconia oxygen probe combined with temperature. To improve process efficiency, it is often necessary to exceed traditional temperatures and explore high carburizing carbon potentials. Under these conditions, free carbon black is easily generated in the furnace and rapidly adheres to the oxygen probe surface, leading to probe measurement distortion. If such historical data is directly used as training data, the model will fall into the trap of invalid input leading to invalid output, incorrectly learning the slowdown in growth caused by the actual carbon potential decay as intrinsic diffusion stagnation under constant carbon potential.

[0005] 2) The unique spatial heterogeneity and local flow field interference of pit furnaces. Components such as the wind turbine internal gear ring and sun gear are huge and are usually hoisted or stacked using large pit carburizing furnaces. Historical data dials often only record the environmental scalars of the main control sensors in the upper part of the furnace, ignoring the significant differences between the effective carbon flux and the actual temperature in the pit furnace due to obstructions from tooling racks and dead zones in the fan circulation flow field.

[0006] Therefore, the present invention provides a training method for a carburized layer growth rate prediction model based on historical processes. Summary of the Invention

[0007] The purpose of this invention is to provide a training method for a carburized layer growth rate prediction model based on historical processes, so as to solve the existing problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for training a carburized layer growth rate prediction model based on historical processes, comprising the following steps: Step S1: Acquire historical process data and perform probe drift compensation; acquire the original time series data in the pit furnace, and perform reverse correction on the sensor distortion characteristics under high carbon potential to generate the corrected carbon potential sequence. Step S2: Construct a spatial environment quantization matrix for loading workpieces; obtain the position parameters and occlusion parameters of the target workpiece in the furnace, and quantify the blocking effect of the local environment in the furnace on carbon flux accordingly, and generate the spatial topological feature vector of the target workpiece. Step S3: Output prediction results based on dual-branch feature fusion; construct a prediction model including dynamic and static branches; input the corrected carbon potential sequence and the spatial topological feature vector into the corresponding branches for feature fusion, and output the diffusion layer prediction results of the target workpiece.

[0009] A further improvement of the present invention is that the dynamic compensation of the true carbon potential value in step S1 is performed by a carbon potential reconstruction anti-drift strategy; the carbon potential reconstruction anti-drift strategy performs reverse correction on the sensor contamination distortion segment in the original carbon potential time series by integrating a multi-source atmosphere cross-validation mechanism and a dynamic error attenuation mechanism to obtain the true carbon diffusion driving force.

[0010] A further improvement of this invention is that the multi-source atmosphere cross-validation mechanism specifically includes: The original carbon potential time series and temperature time series inside the well furnace were obtained, and the moisture content gradient of the furnace gas atmosphere was extracted simultaneously. As a cross-validation feature; Set a set temperature threshold and a carbonization threshold for determining extreme operating conditions; when the temperature time series is detected to be higher than the set temperature threshold, and the target carbon potential of the original carbon potential time series is greater than the carbonization threshold for a preset time window, analyze the moisture content gradient in real time. Abnormal volatility; If the original carbon potential time series tends to be stationary, but the moisture content gradient If the abnormal fluctuation rate indicates a decrease in the actual carbon flux in the furnace, it is determined that the sensor has experienced carbon drift and the dynamic error attenuation mechanism is triggered.

[0011] A further improvement of this invention is that the dynamic error attenuation mechanism specifically includes: At the initial moment of triggering the dynamic error attenuation mechanism, the probe contamination attenuation coefficient is activated. ; The probe contamination attenuation coefficient The calculation is based on the integral accumulation of the excess carbon potential, and the formula is as follows: ,in, The carbon deposition rate constant is obtained based on historical calibration experiments. The carbon deposition threshold is... The starting time is... This is the original carbon potential time series; Using the multiplication correction model Perform reverse correction to output the corrected carbon potential sequence after removing sensor drift error. .

[0012] A further improvement of the present invention is that the spatial environment quantization matrix in step S2 is executed by a spatial heterogeneity quantization strategy; the spatial heterogeneity quantization strategy transforms environmental differences into structured features by integrating a static entity occlusion evaluation mechanism and a dynamic flow field attenuation mapping mechanism.

[0013] A further improvement of this invention is that the static entity occlusion evaluation mechanism includes: establishing a coordinate system based on the physical structure of the pit furnace, and extracting the radial distance of the target workpiece in the coordinate system. With axial height ; Import the three-dimensional bounding box size parameters of the target workpiece and adjacent workpieces; Based on the geometric projection of adjacent workpieces onto the target workpiece, the effective exposed area ratio is calculated, and the local shading coefficient, which characterizes the degree of carbon atom diffusion hindrance, is derived. .

[0014] A further improvement of the present invention is that the dynamic flow field attenuation mapping mechanism includes: obtaining the initial flow field parameters of the main circulating fan of the well furnace; Based on the radial distance r, axial height z, and workpiece density, calculate the kinetic energy dissipation coefficient of the flow field when the fluid penetrates the furnace loading level. The local occlusion coefficient and the flow field kinetic energy dissipation coefficient are combined. The vectors are concatenated to generate the spatial topological feature vector. .

[0015] A further improvement of this invention lies in that the carburized layer prediction result based on dual-branch feature fusion is executed by a carburized layer evolution differential prediction model; the carburized layer evolution differential prediction model integrates a time-series fidelity anti-perturbation strategy and a spatial carbon flux alignment strategy to compare the corrected carbon potential sequence with the static spatial topological feature vector. The process involves fusion and alignment to ultimately generate a carburizing control decision intent for the target workpiece.

[0016] A further improvement of this invention is that the time-series fidelity anti-interference strategy includes: activating the differential writing and generation module within the differential prediction model of the permeation layer evolution during the extraction of nonlinear diffusion time-series features; the differential writing and generation module is based on the moisture content gradient. The changes in this process, using the actual driving forces of carbon diffusion as high-quality features, are differentially incorporated into the current time-series feature sequence to reconstruct a distortion-resistant hidden state vector. .

[0017] A further improvement of this invention is that the spatial carbon flux alignment strategy includes: constructing a three-dimensional hysteresis compensation matrix in tensor space from the local effective carbon flux reduction features; using the three-dimensional hysteresis compensation matrix, applying attenuation penalties of local flow field and environmental occlusion to the extracted temporal latent variables, and calculating the effective carbon atom adhesion rate of the target workpiece at that specific location.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention first solves the problem of input feature distortion caused by carbon drift of oxygen probe under extreme high carbon potential conditions by introducing joint feature preprocessing of dynamic compensation of true carbon potential value and three-dimensional spatial topology mapping. It achieves high-fidelity reconstruction of the unmeasurable true carbon diffusion driving force, transforms the complex spatial heterogeneity into a structured feature vector that the model can directly analyze, and ensures the reliability of accurate prediction of different parts in the same furnace in large-scale industrial production.

[0019] 2. By constructing a differentiated prediction model for infiltration layer evolution that includes dynamic temporal and static spatial branches, the limitations of traditional prediction models, which cannot distinguish local spatial location errors due to the one-furnace-one-value approach, are solved. This model achieves tensor alignment between temporal diffusion patterns and spatial local carbon flux reduction weights, significantly improving the robustness and accuracy of depth prediction under extreme process boundaries. Attached Figure Description

[0020] Figure 1 This is a flowchart of a training method for a carburized layer growth rate prediction model based on historical processes, according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0022] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0023] This embodiment describes a process in which a large-scale pit-type gas carburizing furnace performs deep carburizing heat treatment on core gear components of a wind turbine drive system. In the current wind turbine gearbox manufacturing field, the required depth of the carburized layer on gear components is typically high. During the process, the carbon potential inside the furnace needs to rise to a specified range during the strong carburizing stage. However, under these conditions, the reliability of historical process data is affected by two factors: On the one hand, under high carbon potential conditions, the output electromotive force of the oxygen probe deviates from the true carbon-oxygen balance value due to carbon black adhesion, and the carbon potential curve recorded by the data acquisition system shows a false stability phenomenon. On the other hand, due to the obstruction of the tooling rack and the spatial attenuation of the fan circulation flow field, the actual effective carbon flux at different locations of the workpiece in the pit furnace has a significant gradient difference. Using only a uniform carbon potential curve for the entire furnace as the model input cannot fit the deviation in the diffusion layer depth caused by the furnace loading position.

[0024] To address the two issues mentioned above, this invention proposes a training method for a carburized layer growth rate prediction model based on historical processes: First, probe drift compensation is performed on historical carbon potential time-series data; then, a spatial environment quantification matrix is ​​constructed based on the spatial position and geometric occlusion relationship of the workpiece in the furnace; finally, the two sets of information are fed into a prediction model containing dynamic and static branches for differential fusion, and the prediction results of the carburized layer depth for each workpiece are output.

[0025] Example 1 This embodiment is applicable to the following typical scenario: A wind turbine gearbox manufacturer needs to use the historical process data accumulated over the past three years in its pit-type carburizing furnace to train a carburized layer growth rate prediction model, so as to predict the final carburized layer depth of workpieces at different locations in the furnace piece by piece before the start of a new batch of internal gear ring carburizing tasks, based on the proposed process curve and furnace loading plan.

[0026] Figure 1 This embodiment presents a flowchart of a training method for a carburized layer growth rate prediction model based on historical processes. The steps are as follows: Step S1: Obtain historical process data and perform probe drift compensation.

[0027] Raw time-series data was exported from the data acquisition system of the pit-type carburizing furnace, including... Time series of carbon potential recorded at sampling intervals Temperature time series The data includes methanol flow rate sequences, propane flow rate sequences, and dew point temperature sequences. In high carbon potential conditions where both the carbon potential setpoint and temperature exceed the setpoint, free carbon in the furnace is deposited as carbon black on the surface of the platinum electrode of the oxygen probe, causing the recorded carbon potential to be higher than the true value. If such distorted data is directly used for training, the model will incorrectly attribute the decrease in carburization rate caused by sensor drift to changes in the diffusion kinetics themselves.

[0028] Therefore, this step performs reverse correction on the aforementioned sensor distortion—using an auxiliary atmosphere signal in the furnace gas, independent of the oxygen probe, to cross-verify the probe output. After determining that the probe has entered the carbonization drift state, the attenuation of the probe sensitivity is estimated based on the carbonization rate model, and a multiplicative reduction correction is applied to the original carbon potential to generate a corrected carbon potential sequence. .

[0029] Taking a certain furnace's strong percolation stage as an example: initial carbon potential exist After hours, it remained stable at The corrected carbon potential exist It then showed a slow downward trend, until Time dropped to about This is consistent with the carbon potential trajectory derived from metallographic analysis after the furnace cycle.

[0030] Step S2: Construct the spatial environment quantization matrix for loading the workpiece.

[0031] After step S1 completes the time-domain correction of the carbon potential sequence, the carbon potential time series can reflect the carbon activity changes at the location of the main control probe. However, due to the influence of the airflow circulation path and mutual obstruction of the workpieces in different areas of the furnace, the actual flux of carbon atoms reaching the surface of each workpiece varies spatially. This step obtains the positional parameters of the target workpiece in the furnace, including radial distance. With axial height Based on the shading parameters, the distance between the target workpiece and adjacent workpieces, and the outer contour dimensions of adjacent workpieces, the hindering effect of the local environment within the furnace on carbon flux is quantified, and a spatial topology vector of the target workpiece is generated. .

[0032] In this embodiment for A dimensional vector, whose components include, in order, a normalized radial distance, a normalized axial height, and a local occlusion coefficient. Flow field kinetic energy dissipation coefficient The furnace pack density factor and the radial-axial interaction attenuation factor.

[0033] Step S3: Output prediction results based on dual-branch fusion.

[0034] Corrected carbon potential sequence Belongs to high-dimensional dynamic time-series data, spatial topological vectors This data is low-dimensional and static. Simply concatenating the two data points and feeding them into a single branch would result in the static low-dimensional vector being overwhelmed by the high-dimensional temporal sequence, and the spatial information would vanish during backpropagation. This step constructs a prediction model including dynamic and static branches: the dynamic branch receives the carbon potential and temperature time series to extract the nonlinear temporal evolution of the carbon diffusion process; the static branch receives the spatial topology vector to extract the hindrance modes of the local environment. After the outputs of the two branches are aligned in the fusion layer, the predicted carburized layer depth of the target workpiece is output via a fully connected regression head. .

[0035] In this embodiment, the dynamic branching uses a two-layer gated recurrent unit (GRU) network, with a hidden layer dimension of... The input sequence length is the number of time steps for the entire furnace cycle. Static branches use a three-layer fully connected network, with each layer having a width of... The activation function is ReLU. The fusion layer will dynamically branch the final hidden state ( (dimensional) and static branch output ( (V) spliced ​​together Dimension, then through two fully connected layers ( Output scalar predictions. Training uses the mean squared error loss function, Adam optimizer, and batch size... ,train Rounds. Training and validation sets are arranged by furnace. Division.

[0036] Example 2 Based on the inventive concept of Embodiment 1, the present invention provides a complete preferred embodiment. The following three steps are performed sequentially: S1. Reverse correction is executed by the carbon potential reconstruction anti-drift strategy, which includes two cooperative sub-mechanisms: a multi-source atmosphere cross-verification mechanism, configured to determine whether the sensor has entered the carbon junction drift state and output the trigger time. And a dynamic error attenuation mechanism, to A quantitative correction is performed on the carbon potential, starting from this point.

[0037] (a) Multi-source atmosphere cross-validation mechanism In conventional industrial practice, determining whether an oxygen probe has failed typically relies on periodic calibration using standard gas. However, this method cannot capture progressive drift in real time during the calibration interval, and interrupting the process during the high carbon potential and strong percolation stage disrupts the uniformity of carbon potential within the furnace. This invention introduces a multi-source atmosphere cross-verification mechanism, achieving continuous online drift determination through a second information channel independent of the oxygen probe.

[0038] First, three synchronous timing signals are acquired from the data acquisition system: the original carbon potential timing sequence. Temperature time series and the time series of dew point temperatures measured by a dew point meter. Based on dew point temperature The system converts water vapor partial pressure into furnace gas water vapor partial pressure by using the correlation between water vapor partial pressure and dew point temperature, and calculates the moisture content gradient between adjacent sampling times. That is, the difference in water vapor partial pressure between two adjacent sampling points divided by the sampling interval. .

[0039] The rationale for using a moisture content gradient as the cross-validation signal is as follows: In the carburizing furnace, the carbon potential and water vapor content are thermodynamically coupled through the water-gas reaction. When carbon activity increases, the equilibrium shifts towards consuming water vapor; when carbon activity decreases, the moisture content rises again. Therefore, if the oxygen probe shows a stable carbon potential but the dew point meter indicates an abnormal rise in moisture content, it means that the carbon activity in the furnace has actually decreased. The dew point meter's sensing element does not undergo a carbon deposition reaction with the high-carbon-potential furnace gas, and its drift risk is far lower than that of the oxygen probe, making it suitable as an independent verification channel.

[0040] Two thresholds were then set to define the trigger range. Temperature threshold and carbonization threshold ; By obtaining the target set temperature for the strong permeation stage required by the current process and subtracting the preset bias temperature value, we can ensure that the determination is activated in the early stage of carbonization, while avoiding false triggering caused by normal temperature fluctuations. The critical saturation coefficient is obtained by multiplying the solid solution limit of carbon in austenite at the current furnace temperature by a critical saturation coefficient less than 1. The selection of this critical saturation coefficient aims to balance the sensitivity of the judgment with the false trigger rate, ensuring a timely response when the carbon activity approaches the carbonization boundary. This threshold balances sensitivity and false trigger rate—if it is too low, normal fluctuations will be misjudged; if it is too high, it will not be able to respond in time when the carbon potential just crosses the carbonization boundary.

[0041] When temperature time series consecutive greater than And the original carbon potential Greater than Reaching the preset time window In this embodiment, the preset time window can be based on accelerated aging test data of the same type of oxygen probe, extracting the initial time when the probe output shows a significant deviation due to carbon deposition, and setting it using the lower bound of its statistical distribution, thereby triggering correction immediately at the initial stage of sensor drift. This initiates the analysis of abnormal fluctuations in the moisture content gradient. Abnormal fluctuation rate. Take the moving standard deviation of the moisture content gradient series within the window. With the absolute value of the moving average The ratio, with the addition of a small positive number to the denominator to prevent zero. To avoid calculation failure when the denominator is zero, the abnormal volatility threshold is set to... In this embodiment, the abnormal volatility threshold can be obtained by statistically analyzing the volatility distribution of moisture content gradient in historical steady-state carburizing data and selecting specific high quantiles as the judgment criterion.

[0042] If the original carbon potential sequence tends to be stationary, that is, the sliding standard deviation of the carbon potential is below a certain set value, for example... However, abnormal volatility Greater than If the dew point meter indicates a directional change in the furnace gas composition, but the oxygen probe does not reflect this change, then it is determined that the sensor has undergone carbonization drift, and the current time is recorded as [missing information]. This triggers the dynamic error decay mechanism.

[0043] In some potential application scenarios, multiple stages of alternating strong carburizing and diffusion processes (pulse carburizing) are executed sequentially in the same furnace. During the diffusion stage, when the carbon potential switches from a high value to a low value, the moisture reading on the dew point meter will also rise. To avoid misjudging normal process switching as drift, this invention adds a process stage identifier verification to the triggering logic: before executing the abnormal fluctuation rate judgment, the current stage identifier is read first, and the judgment is only valid when it is in the constant carbon potential strong carburizing stage, thereby avoiding false positive triggers.

[0044] (ii) Dynamic error attenuation mechanism At the output trigger moment of the multi-source atmosphere cross-validation mechanism After that, it is necessary to... The carbon potential in subsequent time periods is quantitatively corrected. Simply discarding or replacing the drift period data with a fixed set value will lead to a break in the time sequence integrity and the introduction of new deviations, respectively. This invention introduces a dynamic error decay mechanism, which achieves continuous correction step by step by establishing an analytical model of how the probe sensitivity decays with the accumulation of carbon deposits.

[0045] At the trigger time Activate the probe contamination attenuation coefficient . Characterizing time ( The gain attenuation ratio of the probe output relative to the true carbon potential, with a value range of [value missing]. ,exist hour . Calculations are based on the integral accumulation of the excess carbon potential: in, The carbon deposition rate constant is expressed in units of 1000 m³ / s. This characterizes the attenuation efficiency of probe sensitivity due to the accumulation of excess carbon potential per unit. The acquisition method can be as follows: place an oxygen probe of the same model in a furnace gas with a known carbon potential calibrated by an independent infrared carbon analyzer, and record the trajectory of the probe output deviation over time under different baseline carbon potential levels and temperature gradients. Subsequently, empirical parameters under multiple test conditions are obtained through least-squares fitting, and the geometric mean of these parameters is taken as the final constant term. Using the geometric mean instead of the arithmetic mean ensures better deviation symmetry under exponential calculations.

[0046] This is the set carbon deposition threshold. It serves as a baseline in the integral formula; only the portion of the carbon potential exceeding this baseline contributes to the attenuation. When the carbon potential is below this value, the probability of free carbon precipitation is extremely low, and carbon deposition on the probe surface is negligible, requiring no correction.

[0047] The drift trigger time (output by the aforementioned cross-validation mechanism). At the current moment. Integrand For any intermediate time The extent to which the initial carbon potential exceeds the carbonization threshold, and the quantity obtained by integrating it over the time axis, has... The dimension of reflects the total accumulation of carbon deposits on the probe surface.

[0048] Using an exponential decay function As The reason for this form is that carbon deposition on the platinum electrode surface of the probe follows first-order reaction kinetics—the deposition rate is proportional to the currently uncovered area, decreasing as the covered area increases, resulting in an exponential decay of sensitivity. Compared to a linear decay model, the exponential form automatically limits the output value to... The interval is monotonically decreasing, requiring no additional truncation operation, thus avoiding meaningless negative gains in linear models after long-term operation.

[0049] In the engineering implementation, the above continuous integral is completed step by step through recursion using the discrete rectangle method: in As of the date The cumulative amount of the excess carbon potential at each sampling time, initial value The option is 0. Optional . Ensure that the carbon potential is accidentally less than The sampling points do not produce negative accumulation. .

[0050] Finally, the multiplication correction model is used: Multiply the original carbon potential by less than The attenuation coefficient reflects the fact that the probe overestimates the carbon potential due to carbon deposition, and the output is a corrected carbon potential sequence after eliminating sensor drift error. .

[0051] In some potential application scenarios, multiple cycles of strong percolation-diffusion may occur within a single furnace. Each time the furnace re-enters the strong percolation stage from the diffusion stage, if the carbon potential triggers a new drift determination, Instead of starting from scratch, this invention builds upon the previous layer, increasing incrementally—the carbon layer already deposited on the probe surface will not completely desorb due to a brief period of low carbon potential. This invention maintains... The continuous increase reflects the irreversible accumulation of carbon deposition. If the probe has undergone high-temperature ablation cleaning during the diffusion stage, the operator can manually remove it. Reset to zero.

[0052] Regarding fault tolerance: If There are missing points due to interruptions in data acquisition communication. These are filled by linear interpolation between two consecutive valid sampling points to ensure the continuity of the integral. If an abnormal jump occurs at a certain sampling moment due to electromagnetic interference (such as a sudden change in carbon potential to greater than a certain value), the missing points are filled by linear interpolation between two consecutive valid sampling points to ensure the continuity of the integral. Median filtering is performed before integration to prevent individual outliers from affecting the overall system. This causes an irreversible jump.

[0053] S2 is executed by a spatial heterogeneity quantification strategy and contains two cascaded sub-mechanisms: a static entity occlusion evaluation mechanism and a dynamic flow field attenuation mapping mechanism.

[0054] (a) Static entity occlusion assessment mechanism In existing methods, researchers have used the layer number of the workpiece in the furnace as a discrete encoding input model. However, this approach cannot distinguish the differences between different radial positions within the same layer, nor can it reflect the influence of changes in workpiece size and spacing. This invention quantifies the degree of carbon atom diffusion hindrance experienced by each workpiece at a continuous value level.

[0055] First, a cylindrical coordinate system is established based on the well-type furnace shell, with the origin set at the center of the bottom of the furnace. The shaft is oriented upwards along the vertical central axis. The axis points radially towards the furnace wall. Each target workpiece... The position is represented as , Radial distance ( ), Axial height ( Then import the 3D bounding box dimensions (outer diameter) of the target workpiece and all adjacent workpieces. ,high , inner diameter In this embodiment, the three-dimensional bounding box of the internal gear ring is a cylindrical annulus.

[0056] Local occlusion coefficient The calculation method is as follows: Each adjacent workpiece is divided along the main transport direction of carbon atoms. The 3D bounding box is projected onto the target workpiece. On the central reference plane, the projected area is obtained. ( The sum of the projected areas of all adjacent workpieces and the total exposed surface area of ​​the target workpiece in an unobstructed state. ( By comparison, we can obtain : Including the sum of the areas of the outer cylindrical surface, the upper and lower end faces, and the inner cylindrical surface, in this embodiment... Indicates no obstruction. The larger the value, the more severe the carbon flux obstruction.

[0057] In some potential application scenarios, the furnace loading layout is highly dense, and the projected areas of adjacent workpieces may overlap. To avoid the overlapping of occlusion areas, this invention performs a Boolean union operation on the projected contours of all adjacent workpieces (taking the union area instead of simple summation) before performing the summation of projected areas, ensuring... It is not too high due to overlapping projections.

[0058] (II) Dynamic flow field attenuation mapping mechanism Static occlusion assessment quantifies the obstruction of geometric occlusion to carbon transport paths, but it does not consider the influence of airflow velocity on the mass transfer rate of carbon carrier molecules—the lower the airflow velocity, the thicker the boundary layer near the workpiece surface, and the greater the mass transfer resistance of carbon atoms across the boundary layer. This mechanism utilizes the flow field kinetic energy dissipation coefficient... Quantify the degree of airflow attenuation at the target workpiece location.

[0059] Obtain the initial flow field parameters of the main circulation fan: rated speed. impeller diameter Initial airflow velocity at the fan outlet First, calculate the workpiece layout density of the layer containing the target workpiece. , that is to The ratio of the sum of the volumes of the three-dimensional bounding boxes of all workpieces within the half-layer spacing above and below the center to the volume of the annular space of the furnace corresponding to that layer, with a value range of... .

[0060] The calculation formula is: in, The initial airflow velocity at the fan outlet.

[0061] The axial attenuation coefficient is expressed in units of... The preferred value represents the velocity attenuation efficiency caused by the packing density when the airflow penetrates the packing layer per unit height along the axial direction downwards. In this embodiment The actual furnace layout and local arrangement density can be used as a reference. As boundary conditions, a computational fluid dynamics (CFD) model of the target pit furnace is established and a steady-state solution is performed. Mathematical regression is then used to obtain the average airflow velocity decay curves of different axial height sections output by the model.

[0062] Radial attenuation coefficient, unit The preferred value represents the velocity attenuation efficiency caused by wall friction and vortex effect when the airflow deflects from the center of the furnace towards the furnace wall. The method of obtaining it is also based on the velocity profile fitting results at different radial positions in the above CFD steady-state solution model.

[0063] Range of values The larger the value, the more severe the airflow attenuation.

[0064] This indicates when the decay term is zero. (No decay), when the decay term approaches infinity Approaching (Complete decay) automatically limits the output to a reasonable range, avoiding potential exceedances that might occur with linear models. Meaningless values. When the target workpiece is at the very top of the furnace ( And when there is no material layer above, Degenerate into only with The relevant single variable is consistent with the actual situation that the airflow attenuation of the top workpiece is mainly determined by the radial wall effect.

[0065] Will and By concatenating the positional normalization value and the interaction factor, a spatial topological vector is generated: in This is the radius of the furnace inner wall. Sixth component. The radial-axial interactive attenuation factor represents the superimposed nonlinear enhancement of airflow dead zone effect and solid blockage effect at locations with simultaneously large radial distance and axial depth (furnace bottom and corner areas), where the airflow dead zone effect and solid blockage effect are superimposed. and The independent components cannot express this second-order interaction; the product term enables the subsequent model to capture the decay of this interaction. After normalization, all components are within... Within the interval, avoid letting components with different dimensions dominate gradient updates due to significant numerical differences.

[0066] S3 is executed by the differential prediction model of infiltration layer evolution, which integrates the temporal fidelity anti-disturbance strategy and the spatial carbon flux alignment strategy to fuse and align the two information streams.

[0067] (a) Timing-fidelity anti-interference strategy When extracting the nonlinear diffusion evolution law in the carbon potential time series in dynamic branching, it is ensured that the anti-drift information carried by the carbon potential sequence after step S1 is effectively preserved, rather than gradually decaying during long sequence propagation.

[0068] In this embodiment, the dynamic branching uses a two-layer GRU network as a temporal encoder (hidden layer dimension). In a standard GRU, the hidden state updates at each time step are controlled by gating mechanisms of update gates and reset gates. This invention builds upon this by introducing a differentiated write and generate module—this module is based on the moisture content gradient. The changes will be written into the hidden state as high-quality inputs to correct the carbon diffusion driving force information carried by the carbon potential, guiding the model to pay more attention to the period when the carbon potential is corrected (the key period of actual change of the carbon potential environment during carburizing) during the training process, rather than the stable illusory period that is covered by probe distortion.

[0069] The differential write and generate module works as follows. At each time step... Construct the input augmentation vector: in This is the carbon potential correction amount. The value is zero during periods when drift correction is not triggered, and positive during periods when drift correction is triggered. (Introduction) This enables the model to distinguish between two different operating conditions: one where the carbon potential is inherently low, and the other where the carbon potential appears high but is reduced after correction.

[0070] Then, the write attention weights are calculated. : in For trainable weight vectors, For bias, This is the Sigmoid function. Range of values ,when and When larger, it approaches This indicates that there is a significant atmosphere anomaly and carbon potential correction at this time step, and it should be written with high priority; otherwise, it should be written with lower priority. This indicates a normal, stable period.

[0071] Will Introducing the hidden state update of GRU, the modified update formula is as follows: in Update the gate output for GRU. This is a candidate hidden state. For element-wise multiplication, The trainable weight matrix is ​​injected. Compared to the standard GRU update formula, a third term is added here in the attention weights. Under regulation, the modified carbon potential and its modification amount are directly injected into the latent state, bypassing the information screening of the gating mechanism.

[0072] Standard GRU gating tends to preserve gradual trends and filter out rapid signals when dealing with long sequences. However, the rapid signals revealed by carbon potential drift correction (the sudden jump in correction amount after drift triggering) precisely reflect the key signals of substantial changes in carburizing kinetics and should not be overlooked by the gating. By directly injecting the signal, bypassing the gating, this signal can be continuously maintained in the hidden state, thereby reconstructing a distortion-resistant hidden state vector. .

[0073] In some potential application scenarios, if a company's carburizing furnace is equipped with multiple oxygen probes (e.g., one each in the upper and lower parts of the furnace), the multi-source atmosphere cross-verification mechanism can simultaneously utilize the differences in readings from multiple probes as supplementary signals for drift determination. Add corresponding inter-probe differential channels in the dimension. The column dimensions are also expanded accordingly.

[0074] (II) Spatial carbon flux alignment strategy Timing hidden state of dynamic branch output ( For the last time step, The spatial topology vector reflects the evolution of the carbon potential throughout the furnace, but the actual effective carbon atom flux received by workpieces at different locations varies due to obstruction and flow field attenuation. This strategy uses spatial topology vectors to reflect the evolution of the carbon potential throughout the furnace. The encoded local environmental hindrance information is transformed into a dimension-wise decay modulation of the temporal hidden state, realizing a dimension-reduction mapping from the carbon potential of the entire furnace to the effective carbon flux of the individual workpiece.

[0075] Firstly based on Constructing a three-dimensional hysteresis compensation matrix The three dimensions refer to the matrix, which integrates the combined attenuation effects of the radial, axial, and occlusion dimensions. Generated from a fully connected network with static branches: The network structure is as follows: input layer ( Dimension), first hidden layer ( Dimension, ReLU), second hidden layer ( Dimension, ReLU), third hidden layer ( (Dimension, ReLU), output layer ( (Dimensional, no activation), the output is reshaped as matrix.

[0076] Then use Apply a decay penalty to the temporally hidden state: For the element-wise Sigmoid function, ensure the decay coefficient is within... Within the range. For locations with severe obstruction and significant flow field attenuation. Each element tends to take the smaller value. Information reflecting the driving force of high carbon potential is attenuated, causing the prediction to shift towards shallower infiltration layers; for open-area workpieces, the elements tend to converge. Temporal hidden states are almost unpunished.

[0077] Compared to the splicing and fusion method in Example 1, here... As a modulation matrix generated from spatial information, an independent attenuation coefficient determined by spatial location is applied to each dimension of the temporal latent state, achieving multiplicative fusion dimension by dimension. This mechanism has stronger cross-modal alignment capability than additive splicing. In the splicing method, the two types of information only interact implicitly through the weight matrix of the subsequent fully connected layer, and the interaction depth is limited; while in multiplicative fusion, spatial information directly modulates each expression dimension of the temporal information, which can be regarded as imposing a channel capacity constraint encoded by spatial information on the temporal information stream.

[0078] Predicted carburized layer depth output by regression head: It is a two-layer fully connected regression head that outputs a scalar. .

[0079] The training configuration includes: Training labels The actual carburized layer depth is obtained by metallographic sectioning after carburizing of each workpiece. The loss function used is Huber loss, with a threshold parameter... : Total loss In this embodiment select This can be obtained by acquiring the baseline measurement uncertainty of the currently used metallographic testing equipment when determining the effective hardened layer depth, and multiplying it by a preset tolerance factor. When the prediction error is less than... When the error is greater than or equal to the specified value, the model parameters are finely adjusted using a squared penalty; when the error is greater than or equal to the specified value... When the network gradient is over-amplified, it automatically switches to linear penalty to prevent extreme label bias or outlier data points from amplifying the network gradient.

[0080] The optimizer used is AdamW.

[0081] The threshold and weight settings involved in this embodiment can be set by default according to the present invention, or can be set by those skilled in the art.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0084] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0085] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0086] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A training method for a carburized layer growth rate prediction model based on historical processes, characterized in that: Includes the following steps: Step S1: Acquire historical process data and perform probe drift compensation; acquire the original time series data in the pit furnace, and perform reverse correction on the sensor distortion characteristics under high carbon potential to generate the corrected carbon potential sequence. Step S2: Construct a spatial environment quantization matrix for loading workpieces; obtain the position parameters and occlusion parameters of the target workpiece in the furnace, and quantify the blocking effect of the local environment in the furnace on carbon flux accordingly, and generate the spatial topological feature vector of the target workpiece. Step S3: Output prediction results based on dual-branch feature fusion; construct a prediction model including dynamic and static branches; input the corrected carbon potential sequence and the spatial topological feature vector into the corresponding branches for feature fusion, and output the diffusion layer prediction results of the target workpiece.

2. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 1, characterized in that: The dynamic compensation of the true carbon potential value in step S1 is performed by the carbon potential reconstruction anti-drift strategy. The carbon potential reconstruction anti-drift strategy reverses the sensor contamination distortion segment in the original carbon potential time series by integrating a multi-source atmosphere cross-validation mechanism and a dynamic error attenuation mechanism to obtain the true carbon diffusion driving force.

3. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 2, characterized in that: The multi-source atmosphere cross-validation mechanism specifically includes: The original carbon potential time series and temperature time series inside the well furnace were obtained, and the moisture content gradient of the furnace gas atmosphere was extracted simultaneously. As a cross-validation feature; Set a set temperature threshold and a carbonization threshold for determining extreme operating conditions; when the temperature time series is detected to be higher than the set temperature threshold, and the target carbon potential of the original carbon potential time series is greater than the carbonization threshold for a preset time window, analyze the moisture content gradient in real time. Abnormal volatility; If the original carbon potential time series tends to be stationary, but the moisture content gradient If the abnormal fluctuation rate indicates a decrease in the actual carbon flux in the furnace, it is determined that the sensor has experienced carbon drift and the dynamic error attenuation mechanism is triggered.

4. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 2, characterized in that: The dynamic error attenuation mechanism specifically includes: At the initial moment of triggering the dynamic error attenuation mechanism, the probe contamination attenuation coefficient is activated. ; The probe contamination attenuation coefficient The calculation is based on the integral accumulation of the excess carbon potential, and the formula is as follows: ,in, The carbon deposition rate constant is obtained based on historical calibration experiments. The carbon deposition threshold is... The starting time is... This is the original carbon potential time series; Using the multiplication correction model Perform reverse correction to output the corrected carbon potential sequence after removing sensor drift error. .

5. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 1, characterized in that: The spatial environment quantization matrix in step S2 is executed by a spatial heterogeneity quantization strategy; the spatial heterogeneity quantization strategy transforms environmental differences into structured features by integrating a static entity occlusion assessment mechanism and a dynamic flow field attenuation mapping mechanism.

6. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 5, characterized in that: The static entity occlusion assessment mechanism includes: establishing a coordinate system based on the physical structure of the well furnace, and extracting the radial distance of the target workpiece in the coordinate system. With axial height ; Import the three-dimensional bounding box size parameters of the target workpiece and adjacent workpieces; Based on the geometric projection of adjacent workpieces onto the target workpiece, the effective exposed area ratio is calculated, and the local shading coefficient, which characterizes the degree of carbon atom diffusion hindrance, is derived. .

7. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 6, characterized in that: The dynamic flow field attenuation mapping mechanism includes: obtaining the initial flow field parameters of the main circulating fan of the well furnace; Based on the radial distance r, axial height z, and workpiece density, calculate the kinetic energy dissipation coefficient of the flow field when the fluid penetrates the furnace loading level. The local occlusion coefficient and the flow field kinetic energy dissipation coefficient are combined. The vectors are concatenated to generate the spatial topological feature vector. .

8. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 3, characterized in that: The carburized layer prediction result based on dual-branch feature fusion is executed by the carburized layer evolution differential prediction model; the carburized layer evolution differential prediction model integrates a time-series fidelity anti-disturbance strategy and a spatial carbon flux alignment strategy to compare the corrected carbon potential sequence with the static spatial topological feature vector. The process involves fusion and alignment to ultimately generate a carburizing control decision intent for the target workpiece.

9. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 8, characterized in that: The time-series fidelity anti-interference strategy includes: activating the differential writing and generation module within the differential prediction model of the infiltration layer evolution during the extraction of nonlinear diffusion time-series features; the differential writing and generation module is based on the moisture content gradient. The changes in this process, using the actual driving forces of carbon diffusion as high-quality features, are differentially incorporated into the current time-series feature sequence to reconstruct a distortion-resistant hidden state vector. .

10. The method for training a carburized layer growth rate prediction model based on historical processes according to claim 8, characterized in that: The spatial carbon flux alignment strategy includes: constructing a three-dimensional hysteresis compensation matrix in tensor space from the local effective carbon flux reduction features; using the three-dimensional hysteresis compensation matrix, applying attenuation penalties from local flow field and environmental occlusion to the extracted temporal latent variables, and calculating the effective carbon atom adhesion rate of the target workpiece at that specific location.

Images