Chain roller continuous heat treatment intelligent control method and system

Abstract

The application belongs to the technical field of heat treatment control, and discloses a chain roller continuous heat treatment intelligent control method and system. The method collects online images of chain rollers, web belt speed, furnace temperature, furnace atmosphere, transfer rhythm and quenching medium state, identifies coverage, accumulated layer thickness, posture dispersion and inner hole exposure rate, generates a virtual batch, calculates equivalent furnace loading heat capacity index and quenching delay index, predicts surface hardening layer depth, core hardness, ovality risk and quenching risk, and accordingly implements cross-station collaborative compensation and backtracking update. The application can improve heat treatment quality consistency and reduce deformation and quenching risk.

Description

Technical Field

[0001] This invention belongs to the field of heat treatment control technology and discloses an intelligent control method and system for continuous heat treatment of chain rollers. Background Technology

[0002] Chain rollers are key wear-resistant components in roller chains, and their heat treatment quality directly affects the chain's wear resistance, fatigue resistance, and service life. In current production, chain rollers typically undergo continuous heating, quenching, and tempering processes to achieve the required surface hardening layer depth, core hardness, and dimensional stability. For controlling the continuous heat treatment process, existing technologies primarily focus on adjusting the overall furnace temperature, workpiece temperature profile, or conveyor cycle time. For example, prior art document CN1055317C discloses an online control method for a continuous annealing furnace, which establishes a dynamic mathematical model of the entire furnace temperature to control and compensate for the furnace temperature in each section online. Furthermore, other existing technologies often use a fixed furnace temperature gradient to control the workpiece's heating and cooling processes by adjusting the workpiece's movement speed or position within the heat treatment system. The aforementioned technical solutions have certain reference value in the field of continuous heat treatment control.

[0003] However, existing technologies primarily control the temperature process of the entire furnace, a section of strip, or a single workpiece. The control logic mainly revolves around furnace temperature distribution, workpiece movement rhythm, or temperature tracking, making it difficult to adapt to the unique operating conditions of chain roller continuous heat treatment, characterized by "discrete feeding, random stacking, posture fluctuations, and the coexistence of hollow internal hole structures." Specifically, the coverage rate, stacking thickness, placement posture, and internal hole exposure state of the chain rollers on the conveyor belt fluctuate with changes in the feeding state, further affecting heating uniformity, furnace atmosphere exchange, temperature drop after exiting the furnace, and the consistency of cooling during quenching. Existing technologies typically lack precise identification of the segmented passage states of such discrete small parts on a continuous line, lack technical means to correlate and track the feeding state with the heating state, furnace exit transport rhythm, and quenching state, and also lack a control mechanism for cross-station collaborative compensation for workpieces in the same segment.

[0004] Therefore, under the current technological conditions, the continuous heat treatment process of chain rollers is still prone to problems such as discrete surface hardening layer depth, core hardness fluctuation, excessive ellipticity, and increased risk of quenching cracks, and further improvements are still necessary. Summary of the Invention

[0005] The technical objective of this invention is to provide an intelligent control method and system for continuous heat treatment of chain rollers, so as to achieve coordinated and optimized control of each stage of heating, transportation, quenching and tempering in the continuous heat treatment process of chain rollers, improve the consistency of surface hardened layer depth, core hardness and dimensional stability, and reduce the risk of ellipticity deviation and quenching crack.

[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solutions.

[0007] In a first aspect, this invention discloses an intelligent control method for continuous heat treatment of chain rollers, comprising: S1, acquiring online images, specifications, conveyor belt speed, furnace temperature, furnace atmosphere, transfer cycle from furnace exit to quenching, quenching medium temperature, and stirring state of the chain rollers at the loading end; S2, identifying the coverage, deposit thickness, orientation dispersion, and inner hole exposure rate of the rollers on the conveyor belt based on the online images, and generating corresponding virtual batches by combining encoder displacement along the conveying direction; S3, calculating the equivalent furnace loading heat capacity index and quenching delay index for each virtual batch, and comparing them with real-time process parameters. The chain roller heat treatment quality prediction model is input together to obtain the predicted values ​​of surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk for the virtual batch; S4, when the predicted value deviates from the target window, the collaborative compensation amount for the heating zone temperature, furnace atmosphere setpoint, mesh belt speed, transfer cycle, quenching medium stirring intensity, and tempering time corresponding to the virtual batch is calculated and implemented when the virtual batch arrives at the corresponding station; S5, the online detection results after quenching and tempering are back-matched according to the virtual batch to generate model deviation and update the quality prediction model and the collaborative compensation amount.

[0008] Specifically, the online image is jointly acquired by an upper region camera and a lateral line laser profilometer; the coverage rate is the ratio of the roller projection area to the effective bearing area of ​​the conveyor belt; the stacking layer thickness is the average height of the roller layer in the direction perpendicular to the conveyor belt; the attitude dispersion is a discrete statistical measure in the direction of the roller axis; and the inner hole exposure rate is the ratio of the number of visible inner hole openings to the total number of rollers.

[0009] Specifically, the virtual batch is synchronously segmented based on the image timestamp and the displacement of the mesh belt encoder. Each virtual batch corresponds to a mesh belt segment of a preset length and inherits the same batch identifier in the furnace exit, quenching, and tempering sections to achieve cross-station tracking.

[0010] Specifically, the equivalent furnace heat capacity index is calculated by weighting the mass of a single roller, coverage, deposit thickness, orientation dispersion, and inner hole exposure rate, and is used to characterize the comprehensive impact of the virtual batch on heat absorption and atmosphere exchange in the heating zone; the quenching delay index is calculated by the time from furnace exit to quenching, transfer path length, quenching medium surface fluctuation, medium temperature, and stirring speed, and is used to characterize the degree of heat loss of the virtual batch before quenching.

[0011] Specifically, the chain roller heat treatment quality prediction model includes a heating response sub-model, a transport temperature drop sub-model, a quenching cooling sub-model, and a tempering recovery sub-model; the heating response sub-model outputs the furnace exit temperature uniformity index, and the quenching cooling sub-model outputs predicted values ​​for surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk.

[0012] Specifically, the online detection results include at least the infrared temperature field after quenching, eddy current detection signal, magnetic Barkhausen noise signal, and hardness or roundness data after tempering, and are mapped back to the corresponding virtual batch according to the timestamp and conveyor belt displacement.

[0013] Specifically, the collaborative compensation is implemented according to the following rules: when the predicted value of the surface hardened layer depth is lower than the target lower limit, the conveyor belt speed is reduced and / or the corresponding heating zone temperature or furnace atmosphere setting value is increased; when the ellipticity risk or quenching crack risk is higher than the threshold, the transfer cycle from furnace exit to quenching is shortened and the stirring gradient of the quenching medium is reduced; when the predicted value of the core hardness is lower than the target lower limit, the temperature of the front heating zone and the conveyor belt speed are adjusted first, and the tempering time is corrected a second time.

[0014] Specifically, the update adopts a recursive correction method with a forgetting factor. Based on the deviation between the online detection results of the virtual batch and the predicted value, the weight parameters of the quality prediction model and the sensitivity parameters of the collaborative compensation amount are updated, and smoothing constraints are set for the control quantities of the most recent virtual batches.

[0015] Secondly, the present invention also discloses an intelligent control system for continuous heat treatment of chain rollers, used to implement the method described in the first aspect, characterized in that it includes: a data acquisition module for acquiring online images, specifications, conveyor belt speed, furnace temperature of each heat treatment zone, furnace atmosphere, transfer cycle from furnace exit to quenching, and quenching medium status at the loading end; a virtual batch generation module for identifying coverage, deposit thickness, orientation dispersion, and inner hole exposure rate, and generating corresponding virtual batches; a quality prediction module for calculating the equivalent furnace loading heat capacity index and quenching delay index and outputting predicted values ​​for surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk; a collaborative compensation control module for generating and executing collaborative compensation amounts for heating zone temperature, furnace atmosphere setpoint, conveyor belt speed, transfer cycle, quenching medium stirring intensity, and tempering time; and a traceability correction module for retrospectively matching online detection results to virtual batches and updating the quality prediction module and the collaborative compensation control module.

[0016] Specifically, the data acquisition module includes an upper area camera, a lateral line laser profilometer, a mesh belt encoder, a furnace temperature sensor, a furnace atmosphere detection sensor, a quenching medium temperature sensor, a stirring speed sensor, and an infrared temperature measurement device; the traceability correction module is also communicatively connected to an eddy current detection device, a magnetic Barkhausen noise detection device, and a hardness / roundness sampling terminal.

[0017] This invention introduces a virtual batch tracking mechanism for chain rollers during continuous heat treatment. It combines characteristic parameters representing the passing state of discrete small parts, such as loading end coverage, deposit thickness, attitude dispersion, and inner hole exposure rate, to construct an equivalent furnace loading heat capacity index and a quenching delay index. Based on these, it predicts in real time the surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk. Furthermore, it implements cross-station collaborative compensation for heating zone temperature, furnace atmosphere, conveyor belt speed, transfer cycle time, quenching medium stirring intensity, and tempering time. This transforms traditional decentralized and lagging single-stage control into a closed-loop control of the entire process oriented towards the same virtual batch. Based on this mechanism, this invention effectively mitigates the adverse effects of random chain roller stacking, attitude fluctuations, and differences in heat loss from furnace exit to quenching on heat treatment quality. It significantly improves the consistency of surface hardened layer depth and core hardness, enhances dimensional stability, reduces ellipticity deviations and quenching crack risks, and improves the first-pass yield and process stability of continuous heat treatment production. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the intelligent control method for continuous heat treatment of chain rollers according to the present invention.

[0019] Figure 2 This is a structural block diagram of the intelligent control system for continuous heat treatment of chain rollers according to the present invention.

[0020] Figure 3 This is a schematic diagram illustrating the virtual batch generation and material loading status feature extraction of the present invention.

[0021] Figure 4 This is a schematic diagram of the heat treatment quality prediction and collaborative compensation control logic of the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the online detection result backtracking matching and model update of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0024] This invention is primarily applicable to continuous heat treatment production lines for chain rollers, and is particularly suitable for scenarios requiring real-time sensing, predictive control, and batch traceability during continuous heating, furnace transfer, quenching, and tempering of chain rollers. The continuous heat treatment is preferably a continuous heating-quenching-tempering process under a protective atmosphere, more preferably a continuous carburizing-quenching-tempering process, but is not limited to these.

[0025] I. Terminology Explanation

[0026] To enable those skilled in the art to accurately understand the technical solution of this invention, the main terms involved in this invention will be explained first.

[0027] Chain rollers are hollow cylindrical parts that are fitted onto the outside of the bushing in a chain assembly and bear contact stress and wear during meshing and rolling.

[0028] Virtual batch refers to defining a group of chain rollers that are in the same section of the conveyor belt and will pass through the same station in subsequent stages as a traceable control unit based on the image timestamp at the feeding end and the displacement of the conveyor belt encoder.

[0029] Coverage rate refers to the percentage of the two-dimensional projected area of ​​the chain rollers on the effective load-bearing area of ​​the mesh belt within a certain virtual batch.

[0030] The thickness of the stacked layer refers to the average stacking height of the chain rollers in the direction perpendicular to the conveyor belt within a certain virtual batch.

[0031] Attitude dispersion refers to the degree of dispersion in the direction, tilt angle, or placement direction of the roller axis of each chain within a virtual batch.

[0032] The inner hole exposure rate refers to the proportion of roller inner hole openings that can be directly identified by the vision system in a certain virtual batch to the total number of rollers in that batch.

[0033] The equivalent furnace heat capacity index refers to a dimensionless index that comprehensively maps the quantity, quality, coverage, deposit thickness, orientation dispersion, and inner hole exposure rate of chain rollers into a characterizing heat absorption and heat transfer behavior.

[0034] The quenching delay index is a dimensionless index that comprehensively maps factors such as the time from furnace exit to quenching, transport path, quenching medium temperature, stirring intensity, and transport fluctuations into a characterizing degree of heat loss before quenching.

[0035] The heat treatment quality prediction model refers to a model that predicts the surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk based on the geometric loading state, real-time process parameters, and environmental parameters of a virtual batch. Specifically, the surface hardened layer depth can be determined using a metallographic sectioning method with a hardness threshold of HV550 in a preferred embodiment; the core hardness can be expressed using Rockwell hardness HRC or Vickers hardness HV conversion in a preferred embodiment; and the ellipticity risk and quenching crack risk can be represented using binary classification or probability values.

[0036] The collaborative compensation amount refers to the adjustment value jointly given to the execution quantities such as heating zone temperature, furnace atmosphere setpoint, conveyor belt speed, transfer cycle, quenching medium stirring intensity, and tempering time in order to bring the predicted quality index of a virtual batch back to the target window.

[0037] II. System Structure

[0038] The system structure of the present invention is as follows: Figure 1 As shown, it mainly includes a feeding and detection unit, a continuous heating unit, a furnace exit and transfer unit, a quenching unit, a tempering unit, an online detection unit, an industrial control unit, and a data traceability unit.

[0039] 2.1 Material Feeding Detection Unit

[0040] Located at the feeding end of the conveyor belt, this device is used to collect online images and outline information of the chain rollers. Preferably, it includes an upper area camera, a lateral line laser profilometer, and a conveyor belt encoder. The upper area camera acquires a planar distribution image of the rollers on the conveyor belt; the lateral line laser profilometer acquires information on the roller stacking height; and the conveyor belt encoder provides conveyor belt displacement and speed data to establish a mapping relationship between image data and material position.

[0041] 2.2 A continuous heating unit is used for continuous heating of the chain rollers. Preferably, it is a multi-zone mesh belt furnace, more preferably including a preheating zone, a heating zone, a holding zone, and a diffusion zone. Each zone is equipped with a furnace temperature sensor and a furnace atmosphere detection sensor, preferably detecting oxygen potential, carbon potential, or protective gas flow rate. This unit is controlled by a PLC or DCS to regulate the heating power, fan speed, and atmosphere supply of each zone.

[0042] 2.3 The furnace exit transfer unit is located between the heating unit and the quenching unit, and is used to transport the chain rollers from the heating furnace outlet to the quenching tank inlet. This unit is equipped with position sensors, cycle time detectors, and transfer mechanical actuators to detect and control the time delay from furnace exit to quenching tank entry.

[0043] 2.4 The quenching unit is used to perform the quenching process. It is preferably an oil quenching tank or a polymer quenching tank, and is equipped with a medium temperature sensor, a liquid level sensor, a stirring speed sensor, and a circulating pump controller. The quenching unit can adjust the stirring intensity, medium temperature, and local flow field state, thereby affecting the cooling rate and deformation behavior of the chain rollers.

[0044] 2.5 The tempering unit is used to perform tempering treatment. A continuous tempering furnace is preferred, equipped with functions for temperature detection and holding time control in the tempering zone.

[0045] 2.6 The online inspection unit is used to perform online or quasi-online inspection of the chain rollers after quenching and tempering. Preferably, it includes an infrared thermal imager, an eddy current testing device, a magnetic Barkhausen noise detection device, and a roundness / ovality measurement device; in feasible embodiments, it also includes an offline sampling terminal for collecting the depth of the metallographic hardened layer and the core hardness results as training or calibration labels.

[0046] 2.7 The industrial control unit, as the core control component of the system, preferably includes an industrial computer, a PLC, an edge acquisition module, and a human-machine interface. The industrial computer deploys a virtual batch generation module, a heat treatment quality prediction model, a collaborative compensation control module, and a traceability correction module; the PLC is responsible for communicating with the furnace body, conveyor belt, transfer mechanism, quenching pump, and tempering furnace actuators and issuing control commands.

[0047] 2.8 The data traceability unit is used to map the results of material loading inspection, process control, online inspection, and sampling inspection to the same virtual batch and save them as traceable records. Preferably, it includes a database server and a production management interface.

[0048] In this invention, the innovation of the system structure does not lie in a single sensor or furnace structure itself, but in using the above structure to form a continuous tracking and collaborative closed-loop control link for virtual batches.

[0049] III. Specific Technical Route for Implementing the Method of the Invention

[0050] The method flow of the present invention is as follows: Figure 2 As shown, the method comprises five steps: data acquisition, virtual batch generation and feature extraction, quality prediction, collaborative compensation control, detection backtracking, and model updating. The specific implementation of the method of this invention is described in detail below.

[0051] 3.1 S1 Chain Roller Data Acquisition and Basic Classification

[0052] Preferably, online images of the chain rollers are collected at the feeding end of the mesh belt, and specifications, mesh belt speed, furnace temperature of each heat treatment zone, furnace atmosphere, transfer cycle from furnace exit to quenching, quenching medium temperature and stirring state are collected simultaneously.

[0053] The specifications include at least the outer diameter, inner diameter, length, unit weight, and material grade of the roller.

[0054] The furnace temperature of each heat treatment zone is preferably sampled at a period of 1s to 5s; the online image is preferably acquired at 5fps to 15fps; and the mesh belt encoder is preferably sampled at a frequency of not less than 500Hz for displacement pulses.

[0055] In this embodiment, traditional machine vision methods are preferred for image processing rather than deep learning image models. This is because the background of the chain roller at the feeding end is fixed, the contrast is high, and the target shape is clear. Background segmentation, edge extraction, contour fitting, and template matching can meet the implementation requirements, thereby simplifying deployment and reducing the dependence on training samples.

[0056] 3.2 S2 Virtual Batch Generation and Feature Index Construction

[0057] 3.2.1 Extraction of Coverage, Stack Thickness, Orientation Dispersion, and Pore Exposure Rate

[0058] For the image obtained by the camera in the upper region, background subtraction and grayscale normalization are performed first, and then the foreground region of the chain rollers is extracted through threshold segmentation to obtain the roller projection region. For example... Figure 3 As shown, the feeding detection unit obtains the planar distribution information, stacking height information and displacement information of the chain rollers through the area camera, the lateral contour detection device and the mesh belt encoder, respectively, and extracts the coverage, stacking layer thickness, attitude dispersion and inner hole exposure rate accordingly, and then generates the corresponding virtual batch.

[0059] Coverage It can be calculated using the following formula:

[0060]

[0061] in, This represents the total area of ​​the roller projection in the current image. This represents the area of ​​the effective carrying area of ​​the mesh belt at the corresponding moment.

[0062] For the profile curve obtained by the lateral laser profilometer, after removing the mesh reference plane, the average height of each sampling point in the current section is calculated to obtain the thickness of the deposited layer. :

[0063]

[0064] in, The number of height sampling points, For the first The height value of each sampling point. This is the reference height for the mesh belt.

[0065] For roller orientation, the axial direction angle of each roller is preferably obtained by fitting the contour ellipse or fitting the cylindrical boundary direction. Posture dispersion It can be represented as:

[0066]

[0067] in, The number of rollers identified in the current segment. This is the average direction angle.

[0068] Inner hole exposure rate It can be calculated using the following formula:

[0069]

[0070] in, To identify the number of rollers with inner hole openings in the image. This represents the total number of rollers in the current section.

[0071] Of the four parameters mentioned above, coverage and deposit thickness mainly reflect the heat capacity load and the relationship between heat shielding; attitude dispersion and inner hole exposure mainly reflect the uniformity of heating, gas exchange in the furnace, and uniformity of quenching. For hollow small parts such as chain rollers, the latter two are particularly critical, which is also an important basis for the present invention to distinguish it from conventional batch furnace control or single-piece temperature measurement control.

[0072] 3.2.2 Generation of Virtual Batches

[0073] This invention does not use the traditional whole basket, whole furnace, or whole shift as the control object, but instead uses virtual batches bound to a section of the conveyor belt as the control object. Let the conveyor belt at time... The cumulative displacement is The virtual batch segment length is Then the virtual batch number It can be defined as:

[0074]

[0075] in, This indicates rounding down to the nearest integer. Preferably, The diameter is 120mm to 250mm; in this embodiment, 180mm is preferred.

[0076] In this way, the same virtual batch maintains the same batch number when passing through the heating zone, furnace transfer, quenching and tempering, thereby achieving continuous tracking across workstations.

[0077] 3.2.3 Equivalent Furnace Loading Heat Capacity Index and Quenching Delay Index

[0078] After obtaining the virtual batch and its image features, an equivalent furnace loading heat capacity index is constructed. The index is preferably determined by the following formula:

[0079]

[0080] in, For the mass of a single roller, This represents the current virtual batch roller quantity. As the benchmark quality item, , , and These are the baseline values ​​for coverage, deposit thickness, orientation dispersion, and pore exposure, respectively. to These are the weighting coefficients.

[0081] Preferred selection , , , , .

[0082] Simultaneously, construct the quenching delay index. The preferred method is to calculate using the following formula:

[0083]

[0084] in, The time from when the furnace is removed to when it is quenched. This is the length of the transit route. The quenching medium temperature, The reference temperature for the quenching medium. The stirring speed during quenching is [specified]. The reference stirring speed is... This refers to the fluctuation in the transport cycle time. , , , and This is the corresponding normalized reference quantity; to These are the weighting coefficients.

[0085] Preferred selection , , , , .

[0086] in, It reflects the combined impact of the furnace loading condition on heat absorption and heat exchange behavior. This reflects the combined effects of heat loss from the furnace to the quenching stage and fluctuations before cooling. These two factors are key bridging variables in the control logic of this invention.

[0087] 3.3 Construction, Training, and Online Prediction of the S3 Heat Treatment Quality Prediction Model

[0088] 3.3.1 Model Input and Output

[0089] In this embodiment, the input feature vector of the heat treatment quality prediction model is: It includes four main categories of characteristics.

[0090] Table 1. Model Input Feature Composition

[0091]

[0092] The model output includes four items:

[0093]

[0094] in, This is the predicted depth of the surface hardened layer. This is a predicted value for heart stiffness. This is the predicted value for ellipticity risk. This represents the predicted risk value for quenching cracks.

[0095] 3.3.2 Model Structure

[0096] Preferably, the heat treatment quality prediction model adopts a multi-task feedforward neural network structure, including a shared backbone network, a regression branch, and a classification branch.

[0097] The preferred shared backbone network architecture is:

[0098] Input layer (28-dimensional) — Batch normalization layer — Fully connected layer FC1 (64 nodes) — ReLU activation — Dropout (0.10) — Fully connected layer FC2 (64 nodes) — ReLU activation.

[0099] The preferred regression branch is:

[0100] FC3 (32 nodes) — ReLU — Output layer (2 nodes), corresponding to and .

[0101] The preferred classification branch is:

[0102] FC4 (32 nodes) — ReLU — Sigmoid output layer (2 nodes), corresponding to and .

[0103] Among them, the regression branch is used to predict continuous quality indicators, and the classification branch is used to predict risk indicators.

[0104] 3.3.3 Generation of Training Labels

[0105] The preferred methods for obtaining training labels are as follows:

[0106] (1) Surface hardened layer depth label: Metallographic sections and microhardness measurements are performed on chain roller samples obtained according to preset sampling rules, and the position corresponding to HV550 is taken as the surface hardened layer depth.

[0107] (2) Core hardness label: The Rockwell hardness (HRC) of the core is measured for the same sample.

[0108] (3) Ellipticity risk label: Roundness or ellipticity is detected for the rollers sampled in the corresponding virtual batch. When the error rate is greater than the set threshold, it is recorded as 1; otherwise, it is recorded as 0.

[0109] (4) Quenching crack risk label: Based on the results of online eddy current detection, magnetic Barkhausen noise detection and offline crack verification, when a cracked part appears in the virtual batch, it is recorded as 1, otherwise it is recorded as 0.

[0110] 3.3.4 Dataset Composition

[0111] In this implementation, production records from the same continuous heat treatment production line over 12 weeks were selected to form 21,600 virtual batch samples. Among them, 6,480 samples were sampled and matched to obtain complete supervisory labels, while the remaining samples were used for online incremental correction and updating of unlabeled statistical features. In the fully labeled samples, the training set, validation set, and test set were divided at 70%:15%:15%.

[0112] Table 2 Dataset Composition

[0113]

[0114] 3.3.5 Data Preprocessing and Training Methods

[0115] Z-score normalization is used for continuous input features, and one-hot encoding or flag bit encoding is used for categorical or discrete process state features.

[0116] The optimizer is preferably Adam, the initial learning rate is preferably 0.001, the batch size is preferably 128, the maximum number of training epochs is preferably 80, and early stopping is performed when there is no improvement on the validation set for 10 consecutive epochs.

[0117] The preferred loss function for the model is as follows:

[0118]

[0119] in, This represents the true depth of the surface hardened layer. This represents the true value of heart hardness. For ellipticity risk, To provide a true label for the risk of quenching cracks. For model parameters, to This is the loss weight.

[0120] Preferred selection , , , , .

[0121] 3.3.6 Online Prediction

[0122] During production line operation, whenever a virtual batch enters a critical control point, such as before entering the heating zone, before entering the holding zone, before exiting the furnace, or before entering the quenching zone, the system reads the current feature vector of the virtual batch and performs forward inference to obtain four quality prediction results. Preferably, the online inference cycle is 5s to 20s; in this embodiment, it is preferable that each virtual batch triggers one inference when it arrives at the corresponding workstation.

[0123] 3.4, S4 Collaborative Compensation Control

[0124] Unlike traditional methods that only adjust furnace temperature or conveyor belt speed, this invention provides cross-station collaborative compensation for the predicted quality results of the same virtual batch. For example... Figure 4 As shown, this invention inputs the equivalent furnace loading heat capacity index and quenching delay index corresponding to the virtual batch into the heat treatment quality prediction model to obtain the predicted values ​​of surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk. Based on the prediction results, it generates a collaborative compensation amount for the heating zone temperature, furnace atmosphere setpoint, mesh belt speed, conveyor cycle, quenching medium stirring intensity, and tempering time.

[0125] Let the control vector be:

[0126]

[0127] in, , , These are the temperature settings for the corresponding heating zones. , Set the furnace atmosphere value. For the speed of the conveyor belt, The transfer cycle from furnace exit to quenching, For the stirring intensity of the quenching medium, This refers to the tempering time.

[0128] Collaborative compensation amount The objective function is obtained as follows:

[0129]

[0130] in, , , and The first The four prediction results for each virtual batch and These refer to the depth of the target's surface hardened layer and the hardness of the target's core, respectively. This is the control increment that was executed in the previous virtual batch. , , , , and These are the weighting coefficients.

[0131] Preferred selection , , , , , .

[0132] The preferred control constraints are as follows:

[0133]

[0134]

[0135]

[0136]

[0137]

[0138] The upper and lower limits can be set according to the rated capacity of the equipment and the process safety window.

[0139] Preferably, the present invention employs a two-stage solution strategy.

[0140] The first stage involves quickly filtering and controlling the direction based on a rule base. For example:

[0141] when When the speed is below the target lower limit, the conveyor belt speed should be reduced first and the temperature or atmosphere setting value of the middle and rear heating zone should be increased appropriately.

[0142] when or When the value is above the threshold, the transfer cycle time should be shortened and the quenching stirring intensity gradient should be reduced.

[0143] when When the temperature is too low, prioritize coordinating the temperature of the front heating zone and the speed of the conveyor belt, and finely adjust the tempering time.

[0144] In the second stage, within the selected feasible control subspace, the optimal solution is obtained using either sequential quadratic programming or constrained gradient descent. .

[0145] Through the above-mentioned coordinated control, the present invention can avoid the local optimality problems that occur in traditional control, such as "deformation increases after increasing furnace temperature" or "hardened layer fluctuates after shortening quenching delay".

[0146] 3.5 S5 Online Result Backtracking Matching and Model Update

[0147] The fifth step of this invention is to trace the online detection results back to the corresponding virtual batch and use them for model bias correction and control sensitivity updates. For example... Figure 5 As shown, the online detection results after quenching and the online detection results after tempering are mapped to the corresponding virtual batches according to the timestamp and the conveyor belt displacement. After completing the backtracking matching, the model deviation is calculated, and the heat treatment quality prediction model and the collaborative compensation amount are updated accordingly to form a closed-loop optimization control for the next virtual batch.

[0148] In this embodiment, the infrared thermographic data, eddy current data, and magnetic Barkhausen noise data after quenching are all timestamped; since each virtual batch has been mapped to the conveyor belt displacement, matching can be performed as follows:

[0149]

[0150] in, This is the virtual batch number corresponding to the testing station. To detect the cumulative displacement of the conveyor belt at any given time, For reference initial displacement, This represents the length of the virtual batch segment.

[0151] After obtaining the backtracking matching results, the model prediction bias is calculated. Taking the depth of the surface hardened layer as an example, the bias can be expressed as:

[0152]

[0153] in, For the first The measured depth of the surface hardened layer in the virtual batch. This is the corresponding predicted value.

[0154] Similarly, the heart stiffness deviation and risk classification deviation can be obtained. Preferably, a recursive update method with a forgetting factor is used to update the model parameters and control sensitivity parameters:

[0155]

[0156] in, For parameters before the update, The new parameters are calculated based on the current deviation. This is for updating coefficients.

[0157] Preferred selection ~ More preferably, it is 0.05.

[0158] Meanwhile, to avoid excessive fluctuations in the control values ​​of consecutive virtual batches, a sliding smoothing process is implemented for the control increments of the most recent five virtual batches to ensure that they meet the process continuity requirements. Specific Implementation

[0159] 4.1 Detection Method

[0160] The surface hardening layer depth was determined by randomly sampling from the corresponding virtual batch, preparing metallographic specimens, and measuring Vickers hardness point by point along the cross section from the surface to the center. When the hardness decreased to the depth corresponding to HV550, it was defined as the surface hardening layer depth.

[0161] Core hardness determination involves measuring the Rockwell hardness (HRC) of the central region of the cross-section.

[0162] Ellipticity is determined by measuring the difference between the maximum and minimum diameters of the outer circle of the roller using a roundness meter or a high-precision pneumatic gauge. This difference is recorded as ellipticity.

[0163] The crack rate was calculated using online eddy current testing for 100% detection. Suspected cracked parts were then verified using magnetic particle testing or cross-sectioning. The crack rate was calculated as the ratio of cracked parts to the total number of parts.

[0164] The energy consumption statistics include the total power consumption of the heating furnace, transfer mechanism, quenching circulation system and tempering furnace, and convert it into energy consumption per 10,000 pieces.

[0165] Experimental setup: 3000 chain rollers were selected for both the example and comparative examples, and the experiment was repeated 3 times, with 1000 pieces each time. All workpieces were inspected online, and 30 pieces were randomly selected for surface hardening layer depth and core hardness testing, and 60 pieces were selected for roundness / ovality precision measurement.

[0166] 4.2, Example 1

[0167] 4.2.1 Production line and workpiece conditions

[0168] This embodiment uses a continuous heat treatment production line for chain rollers for verification. The chain rollers being treated are low-carbon alloy carburized steel rollers with an outer diameter of 10.20 mm, an inner diameter of 6.20 mm, a length of 8.40 mm, and a single piece weight of approximately 2.6 g. The process target window is as follows:

[0169] Table 3 Process Target Window of Example 1

[0170]

[0171] The production line is configured as follows: the upper area camera has a resolution of 4096×3000 and a sampling frequency of 10fps; the lateral line laser profilometer has a sampling frequency of 200Hz; the mesh belt encoder has a resolution of 2048 pulses / revolution; the heating furnace is divided into 3 main control hot zones; the quenching tank is an oil quenching tank; and the tempering furnace is a continuous tempering furnace.

[0172] 4.2.2 Data Acquisition and Model Training

[0173] Production data was collected continuously for 12 weeks, resulting in 21,600 virtual batch samples. One batch out of every 10 virtual batches was selected for offline metallographic and hardness testing, and one batch out of every 5 virtual batches was selected for roundness testing. At the same time, all virtual batches underwent online eddy current and magnetic Barkhausen noise testing.

[0174] The model training parameters are as follows:

[0175] Table 4 Model Training Parameters

[0176]

[0177] After training, the prediction performance on the test set is as follows.

[0178] Table 5 Model Test Performance

[0179]

[0180] The above results show that the heat treatment quality prediction model constructed in this invention can accurately characterize the quality trend of chain rollers during continuous heat treatment, providing a reliable basis for subsequent collaborative compensation control.

[0181] 4.2.3 Online Control Process

[0182] During actual operation, the length of each virtual batch is set to 180mm. The system triggers inference and control decisions before the virtual batch enters the heating zone, the holding zone, before exiting the furnace, and before entering the quenching zone.

[0183] When the system detects that the coverage of a virtual batch increases, the thickness of the deposit layer increases, and the exposure rate of the inner hole decreases, the equivalent furnace loading heat capacity index... Increase; if an extension of the cycle time from furnace exit to quenching is detected simultaneously, then the quenching delay index increases. Increase. At this point, the model will typically give a joint prediction result that is too low in the depth of the surface hardened layer and too high in the risk of ellipticity. Based on this, the system will implement the following compensation:

[0184] The temperature of the middle and rear heating zone is increased by 4℃ to 8℃;

[0185] The conveyor belt speed decreased by 1.5% to 3.0%;

[0186] The cycle time for unloading and transferring materials is shortened by 0.6s to 1.2s;

[0187] The strength decreases by 3% to 6% after quenching and stirring.

[0188] The corresponding tempering time is extended by 20s to 40s.

[0189] The aforementioned collaborative compensation amounts are all implemented within the equipment's allowed window and only take effect when the corresponding virtual batch arrives at the relevant workstation.

[0190] 4.3 Scale settings

[0191] To verify the technical effect of the present invention, the following comparative examples are set up.

[0192] 4.3.1 Comparative Example 1: Fixed Formula Control Group

[0193] This group adopts a conventional continuous heat treatment control method, which controls the temperature of each hot zone, the speed of the conveyor belt, and the intensity of quenching and stirring only according to the preset process formula, without making dynamic adjustments based on the feeding status, virtual batches, or online predictions.

[0194] 4.3.2 Comparative Example 2: Post-test Hardness Feedback Group

[0195] This group uses the conventional "offline sampling inspection - next batch correction" method. That is, the depth of the metallographic hardened layer and the core hardness are sampled every 2 hours. If they deviate from the target, the furnace temperature or conveyor belt speed is manually adjusted. However, no virtual batch is established, and the quenching delay and loading posture are not considered.

[0196] 4.3.3 Comparative Example 3: Image Feature Groups Without Virtual Batch

[0197] This group uses predictive model control, but the model input does not include coverage, deposit thickness, attitude dispersion, internal hole exposure rate, equivalent furnace heat capacity index and quenching delay index. It only uses conventional features such as specification parameters, furnace temperature, mesh belt speed and quenching medium state.

[0198] 4.3.4 Comparison of Results

[0199] The production results of Example 1 and the comparative example are compared, and the final statistical results are shown in Table 6.

[0200] Table 6 Comparison of production results between Example 1 and the comparative example.

[0201]

[0202] As shown in Table 6, the embodiments of the present invention outperform the comparative examples in terms of surface hardening layer depth dispersion, core hardness stability, ellipticity deviation rate, quenching crack rate, and first-pass yield. In particular, compared with Comparative Example 3, even using the same predictive model control, its quality consistency and risk control capabilities are significantly reduced. This indicates that the inventiveness of the present invention does not simply stem from the predictive model, but rather from the introduction of a dedicated state variable for the continuous chain roller line and its combination with cross-station control.

[0203] To further demonstrate the effectiveness of the early warning system, the average advance warning time before the occurrence of ellipticity deviation or quenching cracks in each group was statistically analyzed, and the results are shown in Table 7.

[0204] Table 7 Comparison of Risk Warning Capabilities

[0205]

[0206] As shown in Table 7, the present invention can provide a longer lead time for early warning before the actual occurrence of risks, thereby allowing sufficient time for collaborative compensation to be implemented.

[0207] 4.4, Example 2

[0208] To verify the applicability of this invention to chain rollers of different specifications, another specification of roller was selected for verification. This roller has an outer diameter of 13.50 mm, an inner diameter of 8.00 mm, a length of 11.00 mm, a target surface hardening layer depth of 0.60 ± 0.05 mm, and a target core hardness of 40 ± 3 HRC. The production quantity remained at 3000 chain rollers, and the test was repeated three times, with 1000 rollers produced each time.

[0209] Except for the target window and specification parameters, the system structure, virtual batch generation method, model structure and control method are basically the same as those in Example 1. Only the specification category identifier is added to the model, and the model parameters are retrained or fine-tuned.

[0210] Table 8 Comparison of production results between Example 2 and Comparative Example 1

[0211]

[0212] As can be seen from Table 8, the present invention has good applicability to chain rollers of different sizes.

[0213] 4.5 Overview of Technical Effects

[0214] Based on the above embodiments, it can be seen that:

[0215] First, this invention extracts coverage, stack thickness, attitude dispersion, and internal hole exposure rate at the feeding end, and constructs a virtual batch through conveyor belt displacement. This transforms the controlled object from the traditional whole furnace or sampled batches into a continuous control unit corresponding to the actual passing state of the chain rollers. This method is particularly suitable for the random stacking and attitude fluctuation of hollow small parts such as chain rollers on the conveyor belt, and can significantly improve the resolution of state perception.

[0216] Secondly, by introducing an equivalent furnace loading heat capacity index and a quenching delay index, this invention quantifies the differences in heat absorption during the heating stage and the differences in heat loss from furnace exit to quenching stage of the chain rollers, enabling subsequent quality prediction and control compensation to directly address the physical root causes of quality fluctuations.

[0217] Third, this invention uses a heat treatment quality prediction model to perform multi-task fusion prediction of the loading characteristics of the chain rollers, real-time process parameters and historical feedback, and simultaneously outputs the surface hardened layer depth, core hardness, ellipticity risk and quenching crack risk, thus avoiding the deviation transfer problem caused by traditional single index control.

[0218] Fourth, this invention, through cross-station collaborative compensation, simultaneously adjusts the heating zone temperature, furnace atmosphere setpoint, conveyor belt speed, transfer cycle, quenching medium stirring intensity, and tempering time, extending the control from single-variable correction to full-process linkage correction. This effectively suppresses ellipticity deviation and quenching cracks while ensuring the depth of the hardened layer and the core hardness.

[0219] Fifth, this invention forms an adaptive closed-loop control mechanism for continuous heat treatment production lines by back-matching online detection results according to virtual batches and updating model parameters and control sensitivity parameters accordingly. This enables the system to maintain stable quality output even when raw material fluctuations, loading status changes, and environmental disturbances occur.

[0220] This invention focuses on the continuous heat treatment process of chain rollers. Through a technical approach of "image recognition at the feeding end, virtual batch generation, construction of a dedicated index, prediction of heat treatment quality, cross-station collaborative compensation, and result backtracking and updating," it solves the problem of difficulty in simultaneously considering the depth of the surface hardened layer, core hardness, ellipticity, and quenching crack risk in existing continuous heat treatment control. Furthermore, it can improve the consistency of heat treatment and the first-pass yield of chain rollers without increasing the frequency of complex offline inspections.

[0221] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novelty disclosed herein.

Claims

1. A chain roller continuous heat treatment intelligent control method, characterized in that, include: S1. Collect online images, specifications, belt speed, furnace temperature, furnace atmosphere, transfer rhythm from furnace exit to quenching, quenching medium temperature, and stirring status of the chain rollers at the feeding end. S2. Based on the online image recognition of the roller's coverage, deposit thickness, posture dispersion, and inner hole exposure rate on the conveyor belt, and combined with the encoder displacement along the conveying direction, generate the corresponding virtual batch. S3. Calculate the equivalent furnace loading heat capacity index and quenching delay index for each virtual batch, and input them together with the real-time process parameters into the chain roller heat treatment quality prediction model to obtain the predicted values ​​of surface hardening layer depth, core hardness, ellipticity risk and quenching crack risk for that virtual batch. S4. When the predicted value deviates from the target window, calculate the collaborative compensation amount for the heating zone temperature, furnace atmosphere setpoint, mesh belt speed, conveyor cycle time, quenching medium stirring intensity and tempering time corresponding to the virtual batch, and implement it when the virtual batch arrives at the corresponding station. S5. The online detection results after quenching and tempering are back-matched according to the virtual batch to generate model deviation and update the quality prediction model and the collaborative compensation amount.

2. The chain roller continuous heat treatment intelligent control method according to claim 1, characterized in that, The online images are jointly acquired by an upper region camera and a lateral line laser profilometer; the coverage rate is the ratio of the roller projection area to the effective bearing area of ​​the conveyor belt; the stacking layer thickness is the average height of the roller layer in the direction perpendicular to the conveyor belt; the attitude dispersion is a discrete statistical measure in the direction of the roller axis; and the inner hole exposure rate is the ratio of the number of visible inner hole openings to the total number of rollers.

3. The chain roller continuous heat treatment intelligent control method according to claim 1, characterized in that, The virtual batch is synchronously segmented based on the image timestamp and the displacement of the mesh belt encoder. Each virtual batch corresponds to a mesh belt segment of a preset length and inherits the same batch identifier in the furnace exit, quenching, and tempering sections to achieve cross-station tracking.

4. The chain roller continuous heat treatment intelligent control method according to claim 1, characterized in that, The equivalent furnace heat capacity index is calculated by weighting the mass of a single roller, coverage, deposit thickness, orientation dispersion, and inner hole exposure rate, and is used to characterize the comprehensive impact of the virtual batch on heat absorption and atmosphere exchange in the heating zone; the quenching delay index is calculated by the time from furnace exit to quenching, transfer path length, quenching medium surface fluctuation, medium temperature, and stirring speed, and is used to characterize the degree of heat loss of the virtual batch before quenching.

5. The chain roller continuous heat treatment intelligent control method according to claim 1, characterized in that, The chain roller heat treatment quality prediction model includes a heating response sub-model, a transport temperature drop sub-model, a quenching cooling sub-model, and a tempering recovery sub-model. The heating response sub-model outputs the furnace exit temperature uniformity index, and the quenching cooling sub-model outputs predicted values ​​for the surface hardened layer depth, core hardness, ellipticity risk, and quenching crack risk.

6. The chain roller continuous heat treatment intelligent control method according to claim 1, wherein, The online detection results include at least the infrared temperature field after quenching, eddy current detection signal, magnetic Barkhausen noise signal, and hardness or roundness data after tempering, and are mapped back to the corresponding virtual batch according to the timestamp and conveyor belt displacement.

7. The chain roller continuous heat treatment intelligent control method according to claim 1, characterized in that, The coordinated compensation is implemented according to the following rules: when the predicted value of the surface hardened layer depth is lower than the target lower limit, the conveyor belt speed is reduced and / or the corresponding heating zone temperature or furnace atmosphere setting value is increased; when the ellipticity risk or quenching crack risk is higher than the threshold, the transfer cycle from furnace exit to quenching is shortened and the stirring gradient of the quenching medium is reduced; when the predicted value of the core hardness is lower than the target lower limit, the temperature of the front heating zone and the conveyor belt speed are adjusted first, and the tempering time is corrected a second time.

8. The chain roller continuous heat treatment intelligent control method according to claim 1, wherein, The update adopts a recursive correction method with a forgetting factor. Based on the deviation between the online detection results of the virtual batch and the predicted value, the weight parameters of the quality prediction model and the sensitivity parameters of the collaborative compensation amount are updated, and smoothing constraints are set for the control quantities of the most recent virtual batches.

9. A chain roller continuous heat treatment intelligent control system for implementing the method according to any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect online images, specifications, conveyor belt speed, furnace temperature of each heat treatment zone, furnace atmosphere, transfer cycle from furnace exit to quenching, and quenching medium status at the feeding end. The virtual batch generation module is used to identify coverage, stacking layer thickness, orientation dispersion and internal hole exposure rate, and generate corresponding virtual batches. The quality prediction module is used to calculate the equivalent furnace loading heat capacity index and the quenching delay index, and output the predicted values ​​of surface hardened layer depth, core hardness, ellipticity risk and quenching crack risk. The collaborative compensation control module is used to generate and execute collaborative compensation amounts for heating zone temperature, furnace atmosphere setpoint, conveyor belt speed, conveyor cycle time, quenching medium stirring intensity, and tempering time. The traceability correction module is used to backtrack and match the online detection results to the virtual batch and update the quality prediction module and the collaborative compensation control module.

10. The intelligent control system for continuous heat treatment of chain rollers according to claim 9, wherein, The data acquisition module includes an upper area camera, a lateral line laser profilometer, a mesh belt encoder, a furnace temperature sensor, a furnace atmosphere detection sensor, a quenching medium temperature sensor, a stirring speed sensor, and an infrared temperature measurement device; the traceability and correction module is also connected to an eddy current detection device, a magnetic Barkhausen noise detection device, and a hardness / roundness sampling terminal.

Images