Method for optimizing carburizing and quenching process parameters based on machine learning

Abstract

The application discloses a carburizing and quenching process parameter optimization method based on machine learning, relates to the technical field of carburizing and quenching process optimization, and comprises the following steps: setting a multi-source sensor, collecting environment and process data in real time, and forming a high-quality synchronous data set with a unified time stamp and a process label; using the preprocessed data, constructing an environment interference model through deep learning or time series regression, and quantifying the influence of external disturbance on the depth and hardness of the carburized layer; according to the model prediction result, a self-adaptive controller adjusts the carburizing agent flow, heating power and holding time in real time, realizes closed-loop regulation and control, ensures that the process conditions and product performance are always close to the preset target, detects the carburized layer and hardness of the parts discharged from the furnace, feeds the actual result back to the model for iterative updating, continuously optimizes the control strategy, effectively solves the unstable quality problem caused by environmental fluctuations in the traditional process, significantly improves the production efficiency, and reduces the energy consumption and test cost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carburizing quenching process optimization, in particular to a carburizing quenching process parameter optimization method based on machine learning. BACKGROUND

[0002] In the field of high-end equipment manufacturing and automobile transmission, gears, bearings and various key components widely adopt carburizing quenching process to balance high hardness and good toughness, which plays a decisive role in the service life and reliability of the whole system. In recent years, with the continuous improvement of industrial production line automation, production workshops are gradually integrating online monitoring, intelligent scheduling and other functions, hoping to realize real-time collection of temperature, humidity, air pressure and carbon potential in the furnace and other multi-source environment and process parameters, and then use data analysis means for precise control and quality tracing. However, the actual production environment often cannot maintain ideal conditions for a long time: the temperature and humidity of the workshop will fluctuate greatly with the seasons and circadian rhythms; when the air pressure drops suddenly or the sealing performance of the equipment decays over time, the carbon potential and temperature distribution in the furnace will also be unstable; for small-batch parts with complex shape and high surface quality requirements, flexible control strategies are needed to cope with uncertainties such as part stacking method and tool heat dissipation difference.

[0003] The traditional "experience-based" or "fuzzy interpolation-based" parameter adjustment method usually only relies on past accumulated process curves, and makes small adjustments within the static process window, which is difficult to keep up with the dynamic and changing production rhythm and high-precision performance requirements. At the same time, there is a nonlinear and strong coupling relationship between the carburizing layer uniformity, hardness gradient and quenching deformation, and simply extending the holding time or increasing the heating power often leads to rising energy consumption and deformation risk, which cannot meet the differentiated requirements of batch and small-batch high-end parts.

[0004] However, the current process design and control still cannot completely get rid of the dependence on experience or semi-experience methods, and the core problem lies in the lack of adaptive control means for the coupling effect of multi-source environmental data and process parameters. Specifically, when the external temperature or humidity of the workshop changes rapidly, the heat balance and carbon potential distribution inside the furnace body sealed will produce dynamic deviation, resulting in inconsistency between the target carburizing layer depth and the actual formed carburizing layer; if the initial temperature of the tooling or parts is low, the surface heat absorption difference is large, then the same process curve will produce obvious hardness distribution deviation in different batches. With the accumulation of random external disturbances, in extreme cases, it may cause insufficient carburizing, substandard hardness after quenching or local overburning, etc., which not only increases the scrap rate but also affects the overall line capacity; for high-value parts, the economic loss and delay of delivery caused by such fluctuations are more obvious.

[0005] The traditional static PID control can only make gentle correction in a fixed range, which is still insufficient to quickly respond to process deviations under the influence of multiple variable couplings, and it is difficult to automatically identify and correct the influence of sudden environmental factors on the growth of the carburized layer and the surface structure transformation process. Once the target and the actual result deviate continuously, manual intervention is required for temporary heating or time delay operation, thereby increasing energy consumption and management costs. Therefore, under the influence of the rapidly changing external environment and the complex coupling process requirements, if there is a lack of intelligent, dynamically learning and regulating process optimization system, it is difficult to continuously maintain high-quality consistent carburized layer and accurate hardness distribution in actual production. The key technical problem to be solved by the present application is to solve the above problems.

[0006] Therefore, the present application provides a carburizing and quenching process parameter optimization method based on machine learning. SUMMARY

[0007] (1) Technical problems solved

[0008] In view of the deficiencies in the prior art, the present application provides a carburizing and quenching process parameter optimization method based on machine learning, which collects environmental and process data in real time, and forms a high-quality synchronous data set with a unified time stamp and a process label. Using the preprocessed data, an environmental disturbance model is constructed by deep learning or time series regression to quantify the influence of external disturbances on the depth and hardness of the carburized layer. According to the model prediction results, a self-adaptive controller adjusts the carburizing agent flow, heating power and holding time in real time to realize closed-loop control and ensure that the process conditions and product performance are always close to the preset target. Through detection of the carburized layer and hardness of the parts after being taken out of the furnace, the actual results are fed back to the model for iterative updating to continuously optimize the control strategy, significantly improving production efficiency and reducing energy consumption and test cost, thereby solving the technical problems described in the background art.

[0009] (2) Technical solutions

[0010] To achieve the above purpose, the present application is realized by the following technical solutions:

[0011] The carburizing and quenching process parameter optimization method based on machine learning comprises: when a new batch of parts is detected to be put into the furnace, a sensor array arranged in the workshop and inside the furnace is called and the operation time stamp and batch identification are recorded synchronously. After performing time series alignment and outlier rejection on the collected data, a synchronous labeled data set containing multi-dimensional environmental elements is generated.

[0012] The temperature field and carbon potential variation characteristics in the furnace are fused to extract key feature vectors that affect the depth and hardness of the carburized layer, and the environmental disturbance prediction value is output through a prediction function.

[0013] When the predicted value of environmental disturbance is significantly different from the preset target, the adaptive controller calculates the optimal process control parameters according to the comprehensive deviation function, and adjusts the flow of carburant, heating power and holding time in real time to ensure the uniformity of carburizing and the stability of part performance.

[0014] When the parts are discharged and the hardness and carburizing layer depth detection are completed, the measured values, optimal process control parameters and predicted values of environmental disturbance are stored in the output data set, and incremental training is performed on the environmental disturbance model and adaptive controller to gradually improve the process stability and precision in long-term variable environment.

[0015] Preferably, a sensor array is arranged in each area of the carburizing and quenching workshop and inside the furnace body.

[0016] Each sensor periodically outputs a set of raw measurement values of multi-source environmental data, which is uniformly sent to the data acquisition module, and the data is archived according to the sensor label k to form the raw environmental data set D raw and preliminary quality check and pretreatment are performed.

[0017] Preferably, in order to realize time alignment of the outputs of each sensor, a unified reference clock t * is established, and each piece of raw data is mapped to the corresponding reference clock to form a synchronous data record.

[0018] At key production nodes, operation labels L op are respectively given, and the corresponding time t op is bound. sync The operation process and environmental sensing data are associated to obtain a synchronous marked data set D env ; an environmental reference index I * is introduced to simplify and summarize the overall level of multi-dimensional environmental parameters at a certain time.

[0019] Preferably, from the synchronous marked data set D sync , multi-source environmental data with time stamp t * are obtained, wherein each record carries its operation label L op ; according to the operation label, data segments corresponding to different process stages of the same batch of parts are identified and grouped.

[0020] An abnormality rejection algorithm based on interval envelope or trend comparison is performed on the time sequence readings of each sensor, and abnormal points are marked and rejected or marked for compensation to obtain a stage data set of each process.

[0021] In the stage data set of each process The entropy of the internal, the change amplitude of a certain time window, if the entropy value is less than a certain threshold, it is considered as information poor, and it may need to be spliced with adjacent sections to ensure modeling availability, if the value is too large, it prompts the sensor reading high-speed change, which may need to be encrypted sampling or further aggregation processing;

[0022] Preferably, the corrected stage data set is input, and the sensor readings and operation labels L op in the same batch or the same process are input env (t * ), and the environmental feature vector F(t * ) is constructed in the multidimensional space:

[0023] F(t * ) = [I env (t * ), X (1) (t * ), …, X (K) (t * ), L op (t * )]

[0024] The environmental feature vectors F(t * ) in different time periods are marked according to their corresponding carburized layer depth, hardness detection results or process deviation scalars (Δ c , Δ h , etc.);

[0025] The environmental feature vector F(t * ) is used to fit and output the environmental disturbance prediction value Ω(t * ), and the environmental disturbance model that can accurately map the environmental features to the process error or deviation trend is obtained by minimizing the loss function , so that the environmental feature vector F(t * ) at each time is corresponded to the predicted output of the environmental disturbance prediction value Ω(t * );

[0026] Preferably, the potential influence representation of the carburized layer depth and hardness at the current time is obtained from the environmental disturbance prediction value Ω(t * ) output by the environmental disturbance prediction output; according to the batch identifier and the timestamp t * , the environmental disturbance prediction value Ω(t * ) is correspondingly matched with the target depth and the target hardness of the workpiece, forming a triple

[0027] For the simultaneous introduction of comprehensive deviation value The degree of disturbance that the depth and hardness targets can be subjected to is measured;

[0028] If the integrated deviation value is higher than the preset deviation threshold, strong adaptive control is started; if the integrated deviation value is in the relatively safe interval, only slight or regular adjustment is made to the parameters;

[0029] Preferably, the process variable vector u(t * ) that can be adjusted by the adaptive controller is defined in combination with the integrated deviation value to construct a control target function

[0030] By minimizing the control target function , the optimal solution u * (t * ) of the process variable vector u(t * ) is obtained, and the controller issues real-time instructions to drive the carburizer valve, the heater, and the holding timer, to ensure that the process is adaptive to external disturbances;

[0031] In the sampling period of the adaptive controller, the optimal solution u * (t * ) of the process variable vector is quickly calculated according to the latest batch of environmental disturbance prediction values Ω(t * ) and the integrated deviation value and is sent to the execution end;

[0032] When the integrated deviation value is significantly reduced or the control target function is effectively reduced, it indicates that the control strategy helps to reduce the impact of environmental disturbances on the carburizing layer depth and hardness: if the integrated deviation value continues to rise, the correction amplitude of u(t * ) is increased in the next sampling period, forming a closed-loop control

[0033] Preferably, if the output environmental disturbance prediction value Ω(t * ) or the field sensor suddenly detects extreme disturbances during the control execution process, resulting in an integrated deviation value far exceeding the deviation threshold Θ c , the adaptive controller immediately enters an abnormal compensation mode, and if there is still no significant improvement, the reserved safety operation is forcibly enabled;

[0034] The safety limiting parameter interval is defined, and if the optimal solution u * (t * ) of the process variable vector calculated by the adaptive controller exceeds the interval, it is truncated with the nearest end value;

[0035] Preferably, after the parts are discharged, hardness detection and carburized layer depth determination are performed to obtain the final performance indicators corresponding to the batch of parts, and the detection results are collected as the process variable vector u(t * ) and the corresponding environmental disturbance prediction value Omega(t * ) of the adaptive control parameter, to generate the output data set D out ;

[0036] Preferably, the output data set D out is used to evaluate the accuracy of the environmental disturbance model and the adaptive controller in diversified scenarios, and the difference between the measured depth and the target , and the difference between the measured hardness and the target , and the output item of the environmental prediction Omega(t * ) are combined for segmented deviation analysis.

[0037] The new samples are added to the training set or validation set of the environmental disturbance model, and the model parameters are retrained or fine-tuned.

[0038] Preferably, the detection data of multiple batches and the corresponding environmental prediction and control parameters are regularly summarized to construct a long-term performance curve, and if the long-term performance curve shows that some auxiliary links repeatedly interfere with the carburizing quality, further optimization measures are introduced at the production planning layer or equipment operation layer; the external parameters of the environmental disturbance model are periodically checked.

[0039] (Three) beneficial effects

[0040] The present application provides a carburizing and quenching process parameter optimization method based on machine learning, which has the following beneficial effects:

[0041] Through multi-source environmental data collection and synchronous labeling, key sensing information such as aligned temperature, humidity, air pressure and carbon potential is effectively obtained, so that the target indicators can be bound under the same time reference, avoiding the information distortion caused by the dependence of traditional process on single-point measurement, and a unified batch identification and timestamp is established to ensure that the subsequent processing has sufficient data basis;

[0042] The environmental disturbance model constructed by deep learning or time series regression method maps the fused multi-dimensional features to the environmental disturbance prediction Omega(t * ), and quantifies the potential influence on the carburized layer depth and hardness, capturing the nonlinear effects of factors such as temperature and humidity fluctuations, furnace door opening and closing before process execution, making the production process have foresight and interpretability, and reducing the deviation of hardness and depth indicators;

[0043] According to the environmental disturbance prediction Omega(t *The integrated deviation between the target and the target is corrected in real time, the core process parameters such as carburant flow, heating power, holding time and the like are corrected, when the interference amplitude is large, adaptive compensation and safety limiting are triggered, the deep layer unevenness and hardness overshoot caused by environmental mutation can be avoided, and the product consistency and yield can be improved significantly;

[0044] By detecting the measured values such as carburizing layer depth and hardness, and corresponding to the adaptive control data executed, the model is retrained offline or online, the mechanism can continuously learn the equipment aging, sensor drift and environmental changes in four seasons, and iteratively optimizes the interference model and control logic, so that the long-period production quality and energy consumption balance are guaranteed;

[0045] In summary, the environmental disturbance recognition, parameter adaptive control and model continuous iteration are perfectly combined, the product deviation caused by environmental fluctuation can be significantly reduced, the system robustness and self-learning ability can be continuously improved in long-term operation, and the intelligent upgrading of carburizing and quenching process is supported. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a flowchart of the carburizing and quenching process parameter optimization method of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Please refer to Figure 1 The present application provides a carburizing and quenching process parameter optimization method based on machine learning, comprising,

[0049] Step one, when a new batch of parts is detected to be put into the furnace, the sensor array arranged in the workshop and the furnace body is called and the operation time stamp and batch identification are recorded synchronously, after the collected data is subjected to time alignment and abnormal point elimination, the synchronous labeled data set containing multi-dimensional environmental elements is generated;

[0050] The step one comprises the following contents:

[0051] Step 101, multi-source environmental data sensing arrangement and original information acquisition

[0052] The sensor array is arranged in each area of the carburizing and quenching workshop and the inside of the furnace body, including a plurality of temperature sensors T (k) , humidity sensors H (k) , air pressure sensors P (k) and carbon potential sensors C(k) where k is the sensor label;

[0053] Each sensor periodically outputs a set of raw measurements of multi-source environmental data, denoted as T (k) , H (k) (t), P (k) (t) and C (k) (t) respectively, t is the sampling time, the above raw measurements are sent to the data collection module, the data collection module archives the data according to the sensor label k, forms the raw environmental data set D raw and performs preliminary quality check and preprocessing, if the sensor is offline or the reading exceeds the limit threshold, it is marked as an abnormal point in the data structure;

[0054] Step 102, timestamp synchronization and environmental reference value marking

[0055] To realize the time alignment of the outputs of each sensor, a unified reference clock t * is established, each raw data T (k) (t), H (k) (t), P (k) (t), C (k) (t) is mapped to the corresponding reference clock to form a synchronous data record:

[0056] [t * , T (k) (t * ), H (k) (t * ), P (k) (t * ), C (k) (t * )]

[0057] At key production nodes such as part entering the furnace, temperature rising, carburizing, quenching, etc., operation labels L op are given respectively and the corresponding time t op is bound, the specific operation process is associated with the environmental sensor data, and the synchronous marked data set D sync is obtained;

[0058] An environmental reference index I env (t * ) is introduced to simplify and summarize the overall level of multi-dimensional environmental parameters at a certain time, which is defined as follows:

[0059]

[0060] In the formula: X(u) is the multi-source sensor reading vector collected at time u, which is defined as:

[0061]

[0062] wherein X (k) (u) represents the measurement value of the kth sensor (such as temperature, humidity, air pressure or carbon potential) at time u, K is the total number of sensors; Φ(X(u)) is a nonlinear conversion function vector for X(u), which is used to clip, amplify or map different measurement dimensions or ranges, and is defined as:

[0063]

[0064] Φ k (·) is a single sensor mapping function; a nonlinear mapping is performed on the sensor readings to eliminate the unbalanced value range caused by different dimensions or different scales, which can generally be set as a monotonic increasing function, a piecewise function or a function with interval processing;

[0065] W is a diagonal weighting matrix with size K x K:

[0066] W = diag[w1, w2, …, w K ]

[0067] where w k is the weight coefficient of the kth sensor in the environmental reference index, which is used to adjust the influence of different sensor quantities on the overall index; ||·||p p represents the L p norm of the vector, p is the spatial norm index, and is specifically:

[0068]

[0069] Corresponding to the WΦ(X(u)) vector is:

[0070]

[0071] γ is an additional norm index, and the value range is: γ > 1; Δt is the size of the time window backtracked from the current time t * ;

[0072] The environmental reference index I env (t * ) can be continuously calculated in the time dimension to measure the amplitude of environmental changes in time series, providing more targeted numerical basis for subsequent modeling and real-time regulation;

[0073] When used, calculating the environmental reference index I env (t * ) can extract a single index from high-dimensional sensor data to quickly judge the overall state of the environment, and the key operation label L opThe binding with the timestamp makes the subsequent environmental interference model not only identify the discrete degree of the environmental parameters but also directly trace the specific influence position of the environmental parameters in each step of the process flow, strengthen the correlation analysis of the process and the environment, and provide high-quality input with operation scenarios and comprehensive indexes, so as to realize smooth transition from multi-source sensing to model training.

[0074] Step two, fuse the temperature field in the fusion furnace with the carbon potential variation characteristics, extract the key feature vectors affecting the carburizing layer depth and hardness, and output the environmental interference prediction value through a prediction function;

[0075] The step two includes the following contents:

[0076] Step 201, data consistency processing and noise anomaly identification

[0077] From the synchronization mark data set D sync , obtain multi-source environmental data (temperature, humidity, air pressure, carbon potential, etc.) with a timestamp t * , wherein each record carries its operation label L op ; identify the data segments corresponding to different process stages of the same batch of parts according to the operation label, and group and arrange them into

[0078] Anomaly elimination algorithm based on interval envelope or trend comparison is performed on the time sequence readings of each sensor, wherein:

[0079]

[0080] Wherein, X (k) (t * ) is the value of the kth sensor at time t * , μ k (t * ) is the local smoothing center (which can be obtained by a history window or a regression method), ω k is an adjustable amplification factor, the value is greater than 0, Y(t * ) is a self-definable envelope or trend function, for example, an exponential weighted moving average (EWMA) function can be taken; if ψ = 1, it is marked as an abnormal point and is eliminated or marked for compensation in subsequent analysis, and after consistency alignment, noise filtering and anomaly value elimination processing, the stage data set of each process is obtained

[0081] In the stage data set of each process Inside, introduce information entropy class method (discrete Shannon entropy formula) or interval distance based complexity measurement, statistics its change amplitude in a certain time window, if the entropy value is less than a certain threshold, it is considered as information poor, and may need to be spliced with adjacent section to ensure modeling availability: if the value is too large, it may need to encrypt sampling or further aggregation processing;

[0082] In use, on the one hand, the consistency of environmental data in time and operation label can be ensured, and on the other hand, the suspicious readings or low quality fragments are eliminated through targeted anomaly identification and phased data quality evaluation, so as to lay a high credibility data foundation for the construction of environmental interference model. Flexible and multi-scale anomaly detection is realized; combined with information entropy or interval distance measurement, different fluctuation levels can be adapted in different process stages, showing high scalability.

[0083] Step 202, multi-dimensional construction of environmental interference model and prediction output

[0084] With the corrected stage data set As input, the sensor readings {X (1) (t * ), X (2) (t * ),…, X (K) (t * )} and operation label L op in the same batch or the same process are combined with the environmental reference index I env (t * ), and the environmental feature vector F(t * ) is constructed in the multi-dimensional space:

[0085] F(t * ) = [I env (t * ), X (1) (t * ),…, X (K) (t * ), L op (t * )]

[0086] The environmental feature vectors F(t * ) in different time periods are marked according to their corresponding carburized layer depth, hardness detection results or process deviation scalar (Δ c , Δ h , etc.), which provides supervision target for subsequent environmental interference model training;

[0087] The environmental interference model is constructed through deep neural network (such as time sequence convolution model) or index kernel function based regression method;

[0088] Using environmental feature vector F(t) * Fit and output the predicted value of environmental disturbance Ω(t) * ), Environmental disturbance prediction value Ω(t) * The aim is to quantify at time t * At that time, the overall deviation trend of the environment on the depth and hardness of the carburized layer;

[0089] To reflect the enhancement of the effects of nonlinearity and extrema, the following formula can be used as the loss function.

[0090]

[0091] In the formula: y(u) represents the actual measured depth of carburized layer and hardness index (or process deviation) within the time interval. The truth sequence within; Γ is the corresponding sequence of the model's predicted output; Γ is a predefined or adjustable nonlinear mapping used to highlight the maximum error or the error in a specific interval; η is the power exponent that adjusts the sensitivity of this integral to the overall loss, and its value is greater than 0.

[0092] By minimizing the loss function An environmental disturbance model that can accurately map environmental characteristics to process errors or deviation trends can be obtained, and the environmental feature vector F(t) at each time step can be obtained. * ) corresponds to the predicted environmental disturbance value Ω(t) * The predicted output of ).

[0093] The output environmental disturbance prediction value Ω(t) of the environmental disturbance model * ) Perform buffering smoothing or time-series extrapolation to obtain environmental disturbance prediction results, and calculate the environmental disturbance prediction value Ω(t) for each batch or time period. * Regularly saving data can help with traceability and comparison during continuous process iterations and updates, enabling continuous optimization of the model over the long term.

[0094] In practice, advanced modeling techniques such as deep learning or kernel regression are used to capture complex environment-process relationships in a multi-dimensional feature space, and nonlinear loss is applied. Enhance the ability to fit extreme deviations, enabling the model to more accurately quantify the impact of environmental disturbances on carburizing depth and hardness.

[0095] By analyzing the predicted value of environmental disturbance Ω(t) * By extrapolating the time series, the risk of process deviation caused by environmental fluctuations can be predicted in advance, providing timely decision information for adaptive parameter tuning.

[0096] Step three, when the environmental disturbance prediction value and the preset target exist significant difference, the adaptive controller calculates the optimal process control parameters according to the comprehensive deviation function, and adjusts the carburant flow, heating power and holding time in real time to ensure the carburizing uniformity and the stability of the part performance;

[0097] The step three includes the following contents:

[0098] Step 301, multi-index quantization of target and disturbance mapping

[0099] The current time potential influence representation on the carburized layer depth and hardness is obtained from the environmental disturbance prediction value Ω(t * ) output by the environmental disturbance prediction;

[0100] According to the batch identification and the time stamp t * , the environmental disturbance prediction value Ω(t * ) is matched with the target depth and the target hardness of the workpiece, forming a triple

[0101] In order to measure the disturbance degree of the depth and hardness targets at the same time, the comprehensive deviation value is introduced, which is defined as:

[0102]

[0103] Wherein: is used to calculate the carburized layer depth deviation degree caused by the environmental disturbance; is used to calculate the hardness deviation degree caused by the environmental disturbance; β is the power index, which can amplify the punishment degree of large deviation, and the value is greater than 0;

[0104] Wherein uses the following formula, Therefore:

[0105]

[0106] Wherein: Ω c (t * ) represents the carburized layer depth predicted by the environmental disturbance model at time t * ; represents the target carburized layer depth required by the process; τ c is the depth tolerance parameter, which is used to set the error allowed within the normal process fluctuation range. When the actual deviation is lower than τ c , the function output is zero, indicating that the deviation is within the allowed range;

[0107] If the comprehensive deviation value is higher than the preset deviation threshold, it indicates that under the given parameter setting, environmental disturbance may cause a large deviation of the carburized layer depth or hardness, and strong adaptive regulation needs to be started. If the value is in the relatively safe interval, only slight or regular adjustment is made to the parameters.

[0108] In use, the environmental disturbance prediction Ω(t * ) is closely mapped with the process target (depth and hardness), thereby generating a numerical comprehensive deviation value which provides a clear and adjustable quantitative standard for the trigger mechanism of subsequent real-time adjustment.

[0109] Step 302, parameter update strategy and real-time control implementation

[0110] Define the adjustable process variable vector u(t * ) of the adaptive controller:

[0111] u(t * ) = [u C (t * ), u P (t * ), u T (t * )]

[0112] Wherein: u C (t * ) represents the real-time carburant flow coefficient, used to dynamically adjust the supply amount of carbonizing agent in the carburizing process; u P (t * ) represents the heating power correction amount, indicating the correction amplitude of the heater output power in real-time process regulation; u T (t * ) represents the holding time offset, used to adjust the actual maintenance time of the holding stage, offset relative to the preset holding time; combined with the comprehensive deviation value build the regulation target function as follows:

[0113]

[0114] Wherein, Γ(·) is a nonlinear function that converts the comprehensive deviation value into the contribution to the regulation accuracy;

[0115] γ(·) is used to represent the energy consumption or additional overhead of the adjustment variable, such as excessive increase of heating power which will increase energy consumption and also needs to be balanced in regulation; α1 and α2 are weight coefficients, taking values between 0 and 1;

[0116] By minimizing the regulation target function Obtaining the optimal solution u * (t * ) of the process variable vector u(t * ), and making the controller issue real-time instructions to drive the carburizer valve, the heater, and the holding timer to ensure that the process is adaptive to external disturbances;

[0117] In the adaptive controller, the optimal solution u * (t * ) of the process variable vector is quickly calculated according to the latest batch of environmental disturbance prediction value Ω(t * ) and the comprehensive deviation value , and is sent to the execution end in the sampling period;

[0118] When the comprehensive deviation value is significantly reduced or the control target function is effectively reduced, it indicates that the control strategy helps to reduce the influence of environmental disturbances on the carburizing layer depth and hardness; if the comprehensive deviation value continues to rise, the correction amplitude of u(t * ) is increased in the next sampling period to form a closed-loop control

[0119] In use, a control target that can integrate multiple factors (process quality requirements, energy consumption, disturbance prediction) is established, and nonlinear optimization is performed according to the size of , so that a sensitive and rapid response to the process core parameters is achieved when environmental disturbances occur. The traditional single-dimensional control (only temperature adjustment or only carbon potential adjustment) is changed, and multiple key parameters are scheduled by the process variable vector u(t * ) vector at one time, which is more beneficial to maintaining the carburizing uniformity in complex disturbance situations.

[0120] Step 303, abnormal feedback compensation and dynamic safety limiting

[0121] If the output environmental disturbance prediction value Ω(t * ) or the field sensor suddenly detects extreme disturbances (such as sudden pressure drop, abnormal temperature of tooling) during the control execution process, resulting in a comprehensive deviation value far exceeding the deviation threshold Θ c , the adaptive controller immediately enters the abnormal compensation mode;

[0122] In the abnormal compensation mode, the weight coefficient α1 is temporarily increased to preferentially reduce the deviation caused by the environment. If there is still no significant improvement, the reserved safety operation is forced to be enabled, such as rapidly shortening the holding time, reducing the heating power, to avoid extreme situations such as overheating of the workpiece or over-saturation of carburizing;

[0123] As a further content, to prevent frequent and large-scale parameter adjustment from causing severe fluctuations in furnace temperature or carbon potential, a set of safety limiting parameters u min and u max are defined, which satisfy:

[0124] u min ≤u(t * )≤u max

[0125] wherein: u min and u max are the minimum and maximum operation boundaries set based on the physical performance of the equipment and the process limit requirements; when the optimal solution u * (t * ) of the process variable vector calculated by the adaptive controller exceeds the interval, it is truncated with the nearest end value to ensure the safety of the production line and the service life of the equipment;

[0126] The process variable vector u(t * ) used in the abnormal compensation or limiting regulation process and the corresponding comprehensive deviation value are recorded together.

[0127] In use, on the basis of normal real-time optimization, abnormal compensation and limiting mechanism are introduced to ensure that the system will not lose control due to extreme disturbance or short-term sensor deviation, improve process safety and stability; by recording the compensation process under abnormal conditions, the model can be gradually trained in subsequent iterative updates on how to more effectively respond to rare but highly influential interference scenarios, combining environmental disturbance and process targets, using nonlinear methods to quickly determine the optimal tuning scheme, and comprehensively using compensation mechanism and safety limiting, taking into account the flexibility of responding to extreme disturbances and process safety.

[0128] Step four, after the parts are discharged and the hardness and carburized layer depth detection are completed, the measured values, optimal process control parameters and environmental disturbance prediction values are stored in the output data set together, and the environmental disturbance model and the adaptive controller are incrementally trained to gradually improve the process stability and precision in long-term variable environment;

[0129] The step four includes the following contents:

[0130] Step 401, finished product detection and data integration

[0131] After the parts are discharged, hardness detection (such as Vickers or Rockwell hardness measurement), carburized layer depth measurement (such as metallographic analysis or automatic depth measurement) are performed to obtain the final performance indicators corresponding to the parts in this batch, and the detection results are recorded in the form of , wherein represents the measured carburized layer depth, represents the measured hardness value;

[0132] The process variable vector u(t * ) used as adaptive control parameters and the corresponding environmental disturbance prediction value Ω(t *); and ensure that it is consistent with and Through batch identification L batch or timestamp t * Corresponding; generate output data set D out The basic data elements can be described as follows:

[0133]

[0134] Facilitate accurate traceability and comparison in subsequent model training process; compare the actual performance of this batch of finished products with the environmental parameters (Ω(t * ) and process control u(t*) sampled at the time to form a complete data closed loop, ensuring that subsequent model evaluation and iterative training are more targeted.

[0135] Step 402, model evaluation and online / offline retraining

[0136] Use the output data set D out to evaluate the accuracy of the environmental disturbance model and the adaptive controller in diverse scenarios. The adaptive controller is used to calculate and adjust the key parameters of the carburizing and quenching process in real time, which can be obtained by machine learning algorithm training:

[0137] The difference between the measured depth and the target , and the difference between the measured hardness and the target , combined with the output items of the environmental prediction Ω(t * ) (such as Ω c (t * ), Ω h (t*)) for segmented deviation analysis;

[0138] Add new samples to the training set or validation set of the environmental disturbance model, and retrain or fine-tune the model parameters through offline methods (such as batch updating during non-production periods) or online methods (small batch updating while producing); To prevent the control strategy from being unstable due to frequent model iteration, an adaptive rate κ a can be set to trigger the official version update after a certain amount of data is accumulated in multiple batches;

[0139] When used, by including new finished product test results in the training process, the environmental disturbance model can continuously adapt to seasonal or extreme environmental changes, avoiding gradual decline in model accuracy due to environmental conditions or part batch replacement, while allowing the adaptive controller to continuously evolve to cope with error accumulation or special scenarios in real production conditions, ensuring long-term steady-state performance. Offline retraining and online small batch updating are flexibly combined, and an adaptive rate κa With version management, both timely rectification and frequent updates are avoided, and multiple over-fitting or control shock problems are introduced; For special parts of different batches or extreme environment, higher weight can be given in this process to form a hierarchical retraining mechanism, with more fine adaptability.

[0140] Step 403, long-term performance record and global optimization feedback

[0141] Periodically summarize the multi-batch detection data and the corresponding environment prediction and control parameters, and evaluate the overall stability and energy consumption level of the carburizing and quenching process under different periods and different environmental fluctuations in the form of time series analysis or period comparison;

[0142] On this basis, a long-term performance curve can be constructed to globally observe the trend of carburized layer depth and hardness qualification rate over time and identify potential degradation (such as equipment aging, sensor misalignment) or breakthrough points (process upgrade);

[0143] If long-term performance analysis shows that some auxiliary links (such as cooling oil temperature maintenance, furnace door sealing performance) repeatedly interfere with carburizing quality, further optimization measures (such as replacing sealing components, correcting cooling process) can be introduced at the production planning layer (MES system) or equipment operation and maintenance layer to reduce frequent interference with core carburizing regulation;

[0144] Periodically check the external parameters of the environmental interference model (such as air pressure detection channel, furnace gas supply system) to prevent the model error from accumulating due to hardware aging or sensor drift;

[0145] On the workshop big data platform or MES system, the accumulated time series analysis results are compared with economic indicators (such as energy consumption, production, scrap rate) in multiple dimensions to facilitate better production scheduling or process configuration decisions based on real historical data;

[0146] In use, the single or short-term process results are extended to long-term trend monitoring and global decision optimization, and independent batch local improvement is gradually transformed into whole production line, whole cycle systematic improvement; Through cumulative record and time series analysis, the aging law of equipment and sensors can also be understood, hidden dangers can be found and repaired earlier, and the process can be ensured to run continuously and reliably.

[0147] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0148] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0149] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0150] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0151] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for optimizing carburizing and quenching process parameters based on machine learning, characterized in that: Comprising, When a new batch of parts is detected into the furnace, call the sensor array arranged in the workshop and the furnace, after time alignment and outlier rejection of the collected data, generate a synchronous labeled dataset containing multi-dimensional environmental factors; Fusion of temperature field and carbon potential variation characteristics in the furnace, extraction of key feature vectors affecting carburizing layer depth and hardness, and output of environmental disturbance prediction value through prediction function; When there is a significant difference between the environmental disturbance prediction value and the preset carburizing layer target and hardness target, the adaptive controller calculates the optimal process control parameters according to the comprehensive deviation value, and adjusts the carburizing agent flow, heating power and holding time in real time; After the parts are taken out of the furnace and the hardness and carburizing layer depth detection is completed, the measured values, optimal process control parameters and environmental disturbance prediction values are stored in the output dataset, and the environmental disturbance model and adaptive controller are incrementally trained; Each sensor periodically outputs a set of raw measurement values of multi-source environmental data, which is sent to the data acquisition module. According to the sensor label, the data is archived to form the original environmental dataset and undergo preliminary quality check and preprocessing. At key production nodes, operation labels are assigned and corresponding time points are bound. Correlate the operation process with environmental sensor data to obtain a synchronous labeled dataset; and introduce an environmental reference index that can be continuously calculated and measured in the time dimension to simplify and summarize the overall level of multi-dimensional environmental parameters at a certain time.

2. The machine learning-based carburizing and quenching process parameter optimization method of claim 1, wherein: From the synchronous labeled dataset, obtain multi-source environmental data, identify data segments corresponding to different process stages of the same batch of parts according to the operation label, and group and organize them; The outlier rejection algorithm executed on the time series readings of each sensor marks the outliers and removes or compensates them to obtain the corrected stage dataset of each process, and the change amplitude of the dataset within a certain time window is calculated.

3. The machine learning-based carburizing and quenching process parameter optimization method of claim 2, wherein: Using the corrected stage dataset as input, an environmental feature vector is constructed in a multi-dimensional space by combining sensor readings, operation labels and environmental reference index; The environmental feature vectors in different time periods are labeled according to their corresponding carburizing layer depth, hardness detection results or process deviation scalar; The environmental disturbance prediction value is fitted and output using the environmental feature vector, and each time point environmental feature vector is mapped to the prediction output of the environmental disturbance prediction value.

4. The machine learning-based carburizing and quenching process parameter optimization method of claim 3, wherein: According to the batch identification and time stamp, the environmental disturbance prediction value is matched with the target depth and target hardness of the workpiece to form a triple for subsequent control calculation, and a comprehensive deviation value is introduced to measure the disturbance degree of the depth and hardness targets; If the comprehensive deviation value is higher than the preset deviation threshold, start strong adaptive control, and if the comprehensive deviation value is in a relatively safe range, only slightly or regularly adjust the parameters.

5. The machine learning-based carburizing and quenching process parameter optimization method of claim 4, wherein: Defining the process variable vector adjustable by the adaptive controller, and combining the comprehensive deviation value to build the regulation and control target function; By minimizing the regulation and control target function, the optimal solution of the process variable vector is obtained, and the controller issues real-time instructions to drive the carburizer valve, heater and holding timer, to ensure that the process is adaptive to external disturbances; When the comprehensive deviation value does not significantly decrease or the regulation and control target function is not effectively reduced, increase the correction amplitude of the process variable vector in the next sampling period.

6. The machine learning-based carburizing and quenching process parameter optimization method according to claim 5, characterized in that: If the output environmental disturbance prediction value or the field sensor burst detects extreme disturbance during control execution, resulting in a comprehensive deviation value far exceeding the deviation threshold, the adaptive controller immediately enters the abnormal compensation mode; if there is still no significant improvement, the reserved safety operation is forced to be enabled; Defining a safety amplitude parameter interval, if the optimal solution of the process variable vector calculated by the adaptive controller exceeds this interval, the nearest end value is truncated.

7. The machine learning-based carburizing and quenching process parameter optimization method according to claim 6, characterized in that: After the parts are discharged, hardness detection and carburizing layer depth determination are performed to obtain the performance indicators corresponding to the parts in this batch; Collecting the process variable vector as an adaptive regulation and control parameter and the corresponding environmental disturbance prediction value to generate an output data set.

8. The machine learning-based carburizing and quenching process parameter optimization method according to claim 7, characterized in that: Using the output data set to evaluate the precision performance of the environmental disturbance model and the adaptive controller in diversified scenarios; Comparing the difference between the measured data and the target, and combining the output items of environmental prediction for segmented deviation analysis, adding the new samples to the training set or validation set of the environmental disturbance model, and retraining or fine-tuning the model parameters.

9. The machine learning-based carburizing and quenching process parameter optimization method according to claim 8, characterized in that: Regularly summarize the detection data of multiple batches and the corresponding environmental prediction and control parameters to build a long-term performance curve; If the long-term performance curve shows that some auxiliary links repeatedly cause significant disturbance to the carburizing quality, introduce further optimization measures at the production planning layer or equipment operation and maintenance layer, and periodically verify the external parameters of the environmental disturbance model.

Images