Intelligent speed control method for forging press slide driving motor

Abstract

The present application relates to a kind of forging press slide block drive motor intelligent speed regulation control method, to solve the problem of insufficient physical consistency of predicted speed sequence and process stage logic identification accuracy.The core scheme is to collect real-time signals such as slide block load, displacement and pressure through multi-source sensors, generate high signal-to-noise ratio data stream through synchronous alignment and noise filtering, determine the slide block motion phase and segment the predicted speed curve segment using kinematic feature events and state machine logic mapping, perform symbolic logic verification on each speed curve segment based on the segmented kinematic envelope constraint set, perform template remapping physical repair on abnormal segments, and finally output a physically consistent speed control sequence to drive smooth motion of the slide block.This method can improve the physical feasibility of speed switching in the forging process and the accuracy of process stage identification, and realize model prediction and actual control closed-loop iterative optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent control of servo drive for forging presses and physical consistency governance technology of speed prediction based on deep learning, and in particular to an intelligent speed regulation control method for a slide drive motor of a forging press. Background Technology

[0002] Existing servo drive control systems for forging press slides widely employ intelligent models and predictive methods to improve the accuracy and responsiveness of speed regulation. Typical solutions include deep learning model-driven speed curve prediction, Physical Information Neural Network (PINN), physical weighting of loss functions, multi-model fusion, and simulation data-assisted optimization based on differential equation embedding. These technologies can accurately reproduce the speed motion process of the forging press slide during each stage of downward start-up, pressurization steady state, return acceleration, and braking stop, under conditions of sufficient data and stable operating conditions, and empower the control system to achieve semi-automatic or intelligent speed regulation goals. With the continuous improvement of forging automation, intelligent prediction modules have become an important guarantee for supporting high-speed closed-loop control at the servo end. Industry development trends focus on high-precision, adaptive, and physically consistent intelligent speed sequence prediction and decision-making.

[0003] Current mainstream technologies can be broadly categorized into two types: one is a scheme that directly maps control conditions to velocity sequences end-to-end based on deep neural networks, commonly seen in models such as convolutional neural networks, recurrent neural networks, and LSTM trained on limited typical working condition datasets. This type of method typically obtains time-series predictions of slider velocity through large-scale end-to-end training and directly outputs the predictions to the servo drive. The other type introduces physical knowledge or dynamic models as model constraints. By constructing special loss functions, introducing state variable constraints, or adding simulation trajectories to the training data, it attempts to explicitly enhance physical consistency or boundary feasibility, avoiding significant distortion in the predicted values. Some recent solutions also attempt to "hard-embed" the motion equations into the neural network structure using Physically Consistent Neural Networks (PINNs) to improve robustness and generalization ability under extreme conditions.

[0004] These methods are generally applicable to speed prediction and slider control under common process environments and have a certain degree of robustness to moderately complex load variations. However, in practical engineering applications, when the forging press encounters extreme load changes, abnormal sensor inputs during slider movement, or a sharp shift in historical sample distribution from the current operating conditions, traditional end-to-end models are prone to outputting speed sequences with violent fluctuations, instantaneous jumps, distortion of physical boundaries, or even violations of dynamic laws. While physical information neural network methods can partially reduce physically infeasible anomalies, their training process relies on accurate dynamic models, and their deployment on actual machines is constrained by numerical stability and computational burden. Furthermore, they cannot provide a fallback solution for unforeseen new anomalies during the inference phase. Schemes such as physically weighted loss functions and multi-model fusion usually require retraining the model, have a high dependence on changes in the structure of existing deployed models, and are difficult to respond quickly to changes in the field and meet the realities of heterogeneous industrial deployments.

[0005] The existing technology has the following main drawbacks: When faced with data distribution deviations or extreme operating conditions, deep learning prediction models often experience sudden changes in local velocity sequences or abnormal motion characteristics. In particular, the downward and braking ends are prone to serious instability such as violations of acceleration monotonicity and sudden velocity jumps that do not converge, which directly affect the smoothness of slider motion and execution safety.

[0006] Current popular methods for enhancing physical consistency generally rely on mandatory embedding constraints during the training phase, resulting in a lack of effective fallback mechanisms for unknown oscillations and non-standard anomalies during the inference phase. Physical inconsistency issues during extreme mutations are difficult to completely avoid through end-to-end training.

[0007] In real-world industrial environments, servo drive closed loops are subject to various non-ideal factors such as sensor drift, load disturbance, and aging. Traditional methods such as model optimization, simulation sample expansion, and loss reweighting are slow to respond and cannot quickly implement structured governance measures for individual operating conditions, and lack safety assurance capabilities independent of the model itself.

[0008] The lack of a lightweight, pluggable, and auditable runtime physical consistency verification scheme makes it difficult for the control layer to make closed-loop decisions for extreme anomalies. Once the speed control command is affected by an unstable sequence, the servo motor may oscillate, impact, or even lose synchronization.

[0009] The aforementioned problems make it difficult for existing technologies to meet the rigid requirements of the forging servo drive intelligent control system for the physical stability of the output speed curve under complex working conditions. In particular, under scenarios such as data distribution deviation, extreme loads, or on-site sensor failures, deep learning models cannot guarantee the physical consistency of the slider movement throughout the entire cycle and the response safety of the control system. There is an urgent need for a safety fallback mechanism for extreme working conditions that does not change the original model structure and can be flexibly deployed at the execution link end.

[0010] Therefore, current technology requires a novel physical feasibility management scheme that can perform physical envelope verification and local remapping of the speed curve sequence before execution based on process phase after the deep learning model predicts the output. This would ensure that the speed characteristics of each process stage meet the dynamic boundary and eliminate abrupt behavior, thereby significantly improving the stability, safety, and engineering feasibility of the forging servo drive control system under extreme working conditions and data offset scenarios, and providing key technical support for the continuous and healthy development of forging automation and intelligent manufacturing. Summary of the Invention

[0011] This application provides an intelligent speed control method for a slide drive motor of a forging press, which aims to solve one of the problems or issues of the prior art mentioned in the background.

[0012] This application provides an intelligent speed control method for a slide drive motor of a forging press, specifically including: S1: Acquire multi-source sensor data during the operation of the forging press slide, and generate a slide motion phase label sequence containing several process stage identifiers based on the multi-source sensor data; S2: Based on the dynamic boundary characteristics of each process stage in the slider motion phase label sequence, a segmented kinematic envelope constraint set is constructed. The segmented kinematic envelope constraint set is defined as a set of engineering verifiable rules preset for each phase stage. S3: Based on the time segmentation points of the slider motion phase label sequence, the original predicted rotational speed time sequence output by the preset deep learning model is divided into sub-rotational speed curve segment sequences corresponding to the process stage; S4: Using the segmented kinematic envelope constraint set, perform symbolic logic comparison operation on the sub-speed curve segment sequence to determine whether each sub-speed curve segment satisfies the envelope boundary and qualitative characteristics of the corresponding process stage, and generate physical feasibility verification results; S5: For the abnormal sub-speed curve segments marked as constraint violations in the physical feasibility verification results, the nearest working condition template is retrieved based on the preset phase typical template library, and the abnormal segments are remapped to the corresponding positions of the nearest working condition template to generate reconstructed speed curve segments. S6: Replace the abnormal sub-speed curve segment that fails the verification with the corresponding reconstructed speed curve segment, and keep the verified sub-speed curve segment in its original state, and then splice the sub-speed curve segments in time order to generate the final physically consistent speed control command sequence.

[0013] S7: Input the final physical consistency speed control command sequence into the speed loop controller of the servo driver to drive the forging press slide to perform a smooth speed switching action. At the same time, during the execution process, the phase matching degree, remapping amplitude and constraint violation number are collected in real time as runtime governance indicators.

[0014] S8: Update the sample weights of the online evaluation module of the deep learning model based on the runtime governance metrics.

[0015] The intelligent speed control method for a slide drive motor of a forging press provided in this application has the following beneficial effects: (1) By constructing a runtime verification mechanism deployed at the end of the inference chain, this solution effectively overcomes the problems of weak model generalization ability and poor adaptability to unknown working conditions caused by the forced embedding of physical constraints in the training phase of traditional intelligent speed control methods. Existing technologies usually rely on physical information neural networks or loss functions with constraints to achieve "intrinsic compliance". However, when faced with common industrial scenarios such as data distribution offset, sensor signal drift or sudden load disturbance, abnormal prediction outputs that violate kinematic logic are still prone to occur, which may lead to instability of the drive motor or even mechanical shock. This invention breaks away from this paradigm and decouples the physical feasibility guarantee from the model structure. Instead, it implements runtime governance with strong semantic coupling but structural decoupling on the prediction results at the end of the inference chain by combining phase identification rules with segmented envelope constraints. This mechanism can not only accurately identify the four key process stages of slider movement—downward movement, pressurization, return, and braking—and trigger corresponding verification logic based on the real-time generated phase label sequence, but also efficiently detect whether the speed curve meets the qualitative characteristics of each stage, such as monotonicity, convergence, continuity, and amplitude boundaries, using a symbolic judgment method. Thus, without changing the original prediction model architecture, it can significantly improve the physical rationality and process safety of the overall system output.

[0016] (2) For verification failures, this solution proposes a local remapping strategy based on a typical template library, avoiding the high computational overhead and real-time bottlenecks caused by traditional repair methods that rely on numerical optimization, iterative solutions, or multi-model collaboration. Unlike the complex process of adjusting weights through backpropagation or introducing additional simulation data for retraining, this method uses a combination of lightweight pattern matching and geometric transformation to retrieve the nearest similar cases from a pre-set working condition template library. It then performs deterministic operations such as scaling, translation alignment, or local truncation on abnormal speed segments, grafting them onto the corresponding positions of standard templates. This preserves the overall trend information of the original prediction while forcibly repairing the parts that violate engineering common sense. This remapping process does not require gradient calculation or latent variable solving, and the response delay is only in the hundreds of milliseconds, fully meeting the high dynamic response control requirements of forging presses. More importantly, this mechanism has good interpretability and traceability. Each correction can be attributed to clear phase constraints and template references, providing solid support for fault diagnosis, responsibility identification, and safety auditing of industrial systems.

[0017] (3) By introducing a closed-loop feedback channel for verification indicators, this scheme achieves adaptive linkage between performance evaluation of the prediction model and subsequent iterative optimization, constructing a complete closed-loop governance system of "prediction-verification-repair-learning". The system not only records and feeds back runtime indicators such as phase matching degree, number of constraint violations, and remapping magnitude to the online model evaluation module in real time, but also uses these signals to guide the identification and weighting of difficult samples, enabling the model to actively pay attention to historical high-frequency violation conditions during subsequent training, and gradually enhance its intrinsic stability. This plug-in but semantically closed-loop design allows the model to continuously absorb physical law feedback from the actual operating environment while maintaining lightweight and efficient inference, forming a virtuous cycle of "experience accumulation-behavior correction-capability evolution". Compared with the traditional improvement path that requires reconstructing the network structure or redesigning the loss function, this scheme is more practical in engineering and has greater system scalability, especially suitable for complex industrial sites with multiple models, variable parameters, and non-steady-state operation.

[0018] In summary, this solution innovatively achieves decoupled and collaborative control of intelligent prediction and engineering feasibility by transferring physical constraints from the training phase to the inference phase. This not only significantly improves the robustness and safety of the forging press slide drive motor under extreme conditions, but also takes into account real-time performance, interpretability, and system maintainability, providing a practical and feasible technical path for the intelligent upgrading of high-precision heavy equipment. Attached Figure Description

[0019] Figure 1 This is the main flowchart of an intelligent speed control method for a slide drive motor of a forging press. Figure 2 This is a sub-flowchart of an intelligent speed control method for a slide drive motor of a forging press; Figure 3 This is another sub-flowchart of a method for intelligent speed control of a slide drive motor in a forging press. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0022] like Figure 1 As shown, this application provides an intelligent speed control method for a slide drive motor of a forging press, specifically including: S1: Acquire multi-source sensor data during the operation of the forging press slide, and generate a slide motion phase label sequence containing several process stage identifiers based on the multi-source sensor data. The slide motion phase label sequence includes four process stage identifiers: downward start, pressurization steady state, return acceleration, and braking stop. The multi-source sensor data includes: real-time load torque signal, displacement slope signal, and pressure platform duration signal.

[0023] S2: Based on the dynamic boundary characteristics of each process stage in the slider motion phase label sequence, a piecewise kinematic envelope constraint set is constructed. The piecewise kinematic envelope constraint set is defined as a set of engineering-verifiable rules preset for each phase stage. The set of engineering-verifiable rules includes rotational speed monotonicity, acceleration non-negativity, fluctuation amplitude threshold, and convergence continuity requirements.

[0024] S3: Based on the time segmentation points of the slider motion phase label sequence, the original predicted rotational speed time sequence output by the preset deep learning model is divided into sub-rotational speed curve segments corresponding to the process stage.

[0025] Specifically, the original predicted rotational speed time series output by the deep learning model is used as the input object, and the original predicted rotational speed time series is divided into sub-rotational speed curve segments corresponding to the four process stages based on the time segmentation points of the slider motion phase label sequence.

[0026] S4: Using the segmented kinematic envelope constraint set, perform a symbolic logic comparison operation on the sub-speed curve segment sequence to determine whether each sub-speed curve segment satisfies the envelope boundary and qualitative characteristics of the corresponding process stage, and generate a physical feasibility verification result. The physical feasibility verification result includes: a phase matching degree index and a constraint violation flag.

[0027] S5: For the abnormal sub-speed curve segments marked as constraint violations in the physical feasibility verification results, the nearest operating condition template is retrieved based on a preset phase typical template library, and the abnormal segments are remapped to the corresponding positions of the nearest operating condition template to generate reconstructed speed curve segments that conform to physical behavior logic. Specifically, the abnormal segments are remapped to the corresponding positions of the nearest operating condition template through scaling, translation, or truncation operations.

[0028] S6: Replace the abnormal sub-speed curve segment that fails the verification with the corresponding reconstructed speed curve segment, and keep the verified sub-speed curve segment in its original state, and then splice the sub-speed curve segments in time order to generate the final physically consistent speed control command sequence.

[0029] S7: Input the final physical consistency speed control command sequence into the speed loop controller of the servo driver to drive the forging press slide to perform a smooth speed switching action. At the same time, during the execution process, the phase matching degree, remapping amplitude and constraint violation number are collected in real time as runtime governance indicators.

[0030] S8: Update the sample weights of the online evaluation module of the deep learning model based on the runtime governance metrics. Specifically, use the phase matching degree and constraint violation count to identify hard examples and complete the closed loop of model iterative optimization for data distribution offset scenarios.

[0031] Step S1: Acquire multi-source sensor data during the operation of the forging press slide, and generate a slide motion phase label sequence containing several process stage identifiers based on the multi-source sensor data. The slide motion phase label sequence includes four process stage identifiers: downward start, pressurization steady state, return acceleration, and braking stop. The multi-source sensor data includes: real-time load torque signal, displacement slope signal, and pressure platform duration signal. Specifically, it includes: S1.1: Acquire the real-time load torque signal fed back by the servo drive system of the forging press, the displacement slope signal measured by the grating ruler, and the pressure platform duration signal collected by the pressure sensor, and perform synchronous timestamp alignment and noise filtering on the above multi-source sensor data to generate a high signal-to-noise ratio original working condition data stream with a unified time base.

[0032] In the main step S1 of the intelligent speed control method for the slide drive motor of a forging press, sub-step S1.1 undertakes the fundamental task of multi-source sensor signal acquisition and synchronous preprocessing. Its function is to generate a high signal-to-noise ratio raw operating condition data stream that can be used for subsequent feature extraction by unifying the time base and noise reduction processing of the real-time load torque signal fed back by the servo drive system, the displacement slope signal measured by the grating ruler, and the pressure platform duration signal collected by the pressure sensor. The execution objects include the torque feedback module, the grating ruler displacement measurement device, and the pressure sensor array in the servo drive system. The initial condition is that the equipment is in normal operation and the sampling system clock is stable.

[0033] The load torque signal fed back by the servo drive system is sampled at high speed, and during the sampling process, the hardware trigger interface is called to read the displacement slope signal output by the grating ruler and the pressure platform duration signal output by the pressure sensor, forming a set of multi-source raw signals.

[0034] A unified clock source reference is used to record timestamps on the multi-source raw signal set to ensure that the data at each sampling point has an accurate relative time value. The time axis alignment of each signal is achieved through synchronously triggered sampling indexes.

[0035] A sliding window mean filter is used to smooth short-term fluctuations in the load torque signal. The window length is set according to the natural frequency of the mechanical system to suppress high-frequency noise without weakening the dynamic response capability.

[0036] The displacement slope signal is identified by fast Fourier transform, and high-frequency spikes that are not related to the process are suppressed by a bandpass filter. The bandpass interval is set according to the physical constraints of displacement change.

[0037] The pressure platform duration signal is differentially calculated with the dynamic reference pressure curve, and a threshold comparison is used to eliminate minor disturbances caused by sensor quantization errors, ensuring the accuracy of steady-state range identification.

[0038] Through the above-mentioned synchronization timestamp alignment and multi-domain filtering, the initial multi-source sensor raw signals are mapped into high signal-to-noise ratio data vectors under a unified time base, realizing the clean data source required for subsequent velocity feature extraction and state determination.

[0039] S1.2: Based on the original high signal-to-noise ratio working condition data stream, the first derivative of the displacement change rate is calculated using sliding window differential processing as the instantaneous velocity feature, and a threshold comparator is used to determine the steady-state interval of the pressure platform duration signal, so as to output a set of key feature events including the coordinates of the velocity change point and the start and end times of the pressure steady state.

[0040] Based on the high signal-to-noise ratio (SNR) raw operating condition data stream, a sliding window index set for calculating the displacement change rate is constructed. The displacement slope signal measured by a grating ruler is used as the reference signal input, ensuring that the window size matches the sampling frequency to balance time resolution and smoothing performance. A difference operation is performed on the displacement slope data within the window, and the first derivative of the displacement change rate is calculated using the mathematical formula:

[0041] Where Δs is the displacement slope value and Δt is the sampling time interval, the instantaneous velocity feature sequence is obtained. Local extremum detection is performed on this instantaneous velocity feature sequence. Velocity abrupt change points are extracted by setting a velocity change threshold, and their coordinate positions on the time axis are recorded. The pressure platform duration signal is input into a threshold comparator. Based on the process-preset steady-state pressure value and allowable fluctuation range, the steady-state range is determined using the following formula:

[0042] Where p is the real-time pressure value, p ref Here, δ represents the steady-state reference pressure value, and δ represents the allowable steady-state fluctuation value. The intervals determined to be in steady state are output with their start and end times, and these are combined with the coordinates of velocity abrupt change points to form a set of key feature events. Through the aforementioned differential and threshold determination processing method, the high signal-to-noise ratio raw operating condition data stream from the previous step is transformed into a structured set of key feature events containing the coordinates of velocity abrupt change points and the start and end times of the steady-state pressure, thus preparing supporting data for subsequent phase state machine mapping.

[0043] S1.3: Based on the coordinates of the velocity mutation point and the start and end times of the pressure steady state in the set of key feature events, apply the predefined four-stage state machine transition rules of the forging process to perform logical mapping operations, so as to identify the current state of downlink start, pressurization steady state, return acceleration or braking stop and generate an initial process stage identification sequence.

[0044] Based on the coordinates of velocity mutation points and the start and end times of pressure steady state in the set of key feature events, predefined four-stage state machine transition rules for forging process are loaded as the basis for logical mapping.

[0045] For the coordinate sequence of velocity mutation points, stage boundary location calculations are performed, and the time difference between adjacent mutation points is cross-matched with the steady-state pressure interval to identify potential boundary candidate points at the start and end of the stroke.

[0046] By using the conditional expressions in the state machine transition rules, the mutation point features and pressure range labels are mapped to the state nodes of the four process stages, and candidate state paths including downlink start-up, pressurization steady state, return acceleration, and braking stop are constructed.

[0047] Logical consistency verification is performed based on the candidate state paths. The stroke direction determiner is used to perform symbolic comparison of the relationship between the velocity sign and the displacement slope sign, and illegal transfer nodes that violate the stage order are eliminated.

[0048] Based on the verified state path and corresponding time coordinates, an initial process stage identifier sequence is generated. Each stage identifier is arranged in chronological order and accompanied by a node timestamp for subsequent boundary smoothing and jump correction modules to call.

[0049] Through the above mapping and verification process, the velocity mutation point and pressure steady-state range of the previous step are transformed into initial process stage identification data with clear stage semantics, thereby realizing the structured generation of slider motion phase labels.

[0050] S1.4: Perform boundary smoothing and illegal state transition correction operations on the initial process stage identifier sequence to eliminate phase misjudgment caused by sensor jitter, so as to generate a corrected process stage identifier sequence with time continuity and logical consistency.

[0051] S1.5: The corrected process stage identifier sequence is fused and encapsulated with the original sampling time axis, and a precise timestamp is added to each stage identifier to finally generate a slider motion phase label sequence with timestamp as the output of this step.

[0052] Step S2: Based on the dynamic boundary characteristics of each process stage in the slider motion phase label sequence, a piecewise kinematic envelope constraint set is constructed. This piecewise kinematic envelope constraint set is defined as a pre-defined set of engineering-verifiable rules for each phase stage. Specifically, the engineering-verifiable rule set includes rotational speed monotonicity, acceleration non-negativity, fluctuation amplitude threshold, and convergence continuity requirements. S2.1: Obtain the down-start stage identifier and the pressurization steady-state stage identifier from the slider motion phase label sequence, and extract the corresponding speed change rate threshold and load disturbance tolerance parameters based on the dynamic equation of the forging press mechanical transmission system, and generate a primary stage dynamic boundary feature set containing the non-negativity constraint condition of the down-start stage acceleration and the upper limit value of the speed fluctuation amplitude in the pressurization stage.

[0053] Based on the downlink start-up phase identifier and the pressurized steady-state phase identifier in the slider motion phase label sequence, the phase identifier parsing module is called to extract the corresponding timestamp index interval, and a phase data index table for dynamic calculation is constructed.

[0054] Using the speed trajectory data during the downward start-up phase as input, the instantaneous speed change rate curve is calculated based on the dynamic equation of the forging press mechanical transmission system, and the peak and average speed change rate are identified to form a statistical feature set of speed change rate.

[0055] Based on the peak value and the average value of the steady-state first phase in the statistical feature set of the rate of change of velocity, a non-negativity constraint condition for acceleration in the downshift phase is set, and the threshold of the rate of change of velocity is determined by the formula:

[0056] Where Δv is the velocity increment and Δt is the corresponding time increment, this condition is used to determine the dynamic rationality of the downhill phase.

[0057] The speed trajectory data during the pressurized steady-state phase is matched with the pressure signal during the same phase. After eliminating high-frequency noise using bandpass filtering, the speed fluctuation amplitude is calculated using the extreme value difference formula:

[0058] Where max(v) and min(v) are the maximum and minimum speeds during the steady-state phase, respectively, thus yielding the speed fluctuation amplitude A during this phase.

[0059] By comparing the fluctuation amplitude A with the load disturbance tolerance parameter, an upper limit constraint condition for the speed fluctuation amplitude during the pressurization stage is generated to ensure speed stability during this stage.

[0060] A set of dynamic boundary features for the initial stage is established by combining the threshold of the rate of change of velocity with the upper limit of the fluctuation amplitude, providing basic data for the subsequent construction of full-cycle constraints.

[0061] Through the above processing method, the stage identifier and transmission system dynamic parameters are transformed into executable speed change rate thresholds and fluctuation amplitude upper limits, realizing the construction of the primary dynamic boundary feature set, and providing engineering verifiable data basis for subsequent constraint rule definition and physical feasibility verification.

[0062] S2.2: Utilizing the velocity change rate threshold and load disturbance tolerance parameters in the initial stage dynamic boundary feature set, combined with the return acceleration stage identifier and braking parking stage identifier, the minimum starting speed limit of the return stage and the deceleration continuous decay curve model of the braking stage are calculated through the inverse kinematics algorithm, generating a complete set of stage dynamic boundary feature vectors covering the four stages of the entire cycle.

[0063] Based on the velocity change rate threshold and load disturbance tolerance parameters in the initial stage dynamic boundary feature set, and combined with the return acceleration stage and braking / stopping stage identifiers in the slider motion phase label sequence, the calculation conditions for the minimum starting speed limit in the return stage are constructed. For the return stage, the inverse kinematics algorithm is invoked, and based on the drive system moment of inertia and load disturbance tolerance, the velocity change rate threshold is inversely derived to the starting speed limit. Initial condition constraints are established through the differential form of the displacement-time relationship. In the braking / stopping stage, the initial boundary values ​​of the deceleration continuous decay curve model are established using the combined parameters of the velocity change rate threshold and load disturbance tolerance. The deceleration law curve is inversely solved from the known velocity-time function to ensure that the velocity convergence logic is satisfied. The following inverse kinematics formula maps the velocity change rate threshold to the starting speed limit:

[0064] Where v0 is the lower limit of the return stroke starting speed, v1 is the target steady-state speed during the return stroke, and a m Let Δt be the threshold value for the rate of change of velocity, and Δt be the initial acceleration time span. For the braking phase, a fitting expression for the continuous deceleration curve model is established:

[0065] Where a(t) is the instantaneous deceleration at time t, a0 is the initial deceleration during braking, and k is the attenuation coefficient, obtained through fitting real-time operational data. The derivation results of the return and braking phases are combined to form a complete set of dynamic boundary feature vectors covering the four phases of downhill initiation, pressurized steady state, return acceleration, and braking parking, while maintaining numerical and logical consistency of the indicators for each phase. Through the aforementioned inverse kinematics solution and curve fitting processing, the initial phase dynamic boundary feature set results from the previous step are transformed into a complete set of dynamic boundary feature vectors for the entire cycle, achieving full-stage coverage of the physical constraints of the operating conditions.

[0066] S2.3: Based on the non-negativity constraint of acceleration, the upper limit of speed fluctuation amplitude, the lower limit of minimum starting speed and the continuous decay curve model of deceleration in the complete stage dynamic boundary feature vector group, the symbolic logic rule mapping method is used to define the speed monotonicity discriminant and convergence criterion for each stage, and generate a standardized kinematic envelope constraint rule library for each phase stage.

[0067] Based on the non-negativity of acceleration constraints, the upper limit of rotational speed fluctuation, the lower limit of minimum starting rotational speed, and the continuous decay curve model of deceleration in the complete stage dynamic boundary feature vector group, the above feature parameters are loaded as input objects into the symbolic logic rule mapping processing module to establish the mapping relationship between physical constraints and logical discriminant expressions.

[0068] A logical symbolization operation is performed on the non-negativity constraint of acceleration to construct a monotonicity criterion based on the instantaneous acceleration sequence. The criterion is defined as follows: if the difference between any two adjacent discrete speed sampling points is less than zero, the constraint is considered violated. This operation relies on the difference operator to calculate the speed sampling sequence point by point. The difference formula is as follows:

[0069] Where a is the instantaneous acceleration, v(t) is the rotational speed at the current sampling point, and Δt is the sampling interval.

[0070] A symbolic logical mapping is performed on the upper limit parameter of the speed fluctuation amplitude, and a comparison rule is defined between the absolute value of the extreme value difference and the preset fluctuation threshold. The construction form is as follows:

[0071] Among them, v max With v min These are the maximum and minimum rotational speeds of the subsequence in this stage, respectively, and ΔV is the preset upper limit of the fluctuation amplitude.

[0072] A thresholding logic mapping is performed on the minimum starting speed lower limit, defining a comparison rule between the starting speed value and the lower limit value, in the form of:

[0073] Among them, v start V represents the starting speed value for the current stage. min This is the minimum starting speed allowed for this stage.

[0074] A trend symbolic mapping is performed on the continuous deceleration curve model, and a criterion is defined for the final acceleration sequence to have a strictly decreasing absolute value and converge to zero. The criterion formula is as follows:

[0075] Where T is the end time of the phase, and a(t) is the instantaneous acceleration value at time t.

[0076] Through the above symbolic logic rule mapping method, all physical constraints are transformed into rotational speed monotonicity discriminants and convergence criteria with executable symbolic comparison and logical operation capabilities. These are then stored in the standardized kinematic envelope constraint rule library in stages, thereby achieving the structuring and unification of multi-stage physical constraints.

[0077] By mapping symbolic logic rules, the complete stage dynamic boundary feature vector set of the previous step is transformed into a set of executable discrimination rules for each phase stage, thereby realizing the ability to determine the physical consistency of the predicted speed curve in different process stages.

[0078] S2.4: Perform engineering verifiability quantification on the speed monotonicity discriminant and convergence criterion in the standardized kinematic envelope constraint rule base, transform the qualitative features into an executable verification instruction set containing numerical comparison operators and logical state flags, and generate the final piecewise kinematic envelope constraint set to support the symbolic logical comparison operation of subsequent sub-speed curve segments.

[0079] Using the rotational speed monotonicity discriminant and convergence criterion from the standardized kinematic envelope constraint rule base generated by the preceding sub-steps as input objects, the rule base index interface is called to extract each constraint formula containing qualitative description, and it is split into three sets of elements: discriminant variables, relational operators, and logical conditions.

[0080] The physical quantities involved in the discriminant variable set, such as the rate of change of rotational speed, rotational speed, and acceleration continuity, are uniformly mapped to a combination of sampling data index and calculation function. During the mapping process, a unit system and magnitude coefficient are added to each physical quantity to ensure the consistency of the physical dimensions in subsequent numerical comparisons.

[0081] Perform symbol-to-operator encoding conversion on the set of relational operators, converting logical relations such as "greater than", "less than", "equal to", and "approaching" into machine-readable operator identifiers, and construct differential threshold comparison operators for continuous conditions.

[0082] Based on the stage characteristics in the logical condition set, a condition trigger middleware is constructed to numerically define qualitative descriptions such as "strictly monotonically increasing" and "converging to zero in the final stage". For example, a judgment rule is generated for the instantaneous acceleration polarity of the entire sequence to be always non-negative for monotonicity constraints, and a fitting accuracy threshold rule is generated for the final stage speed slope to approach zero for convergence constraints.

[0083] The symbolic syntax tree parsing method is adopted to assemble the mapped discriminant variables, encoding operators, and numerical conditions into a structured executable verification instruction set. Stage flags and constraint number identifiers are added to the syntax tree nodes, thereby transforming the original qualitative features into a final piecewise kinematic envelope constraint set with numerical comparison operators and logical state flags.

[0084] Through the above-mentioned engineering verifiability quantification process, the rule base output of the previous step is transformed into an executable set of verification instructions, thereby achieving the effect of parameterized rule application that supports subsequent symbolic logic comparison of sub-speed curve segments.

[0085] like Figure 2 As shown, step S3: Based on the time segmentation points of the slider motion phase label sequence, the original predicted rotational speed time sequence output by the preset deep learning model is divided into sub-rotational speed curve segments corresponding to the process stage.

[0086] Specifically, the original predicted rotational speed time series output by the deep learning model is used as input. Based on the time segmentation points of the slider motion phase label sequence, the original predicted rotational speed time series is divided into sub-rotational speed curve segment sequences corresponding to the four process stages. This includes: S3.1: Obtain the original predicted rotational speed time series and slider motion phase label sequence output by the deep learning model. Based on the timestamps contained in the slider motion phase label sequence, extract the start and end times of each of the four process stages: downlink start, pressurized steady state, return acceleration, and braking stop, and generate a process stage time window set containing four sets of time boundary parameters.

[0087] In this embodiment, the deep learning model adopts a multi-task temporal regression model architecture based on a bidirectional gated recurrent unit (Bi-GRU). The model construction process specifically includes: First, feature encoding is performed on the preprocessed historical process parameter data, mapping the slider motion phase label sequence and the corresponding time series to a high-dimensional tensor that the model can recognize as input features; Second, deep feature extraction is performed in the Bi-GRU network layer, using its unique gating mechanism to capture the nonlinear dynamic evolution law and long-range time dependence of each process stage during the slider motion, namely, downhill start-up, pressurized steady state, return acceleration, and braking parking, and outputting the temporal feature vector of the hidden layer; Finally, through a fully connected regression layer connected to the back end of the Bi-GRU layer, the temporal feature vector is mapped to continuous time series values ​​to generate the final original predicted speed time series.

[0088] S3.2: Based on the start and end time points in the set of process stage time windows, perform a time-series indexing operation on the original predicted speed time series, calculate the discrete data point index range corresponding to each process stage time window in the original predicted speed time series, and generate a process stage data mapping table containing four sets of index intervals.

[0089] The original predicted rotational speed time series output by the deep learning model is used as the input for index localization. Combined with the start and end time points in the process stage time window set, each set of boundary parameters in the time window set is first converted into absolute sampling position information on the time axis according to a unified sampling frequency benchmark, so as to establish a one-to-one mapping relationship in the original time series. Interpolation correction is performed on the converted absolute sampling position information, using linear or polynomial interpolation methods to correct the time boundary point deviation caused by non-divisible sampling frequency, ensuring that the localization result is consistent with the time resolution of the actual physical process. The corrected absolute sampling position information is matched with the time index matrix of the original predicted rotational speed time series. A threshold tolerance range judgment method is used to determine the index value of each start and end time point in the time series, and this index value is recorded in the initial draft of the stage index interval. Continuity detection is performed on the initial draft of the stage index interval, using logical judgment to remove abnormal segments with index jumps or overlaps, and the index interval boundaries are adjusted according to the detection results to ensure that the interval division of each process stage in the time series does not overlap and covers the global time axis. After continuous detection and boundary adjustment, the four sets of index intervals are arranged in the order of process stages. The output is a process stage data mapping table containing four sets of index ranges: downlink start, pressurized steady state, return acceleration, and braking stop. Through index positioning and boundary adjustment processing, the time window set of the previous step is transformed into discrete index intervals that can be directly used for data extraction, so as to achieve a precise correspondence between the original speed time series and the process stages.

[0090] For example, in a speed control scenario for a forging press slide drive motor, the deep learning model outputs a predicted rotational speed time series with a sampling frequency of 1000Hz and a total duration of 8 seconds, resulting in 8000 discrete rotational speed data points. The process stage time window set contains four sets of boundary parameters: the starting time of the down-draft start-up stage is 0.00 seconds, and the ending time is 1.85 seconds; the starting time of the pressurization steady-state stage is 1.85 seconds, and the ending time is 3.50 seconds; the starting time of the return acceleration stage is 3.50 seconds, and the ending time is 6.20 seconds; the starting time of the braking and parking stage is 6.20 seconds, and the ending time is 8.00 seconds. Each start and end time is multiplied by the sampling frequency to obtain an initial index value; for example, the starting index for the down-draft start-up is 0, and the ending index is 1850. Polynomial interpolation is used to correct the time boundary deviation caused by sensor delay, adjusting the down-draft start-up and ending indexes to 1852. Continuity detection reveals two overlapping points between the ending index of the return acceleration stage and the starting index of the braking stage; after removing the overlapping part, the starting index of the braking and parking stage is adjusted to 6201. The final generated data mapping table is as follows: Downward start [0, 1852], Pressurized steady state [1853, 3500], Return acceleration [3501, 6200], Braking stop [6201, 8000]. Using this table to extract data from each stage of the predicted rotational speed time series ensures accurate correspondence between the physical process time interfaces, significantly improving the accuracy of subsequent envelope constraint verification.

[0091] S3.3: Using the discrete data point index range in the process stage data mapping table, extract the speed value segments within the corresponding index interval from the original predicted speed time series, and construct the down-start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence respectively, generating an initial sub-speed curve segment set containing the down-start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence.

[0092] The discrete data point index range in the process stage data mapping table is used as the index retrieval parameter input into the original predicted speed time series. The speed data of the corresponding index interval is accurately truncated using time series slicing operation to ensure that the truncated range is completely consistent with the time boundary parameter.

[0093] The extracted speed data segments within the downlink start-up phase index range are arranged in chronological order to form a downlink start-up speed subsequence containing continuous sampling points. The original predicted values ​​are retained to support subsequent physical envelope comparison.

[0094] Arrange the extracted speed data segments within the index range of the pressurized steady-state stage in chronological order to form a pressurized steady-state speed subsequence containing continuous sampling points, and ensure that the start and end points of this subsequence are completely aligned with the pressure steady-state interval.

[0095] The extracted rotational speed data segments within the index range of the return acceleration stage are arranged in chronological order to form a return acceleration rotational speed subsequence containing continuous sampling points, and features such as the initial speed of this stage are retained for subsequent stage dynamics determination.

[0096] The extracted speed data segments within the index range of the braking and parking phase are arranged in chronological order to form a braking and parking speed subsequence containing continuous sampling points, ensuring that the final data segment covers the entire deceleration and convergence process.

[0097] By using the above-mentioned partitioning and subsequence construction processing method, the process stage data mapping table is transformed into a set of initial sub-speed curve segments containing four independent elements: downlink start-up, pressurized steady state, return acceleration, and braking stop, thereby realizing the structured decomposition and stage data binding of the original predicted speed time series.

[0098] For example, in the movement cycle of the forging press slide, the deep learning prediction outputs a rotational speed time series with 2000 sampling points and a sampling interval of 2ms. In the process stage data mapping table, the index range for the downlink start-up stage is defined as 0 to 450, the index range for the pressurization steady-state stage is defined as 451 to 900, the index range for the return acceleration stage is defined as 901 to 1550, and the index range for the braking and stopping stage is defined as 1551 to 2000. In the execution step, the original rotational speed time series is truncated from 0 to 450 to generate a downlink start-up rotational speed subsequence with a length of 451 points; the range from 451 to 900 is truncated to generate a pressurization steady-state rotational speed subsequence with a length of 450 points; the range from 901 to 1550 is truncated to generate a return acceleration rotational speed subsequence with a length of 650 points; and the range from 1551 to 2000 is truncated to generate a braking and stopping rotational speed subsequence with a length of 450 points. Using a rotational speed numerical matrix slicing operation, the four subsequences are encapsulated according to their index ranges, and their respective timestamp information is retained. In mathematical verification, the average initial velocity of the subsequence in the retrace acceleration phase is calculated using the following formula:

[0099] Among them, v i Let be the velocity value of the i-th sampling point within the initial index interval, and n be the total number of sampling points. The average initial velocity value in this stage is significantly higher than the average velocity over the entire cycle, indicating that the prediction model has a significant acceleration characteristic in this stage. The final output set of initial sub-rotation speed curve segments demonstrates good performance in subsequent physical envelope constraint verification, exhibiting continuous data, accurate stage boundaries, and extractable physical features.

[0100] S3.4: Perform phase attribute binding processing on the downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence in the initial sub-speed curve segment set to generate downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence with phase semantic tags.

[0101] The step of performing phase attribute binding processing on the downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence in the initial sub-speed curve segment set includes: appending the process stage identifier of the downlink start as metadata to the header of the downlink start speed subsequence; appending the process stage identifier of the pressurized steady-state speed subsequence as metadata to the header of the pressurized steady-state speed subsequence; appending the process stage identifier corresponding to the return acceleration as metadata to the header of the return acceleration speed subsequence; and appending the process stage identifier corresponding to the braking stop as metadata to the header of the braking stop speed subsequence.

[0102] The initial set of sub-speed curve segments is used as the input object for this step, and a phase attribute binding process is established for each independent element in the set. The process stage identifier in the slider motion phase label sequence is read, and the process stage semantic information of the current speed sub-sequence is established based on the time boundary parameter index corresponding to the identifier. The metadata construction module is called to encapsulate the acquired process stage identifier into a structured metadata field, and this field is appended to the beginning of the speed sub-sequence to form a curve segment data unit with clear phase semantics. A data structure integrity check is performed on the speed sub-sequence after appending metadata to verify the consistency between the metadata field and the time axis index, preventing label misalignment or semantic binding errors. A symbolic mapping method is used to map different process stage identifiers to standard phase semantic labels. For example, the identifier corresponding to the downlink start-up stage is mapped to "phase label-DS", the pressurized steady-state stage to "phase label-HP", the return acceleration stage to "phase label-RA", and the braking and parking stage to "phase label-BS", ensuring that the subsequent verification module can directly call the label semantic execution logic for judgment. The above processing method transforms the initial sub-speed curve segment set output in the previous step into a structured speed sub-sequence with phase semantic labels, realizing bidirectional binding between data and process stages, and providing accurate semantic indexes for subsequent segmented kinematic envelope constraint verification.

[0103] For example, in the working cycle of a forging press, the initial sub-speed curve segment set contains four independent elements: the first time segment spans from 0.00 to 1.25 seconds, the second time segment spans from 1.25 to 2.80 seconds, the third time segment spans from 2.80 to 4.10 seconds, and the fourth time segment spans from 4.10 to 5.00 seconds. The identifiers of the slider motion phase label sequence in the corresponding time intervals are downhill start, pressurized steady state, return acceleration, and braking stop, respectively. The start time index of the first segment is retrieved and matched with the label sequence. The identifier "downhill start" is encapsulated as a metadata field and appended to the head of the first segment curve; the second segment is appended with the label "pressurized steady state", the third segment with the label "return acceleration", and the fourth segment with the label "braking stop". Symbolic mapping converts the downhill start label to "phase label-DS", the pressurized steady state to "phase label-HP", the return acceleration to "phase label-RA", and the braking stop to "phase label-BS". In the integrity check, the alignment of the labels with the time span is verified to be without deviation. Through strict semantic binding, the subsequent physical feasibility verification module can perform envelope boundary comparison of the corresponding process stage for each curve according to the phase label. The final output results show that the accuracy of stage binding is significantly improved and the call latency of the verification module is greatly reduced.

[0104] S3.5: Based on the downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking parking speed subsequence with phase semantic tags, the data structure is reorganized in chronological order and encapsulated to form a standard sub-speed curve segment sequence. This object is used as the final output and directly supplied to the subsequent piecewise kinematic envelope constraint set verification module.

[0105] like Figure 3 As shown, step S4: The segmented kinematic envelope constraint set is used to perform a symbolic logic comparison operation on the sub-speed curve segment sequence to determine whether each sub-speed curve segment satisfies the envelope boundary and qualitative characteristics of the corresponding process stage, and a physical feasibility verification result is generated. The physical feasibility verification result includes: a phase matching degree index and a constraint violation flag. Specifically, it includes: S4.1: Based on the segmented kinematic envelope constraint set, extract the dynamic boundary threshold set of the current process stage, and take the discrete speed sampling points in the sub-speed curve segment sequence as input objects. Calculate the instantaneous acceleration numerical sequence between adjacent discrete speed sampling points through differential operation to obtain the instantaneous acceleration numerical sequence characterizing the speed change rate.

[0106] S4.2: Perform sign polarity comparison processing using the instantaneous acceleration numerical sequence and the acceleration non-negativity requirement in the dynamic boundary threshold set, and execute monotonicity determination logic for the sub-speed curve segments of the downlink start-up phase and the return acceleration phase to generate a phase monotonicity determination result containing a monotonicity compliance status identifier.

[0107] The instantaneous acceleration numerical sequence is compared with the acceleration non-negativity constraint conditions in the dynamic boundary threshold set by performing sign polarity comparison. The input objects are the instantaneous acceleration numerical sequence calculated according to the previous step S4.1 and the pre-extracted stage dynamic boundary threshold set.

[0108] Based on the symbolic logic rules, the symbolic attribute corresponding to each sampling point in the instantaneous acceleration numerical sequence is extracted as a positive, zero, or negative polarity label, and a point-by-point matching operation is performed with the symbol requirements in the non-negativity constraint of acceleration.

[0109] Scan all sampling points of the sub-speed curve segment during the downlink start-up phase and the sub-speed curve segment during the return acceleration phase, and check whether there are any illegal negative values ​​in the symbol labels. If a continuous negative value interval is detected, record the illegal event index.

[0110] For each sampling point that passes symbol matching, the proportion is statistically analyzed and the stage compliance metric is calculated using the following formula:

[0111] Where, N pos The number of sampling points required to satisfy the non-negativity condition, N is the total number of sampling points in this stage.

[0112] Based on the numerical value of the phased compliance metric, a monotonic compliance status identifier is constructed. When R reaches the preset compliance threshold, it is marked as compliant; otherwise, it is marked as non-compliant.

[0113] The stage compliance status identifier is attached to the metadata of the corresponding downlink start-up stage and back-end acceleration stage sub-speed curve segments to form a complete stage monotonicity determination result.

[0114] Through the above symbol polarity comparison processing and threshold judgment logic, the instantaneous acceleration sequence of the previous step is transformed into a data structure containing quantitative compliance indicators and binary state flags, thereby realizing the judgment that the monotonicity of the rotational speed curves at each stage conforms to physical constraints.

[0115] For example, during the return acceleration phase of a forging press, the sampling frequency is configured to 1000Hz, with a total of 500 sampling points. The non-negativity condition for acceleration in the dynamic boundary threshold is set to ≥0 m / s². The instantaneous acceleration sequence is obtained using differential operations, with a maximum value of 2.4 m / s² and a minimum value of -0.1 m / s². After sign polarity comparison processing, the number N of sampling points satisfying the non-negativity condition is counted. pos =495, substituting into the formula, the compliance metric value R=0.99 is calculated. The compliance threshold is set to 0.98, the judgment result is compliant, and the monotonicity status flag is set to 1. In this embodiment, the output stage monotonicity judgment result shows that the return stage speed curve fully conforms to the dynamic non-negative acceleration constraint, effectively ensuring the accuracy of the physical feasibility verification.

[0116] S4.3: Based on the sub-speed curve segments marked as compliant in the stage monotonicity determination results, the fluctuation amplitude threshold rule is called to perform extreme value difference calculation on the sub-speed curve segments in the pressurized steady state stage. By comparing the magnitude relationship between the maximum speed deviation value and the preset fluctuation amplitude threshold, the fluctuation amplitude compliance flag bit characterizing speed stability is obtained.

[0117] Based on the sub-speed curve segments marked as compliant in the monotonicity determination results of the aforementioned stage, the execution object of the fluctuation amplitude threshold rule is the speed subsequence of the pressurized steady-state stage. All discrete sampling point values ​​in this speed subsequence are used as input for extreme value calculation, and the maximum and minimum speed values ​​of the subsequence are obtained through an extreme value extractor. The amplitude difference is calculated using a difference calculation module to obtain a value characterizing the speed fluctuation range of this stage. A symbolic logic comparator is used to determine the magnitude relationship between this amplitude difference value and a preset fluctuation amplitude threshold for the pressurized stage. The preset threshold is derived from the stability boundary conditions defined in the piecewise kinematic envelope constraint set. A non-compliant state flag is generated for differences greater than the threshold, and a compliant flag is generated for differences less than or equal to the threshold. Through the above difference calculation and logic comparison processing, the monotonicity compliant sub-segment of the previous step is transformed into a fluctuation amplitude compliance flag with stability qualitative characteristics, realizing the quantitative verification of the predicted speed stability in the pressurized steady-state stage.

[0118] For example, in one forging cycle, a subsequence of rotational speeds during the steady-state pressurization phase is obtained with a length of 200 sampling points and a sampling period of 5ms. The maximum rotational speed is 1520 rpm, and the minimum rotational speed is 1485 rpm. The fluctuation amplitude threshold is set to 50 rpm. The amplitude difference is calculated using differential calculation: 1520-1485 yields a difference of 35 rpm. A comparator is used to determine the order of magnitude of the difference and the threshold, using the formula: 35≤50. Since the determination result is true, a fluctuation amplitude compliance flag is generated. If, under another operating condition, the maximum rotational speed is 1555 rpm and the minimum is 1480 rpm, the difference is 75 rpm obtained using the differential formula 1555-1480. The result of the formula 75≤50 is determined to be false, and a non-compliant status flag is generated. In the two scenarios described above, the compliance flag for the first scenario will directly enter the convergence continuity verification, significantly improving the smooth operation performance during the verification phase; the non-compliance flag for the second scenario will be marked as an abnormal segment in subsequent steps and enter the remapping repair, thereby ensuring the physical consistency of the control command sequence.

[0119] S4.4: Based on the aforementioned fluctuation amplitude compliance flag and convergence continuity requirements, perform a final slope decay trend analysis on the sub-speed curve segment during the braking and parking phase. By fitting the tangent slope parameter of the final speed trajectory and determining whether it strictly converges to zero, a convergence continuity verification index characterizing the smoothness of the braking process is generated.

[0120] For the sub-speed curve segment during the braking and parking phase, the previously generated fluctuation amplitude compliance flag and convergence continuity requirement are used as input conditions to organize the speed trajectory data of the final segment of this phase into a discrete sampling sequence in chronological order. Based on this discrete sampling sequence, the tangent slope of the final segment interval is calculated. The speed difference between adjacent discrete speed sampling points is divided by the corresponding time difference to generate the instantaneous slope change sequence of the final segment. Linear least squares fitting is applied to the instantaneous slope change sequence to extract the slope parameter of the fitted equation. The absolute difference between the slope parameter and the zero value target in the convergence continuity requirement is calculated, i.e., the degree of deviation. Based on the deviation value and the preset continuity threshold, the judgment logic is executed to generate a convergence continuity verification index, which is then attached to the verification result data structure of the current stage. Through linear fitting and deviation judgment processing, the fluctuation amplitude compliance flag and the final slope change sequence of the previous step are transformed into a quantitative convergence continuity verification index, realizing the physical feasibility verification of the smoothness of the braking process.

[0121] S4.5: Combining the monotonicity judgment results of the aforementioned stages, the compliance flag of the fluctuation amplitude, and the convergence continuity verification index, a weighted aggregation process is used to calculate the degree of deviation between each sub-speed curve segment and the ideal physical behavior model, so as to output a physical feasibility verification result that includes a quantized phase matching degree index and a binary constraint violation flag.

[0122] In this embodiment, the ideal physical behavior model refers to the standard speed change reference benchmark constructed based on the mechanical kinematics principle and ideal power transmission characteristics of the forging press slide for each process stage (downward start-up, pressurized steady state, return acceleration, and braking stop). This model is not a single mathematical formula, but a comprehensive physical behavior map formed by integrating the requirements that each stage must meet, namely, speed monotonicity, acceleration non-negativity, fluctuation amplitude threshold, and convergence continuity.

[0123] Specifically, during the downhill start-up and return acceleration phases, the model exhibits a smooth upward curve with strictly monotonically increasing acceleration and no negative acceleration; during the pressurized steady-state phase, it exhibits a horizontal straight line with slight fluctuations around the target speed and meeting preset amplitude limits; and during the braking and parking phase, it exhibits a smooth downward trajectory with a continuously decreasing slope that eventually converges strictly to zero speed.

[0124] Based on the stage monotonicity judgment results, fluctuation amplitude compliance flags, and convergence continuity verification indicators generated in the previous sub-steps, a dataset containing three types of physical consistency features is constructed as the input object for this sub-step. This dataset is grouped according to process stages, and a weighted vector is constructed for the three features of each stage. The weight coefficients are assigned based on the previous experimental calibration results and the importance of energy efficiency and stability of the process stage throughout the entire cycle. Linear weighted aggregation is used to perform numerical fusion on the physical consistency features of each stage, forming a quantitative index vector characterizing the deviation of the sub-speed curve segment from the ideal physical behavior model. The phase matching degree index is calculated according to the formula:

[0125] Among them, w i p represents the weight coefficient of the i-th feature. i This determines the compliance score for this feature. The matching degree index is numerically compared to a preset physical consistency threshold. If the matching degree is lower than the threshold, a constraint violation flag is set to 1 in the output; otherwise, it is set to 0. The matching degree indices and constraint violation flags from all stages are combined and encapsulated into a complete physical feasibility verification result data structure, and then passed to the anomaly repair module in the next step via a data interface. Through weighted aggregation and threshold judgment processing, the results of the previous step are transformed into executable engineering data containing phase matching degree indices and binary constraint violation flags, enabling a quantitative assessment of the physical feasibility of the predicted rotational speed curve.

[0126] Specifically, for the downlink start-up and return acceleration phases: if the phase monotonicity determination result is non-compliant, the corresponding sub-speed curve segment is marked as an abnormal sub-speed curve segment, and its violation of speed monotonicity constraints is recorded; otherwise, it is marked as a compliant sub-speed curve segment.

[0127] For the pressurized steady-state stage: if the compliance flag of its fluctuation amplitude is not met, the sub-speed curve segment corresponding to this stage is marked as an abnormal sub-speed curve segment, and its violation of the fluctuation amplitude threshold constraint is recorded; otherwise, it is marked as a compliant sub-speed curve segment.

[0128] For the braking and parking phase: if the convergence continuity verification index is not met, the sub-speed curve segment corresponding to this phase is marked as an abnormal sub-speed curve segment, and its violation of the convergence continuity requirement is recorded; otherwise, it is marked as a compliant sub-speed curve segment.

[0129] Step S5: For the abnormal sub-speed curve segments marked as constraint violations in the physical feasibility verification results, retrieve the nearest operating condition template based on the preset phase typical template library, and remap the abnormal segments to the corresponding positions of the nearest operating condition templates to generate reconstructed speed curve segments that conform to the physical behavior logic. Specifically, the abnormal segments are remapped to the corresponding positions of the nearest operating condition templates through scaling, translation, or truncation operations. This includes: S5.1: Obtain the constraint violation flag and the corresponding abnormal sub-speed curve segment from the physical feasibility verification result. Based on the constraint violation flag, extract the speed data block to be corrected containing phase semantic tags, and perform feature vector standardization processing on the speed data block to be corrected to generate a normalized abnormal operating condition feature vector for template retrieval.

[0130] S5.2: Based on the normalized abnormal working condition feature vector, call the preset phase typical template library to perform multidimensional Euclidean distance similarity matching processing, calculate the distance metric value between the normalized abnormal working condition feature vector and the feature vector of each standard working condition template in the phase typical template library, and select the nearest working condition template with the smallest distance metric value as the benchmark reference object.

[0131] In this embodiment, the phase typical template library is a pre-built reference database used to store and retrieve slider motion characteristics under standard operating conditions. Specifically, it is constructed as follows: based on samples marked as compliant sub-speed curve segments from a large amount of historical operating data, classification and clustering analysis are performed according to the four process stages (downward start-up, pressurized steady state, return acceleration, and braking stop). Typical speed change patterns under ideal physical behavior in each stage are extracted and solidified to form corresponding standard operating condition templates. Each standard operating condition template is stored in the form of a feature vector.

[0132] The normalized abnormal working condition feature vector is used as the input object for similarity matching, and the feature vector set of each standard working condition template in the pre-set phase typical template library is loaded to form a template feature matrix with uniform scale and dimension.

[0133] Element-wise difference operations are performed on the normalized abnormal working condition feature vector and each standard working condition template feature vector in the template feature matrix, and the numerator of the Euclidean distance is obtained by accumulating the sum of squares.

[0134] Taking the square root of the sum of squared differences above, a list of Euclidean distance metrics between the normalized abnormal working condition feature vectors and the feature vectors of each standard working condition template is generated, expressed by the formula:

[0135] Where x is a certain dimension value of the normalized abnormal working condition feature vector, t is the corresponding dimension value of the template feature vector, and the summation symbol represents the accumulation over all feature dimensions.

[0136] Perform a minimum value search operation on the list of Euclidean distance metrics to determine the template index number with the smallest distance metric value, and select the corresponding standard working condition template as the nearest working condition template.

[0137] The selected nearest working condition template is used as the benchmark reference object to provide a standard trajectory reference for the calculation of geometric transformation parameters such as subsequent scaling and time axis translation.

[0138] Through the above similarity matching processing method, the normalized abnormal operating condition feature vector of the previous step is transformed into a set of template matching information with the minimum distance metric value, so as to realize a high-precision correspondence between the abnormal sub-speed curve segment and the historical standard operating condition.

[0139] S5.3: Using the standard speed trajectory data in the nearest working condition template and the time span parameter of the abnormal sub-speed curve segment, construct a linear affine transformation matrix, and solve the scaling factor and time axis translation offset by the least squares method to generate a remapping control parameter set containing geometric transformation parameters.

[0140] The time span parameter refers to the effective coverage area of ​​the abnormal sub-speed curve segment on the time axis, i.e., the difference between the start and end timestamps, which serves as the benchmark for subsequent coordinate system alignment and time axis translation calculations. The standard speed trajectory data in the nearest neighbor operating condition template specifically refers to the specific data instance corresponding to the standard operating condition template most similar to the current abnormal operating condition, retrieved from the phase typical template library through similarity matching in step S5.2. The standard speed trajectory data is a set of discrete, time-ordered "timestamp-speed value" data pairs, completely depicting the trajectory of the ideal speed changing over time under the corresponding process stage, verified by a large amount of historical compliant data.

[0141] Using the standard speed trajectory data in the nearest-neighbor operating condition template as a reference and the time span parameter of the abnormal sub-speed curve segment as a constraint benchmark, a set of input elements for a two-dimensional linear affine transformation matrix is ​​constructed. The standard template speed trajectory and the abnormal speed curve segment are mapped to a unified time axis and amplitude coordinate system, respectively, to ensure the consistency of the data point correspondence. Under the unified coordinate system, the set of corresponding points between the amplitude vector of the abnormal curve segment and the amplitude vector of the standard template is extracted to form an amplitude difference matrix for calculating the scaling factor. For the time axis offset, the difference sequence between the start time of the abnormal curve segment and the start time of the standard template is extracted, and a time offset difference matrix is ​​constructed. The least squares method is used to solve for the scaling factor and time axis translation offset of the linear affine transformation in the joint space of the amplitude difference matrix and the time difference matrix, where the scaling factor is calculated by the following formula:

[0142] Among them, v template For standard template magnitude vector elements, v abnormal These are the amplitude vector elements of the abnormal curve segment; the time axis translation offset is calculated using the following formula:

[0143] Among them, t template For template time series elements, t abnormal For the time series elements of the abnormal curve segment, N is the total number of corresponding data points; the obtained scaling factor and time axis translation offset are encapsulated into a geometric transformation parameter group and output as a remapping control parameter group to the next data reconstruction stage. Through the above processing method, the template matching result of the previous step is transformed into an amplitude adjustment factor and time alignment amount with numerical executable, so as to realize the physical consistency remapping of the abnormal curve segment in amplitude and time.

[0144] S5.4: Based on the remapping control parameter group, perform point-by-point numerical reconstruction operation on the abnormal sub-speed curve segment, apply the scaling factor to adjust the speed amplitude and apply the time axis translation offset to align the phase start point, and at the same time perform truncation operation on redundant data segments that exceed the process stage boundary to generate candidate reconstructed speed curve segments that initially conform to the physical behavior logic.

[0145] Based on the scaling factor and time axis translation offset in the remapping control parameter group, a single-point amplitude adjustment operation is performed on the discrete speed sampling points of the abnormal sub-speed curve segment, and the original speed value is multiplied by the scaling factor to correct the amplitude.

[0146] The amplitude-corrected rotational speed sampling points are shifted by the time sampling index, and the adjusted time index is added to the whole to achieve phase start point alignment.

[0147] Boundary condition detection is performed on the speed curve segments processed by scaling and time translation. The sampling point time index is compared with the start and end times of the process stage to determine whether there is redundant data that exceeds the stage boundary.

[0148] Based on the boundary condition detection results, redundant data segments that exceed the stage range are truncated, and sampling points that do not conform to the stage timing logic are removed to avoid velocity jumps between phases.

[0149] The continuity of the curve segment after boundary truncation is calculated, and the smoothness of the rate of change of acceleration is checked by the difference of nearby sampling points to ensure that the corrected curve segment conforms to the trend of the kinematic envelope constraint set in physical implementation.

[0150] By using scaling, time shifting, and redundancy truncation, the abnormal sub-speed curve segment results from the previous step are transformed into candidate reconstructed speed curve segments that preliminarily conform to the physical behavior logic, thus achieving physical feasibility repair of the predicted curve under extreme operating conditions.

[0151] S5.5: Perform dynamic consistency final verification processing on the candidate reconstructed speed curve segment to verify whether the acceleration continuity and speed monotonicity of the candidate reconstructed speed curve segment meet the requirements of the piecewise kinematic envelope constraint set, and encapsulate the verified final data stream into a reconstructed speed curve segment with a phase verification qualified identifier to complete the physical feasibility repair of the abnormal segment.

[0152] Using the candidate reconstructed rotational speed curve segment as input, the dynamic boundary thresholds of each stage in the piecewise kinematic envelope constraint set are called and the current phase semantic label is loaded. A one-time difference operation is performed on the discrete sampling points of the whole sequence to generate an instantaneous acceleration numerical sequence.

[0153] The instantaneous acceleration numerical sequence is compared with threshold parameters such as acceleration non-negativity and deceleration continuous decay curve model to determine the sign polarity and slope trend, and a monotonicity determination matrix is ​​constructed and the acceleration sign change interval is marked.

[0154] Perform a continuous violation scan within the marked interval, call the velocity difference features of adjacent discrete rotation speed sampling points, apply the convergence judgment criterion, evaluate whether the final velocity has stably converged to zero, and form a convergence judgment matrix.

[0155] Perform a logical AND and aggregation on the monotonicity judgment matrix and the convergence judgment matrix to generate a stage dynamics consistency qualification flag, and attach the qualification flag to the metadata of the corresponding speed curve segment.

[0156] A format wrapper is used to merge the curve segments with the qualification mark with the original timestamp index, and the final data stream after wrapping is output as the reconstructed speed curve segment.

[0157] By segmented logical comparison and final verification, the candidate reconstruction curves from the previous step are transformed into a data stream that passes dynamic consistency verification, thereby achieving the physical feasibility repair of the abnormal sub-speed curve segment.

[0158] Step S6: Replace the failed sub-speed curve segments with the corresponding reconstructed speed curve segments, and keep the passed sub-speed curve segments in their original state. Then, concatenate the sub-speed curve segments in chronological order to generate the final physically consistent speed control command sequence. Specifically, this includes: S6.1: Obtain the constraint violation flag and reconstruct the speed curve segment sequence from the physical feasibility verification results. Based on the flag status, perform a conditional branch filtering operation on the original sub-speed curve segment sequence to separate the set of sub-speed curve segments marked as normal to be retained and the placeholder sequence of sub-speed curve segments marked as abnormal to be replaced.

[0159] The constraint violation flag from the physical feasibility verification result is used as the criterion input for the decision branch. Simultaneously, the complete data objects of the reconstructed speed curve segment sequence and the original sub-speed curve segment sequence obtained after processing in step S5 are acquired. Based on the constraint violation flag, the physical consistency state of each segment in the original sub-speed curve segment sequence is detected segment by segment using a Boolean state determination module, and the detection results are mapped to binary decision signals. The binary decision signals are combined with the corresponding sub-segment index positions to form a structured state index table. According to the state index table, a conditional branch filtering operation is performed on the original sub-speed curve segment sequence. Sub-segments with compliant flags are added to the set of sub-speed curve segments to be retained, and sub-segments with non-compliant flags are replaced with placeholders and added to the placeholder sequence of sub-speed curve segments to be replaced. Data integrity verification is performed on the set of sub-speed curve segments to be retained to ensure that their timestamp continuity and phase label binding relationship are not disrupted, serving as the normal prediction segment output set for this sub-step. By using the above-mentioned conditional branching and index mapping processing method, the verification result of the previous step is transformed into two types of sub-rotation curve segment datasets with clear state classification, thus realizing the preparation conditions for subsequent segment reconstruction mapping and sequence splicing.

[0160] S6.2: Using the timestamp index information of the placeholder sequence of the sub-speed curve segment to be replaced, retrieve the corresponding reconstructed speed curve segment that conforms to the physical behavior logic from the reconstructed speed curve segment sequence, and map the retrieved reconstructed speed curve segment to the placeholder position to generate a complete set of mixed speed curve segments containing normal prediction segments and corrected segments.

[0161] The timestamp index information of the placeholder sequence of the sub-speed curve segment to be replaced is used as the retrieval condition input to the matching control module. The start and end time boundaries recorded in the stage segmentation table are used to perform index matching operation on the reconstructed speed curve segment sequence to locate the target reconstructed segment that matches the time span.

[0162] Phase semantic verification is performed on the target reconstructed fragments obtained by matching. The process stage identifier bound to the placeholder is compared with the stage identifier in the metadata of the reconstructed fragment. Candidate fragments with inconsistent stages are eliminated to ensure the logical consistency of the replacement fragments.

[0163] Based on the target reconstructed fragment that has passed phase semantic verification, the data mapper is called to perform position mapping processing, so that the timestamp of the fragment is matched with the timestamp sequence of the placeholder point by point, and the consistency of the sampling point index is maintained for subsequent smooth splicing operations.

[0164] Amplitude consistency adjustment is performed on the reconstructed target segment after mapping. The segment rotation speed amplitude is adjusted by a scaling factor to match the boundary values ​​of the start and end segments of the placeholder, so as to avoid sudden jumps or amplitude imbalances after replacement.

[0165] The target reconstructed segment, after position mapping and amplitude adjustment, is filled into the placeholder position and together with the retained normal prediction segment, forms a complete set of mixed speed curve segments containing normal prediction segments and corrected segments. Through the above processing method, the result of the previous step is transformed into a curve segment dataset with time index consistency and amplitude coherence, achieving the physical consistency replacement effect of the corresponding abnormal segments.

[0166] S6.3: Based on the process stage switching time points defined in the slider motion phase label sequence, perform continuous smoothing processing on the connection boundaries of adjacent sub-segments in the full set of the mixed speed curve segments to eliminate speed steps or discontinuous derivatives that may be caused by segment replacement, and generate a spliced ​​speed curve sequence with time continuity and acceleration smoothness.

[0167] For the connection boundaries of adjacent sub-segments in the full set of mixed speed curve segments, based on the process stage switching time points defined in the slider motion phase label sequence, the speed and acceleration values ​​of the last sampling point of the front segment and the first sampling point of the back segment corresponding to each connection boundary are extracted as processing inputs.

[0168] Gradient detection is performed on the speed difference between the last sampling point of the front section and the first sampling point of the rear section. The speed step amplitude is calculated and a continuity error index matrix is ​​established. The elements in the error index matrix are composed of speed difference and acceleration difference, which are used to quantify the degree of discontinuity at the connection.

[0169] For the connection boundary in the error index matrix where the velocity step amplitude exceeds the preset continuity threshold, the local smoothing interpolation operator is called, and several sampling points on both sides of the boundary are selected as interpolation support nodes. Cubic spline interpolation is used to construct a smooth transition curve to ensure that the interpolation curve is completely continuous at the position and the first derivative.

[0170] Based on the acceleration difference on both sides of the boundary, an acceleration smoothing constraint equation is established. By solving the constraint, the second derivative of the interpolation curve is adjusted to ensure that the generated spliced ​​curve has no abrupt changes in acceleration. This constraint can be expressed as:

[0171] Where a is the acceleration function, t is the boundary connection time, and t + With t - These represent the instantaneous values ​​on the right and left sides of the boundary, respectively.

[0172] The smooth transition segment, adjusted by acceleration continuity constraints, is embedded into the original curve data stream, replacing the corresponding connection boundary data segment. Local weighted averaging is then performed at the embedding point to eliminate minor numerical differences caused by interpolation embedding.

[0173] The output is a sequence of spliced ​​speed curves after the above continuous smoothing process. This sequence has smoothness in velocity and acceleration on the global time axis, providing high-quality input for subsequent global time axis alignment and sampling rate normalization. Through continuous smoothing interpolation and acceleration constraint processing, the entire set of mixed speed curve segments from the previous step is transformed into spliced ​​curve data with temporal continuity and dynamic consistency, eliminating velocity steps and discontinuous derivatives at the connection points.

[0174] S6.4: Perform global time axis alignment and sampling rate standardization on the spliced ​​speed curve sequence. According to the communication protocol requirements of the servo drive speed loop controller, encapsulate the standardized data stream into a final physically consistent speed control command sequence with a synchronization frame header to complete the format conversion from internally calculated data to externally executed commands.

[0175] Step S7: Input the final physical consistency speed control command sequence into the speed loop controller of the servo driver to drive the forging press slide to perform a smooth speed switching action. Simultaneously, during execution, the phase matching degree, remapping amplitude, and constraint violation count are collected in real time as runtime governance indicators. Specifically, this includes: S7.1: Obtain the discrete time point speed setpoint in the final physical consistency speed control command sequence, and use the digital-to-analog conversion module to perform high-frequency interpolation and smoothing filtering on the discrete time point speed setpoint to generate a continuous and step-free analog speed reference voltage signal.

[0176] S7.2: Receive the analog speed reference voltage signal as the speed loop input, combine it with the actual rotor position signal fed back by the encoder in real time, and use the proportional-integral-derivative adjustment algorithm to perform dynamic compensation calculation on the speed deviation, so as to output an instantaneous torque current command with anti-disturbance capability.

[0177] The simulated speed reference voltage signal is used as the given input of the speed loop. The speed loop sampling module is called to collect the motor rotor position signal fed back by the encoder in real time within a unified sampling period, and the position-to-speed conversion operation is performed to obtain the actual speed value sequence.

[0178] Based on the actual rotational speed numerical sequence and the set value of the speed loop given signal, a point-by-point differential operation is performed in the delay compensation module to generate a speed deviation data stream, and the data stream is synchronously input into the proportional-integral-derivative control unit.

[0179] In the proportional control stage, the speed deviation value is multiplied by the set proportional coefficient to generate a proportional control output, so as to realize the real-time correction of speed error.

[0180] In the integral control stage, the speed deviation value is accumulated and summed, and then multiplied by a preset integral coefficient to form the integral control output, so as to eliminate the steady-state error caused by long-term deviation.

[0181] In the differential control stage, the first-order difference of the speed deviation value is calculated and multiplied by the set differential coefficient to generate the differential control output, which is used to suppress dynamic oscillations during speed change.

[0182] The proportional control output, integral control output, and derivative control output are combined using a weighted summation method to form a sequence of instantaneous torque current command values.

[0183] The proportional-integral-differential formula used to calculate the output term of each regulating element is defined as follows:

[0184] Where u is the instantaneous torque current command value, K p K is the proportionality coefficient. i K is the integral coefficient. d denoted as the differential coefficient, e is the velocity deviation value, and Δe is the change in deviation.

[0185] By combining proportional-integral-derivative (PID) control algorithms with encoder feedback, the analog speed reference voltage signal is converted into an instantaneous torque current command with disturbance rejection capability, achieving a smooth and precise speed control effect.

[0186] S7.3: Based on the instantaneous torque current command, drive the power inverter bridge arm to perform space vector pulse width modulation switching action, convert DC bus power into three-phase sinusoidal AC power and apply it to the stator winding of the forging press slide drive motor, so as to force the motor rotor to follow the target speed curve to complete the smooth speed switching action of the down-going, pressurizing, return and braking stages.

[0187] The input conditions are the instantaneous torque current command generated by the proportional-integral-derivative (PID) control algorithm and the drive channel status of the switching transistors of the power inverter bridge arm, with DC bus power as the energy input source. The drive control logic calls the space vector pulse width modulation (SPWM) algorithm to decompose the amplitude and phase components of the three-phase reference voltage based on the instantaneous torque current command, and calculates the corresponding space vector position in a two-dimensional α-β stationary coordinate system. Based on the calculated space vector position and the DC bus voltage value of the inverter bridge arm, the reference voltage vector is mapped to the six nearest sectors, and the selection of the switching state combination is achieved by determining the sector number. A customized vector synthesis strategy is adopted to determine the duty cycle of the two sets of active and zero vectors, and the duty cycle result is converted into a PWM trigger pulse sequence for each bridge arm to ensure the symmetry of the three-phase voltage waveform and minimum harmonic distortion. The PWM trigger pulses are applied to the three-phase bridge arm of the power inverter through the driver hardware layer to control the on and off behavior of the corresponding IGBT or MOSFET switches, so as to convert the DC bus power into a three-phase sinusoidal AC current with the target amplitude and frequency according to the space vector pulse mode. The generated three-phase sinusoidal alternating current is applied to the stator windings of the slider drive motor. Utilizing the principle of electromagnetic induction, the motor rotor is forced to produce a mechanical response consistent with the target speed curve, achieving smooth speed switching during the downward, pressurized, return, and braking phases. Through the aforementioned vector synthesis and PWM drive processing, the instantaneous torque current command from the previous step is converted into executable power electronic switch control data, achieving high-precision tracking and smooth switching of the target speed curve.

[0188] S7.4: During the speed switching action, the phase tag sequence and remapping operation log are captured synchronously. The actual running trajectory and the original predicted trajectory are compared and analyzed point by point using the timestamp alignment mechanism to calculate the phase matching degree value representing the prediction accuracy and the remapping amplitude value representing the correction strength.

[0189] During the speed switching operation, the actual running trajectory data sequence obtained from the servo drive encoder feedback and the slider motion phase label sequence are used as input objects. The phase label sequence and the remapping operation log are timestamped according to a unified sampling time axis to ensure that the sampling points of different data sources correspond completely in the time domain. Based on the timestamp alignment result, the trajectory comparison module is called to extract the speed values ​​in the actual running trajectory and the corresponding speed values ​​in the original predicted trajectory according to the discrete sampling point index positions, forming a point-by-point matching speed comparison table. The difference is calculated on the speed comparison table, and the difference between the actual speed value and the predicted speed value is used as the instantaneous deviation sequence. The instantaneous deviation sequence is segmented and stored according to the process stage using the phase label information for phase matching degree calculation. The normalized correlation coefficient method is used to calculate the phase matching degree of the deviation sequence of each process stage, and the matching degree value is obtained by the following formula:

[0190] Where v pred,i v represents the predicted velocity value for the i-th data point. act,i μ represents the actual velocity value of the i-th data point. pred Let μ be the mean of the predicted value sequence. act This is the mean of the actual value sequence.

[0191] Using the scaling factor and time axis shift offset recorded in the remapping operation log, and combining the difference between the original and corrected amplitudes of the aberration segment, the remapping amplitude value is calculated using the following formula:

[0192] Among them, v remap (j) represents the velocity value of the remapped segment at the j-th sampling point, v orig (j) represents the velocity value of the original anomalous segment at the j-th sampling point, m is the number of data points, and M is the remapping amplitude value. Through trajectory differential analysis and formula calculation based on timestamp alignment, the velocity matching accuracy and correction operation amplitude generated during real-time operation are transformed into quantified phase matching degree values ​​and remapping amplitude values, realizing a dual-index evaluation of prediction accuracy and correction strength.

[0193] S7.5: Summarize the phase matching degree value, remapping amplitude value, and constraint violation flag generated in the previous steps, and accumulate the frequency of abnormal events in the current work cycle through statistical counting logic to generate a set of runtime governance indicators with multiple dimensions and send it to the online evaluation module.

[0194] Step S8: Update the sample weights of the online evaluation module of the deep learning model based on the runtime governance metrics. Specifically, difficult examples are identified using the phase matching degree and the number of constraint violations, completing the closed-loop iterative optimization of the model under the data distribution offset scenario. This includes: S8.1: Obtain the phase matching degree value, remapping amplitude value and constraint violation count collected in real time during the execution of the servo driver, and perform timestamp alignment and normalization on the above multi-source runtime governance indicators to generate a standardized multi-dimensional runtime governance feature vector.

[0195] S8.2: Based on the standardized multi-dimensional runtime governance feature vector, execute the difficult example sample discrimination logic, and mark the corresponding historical working condition data with the constraint violation count exceeding the preset threshold or the remapping magnitude value exceeding the allowable range as high-weight redistribution offset difficult example samples, so as to construct a difficult example sample set containing data distribution offset features.

[0196] S8.3: Calculate the confidence decay coefficient of each sample using the phase matching degree value and the constraint violation count in the difficult sample set, and perform a reverse proportional correction operation on the initial weight value of the corresponding sample in the original training sample library based on the confidence decay coefficient to generate a dynamic sample weight matrix that reflects the degree of data distribution offset.

[0197] Using the phase matching degree values ​​and constraint violation counts from the difficult example sample set as input vectors, a confidence decay coefficient calculation model is constructed, and weight influence factor matrices for different stage conditions are loaded. Normalization is performed based on the input vector to ensure that each feature is calculated on a uniform scale, avoiding weight correction biases caused by differences in numerical ranges. Based on the normalized feature vector, an exponential decay function is used to generate the corresponding confidence decay coefficient for each sample. The specific calculation formula is as follows:

[0198] Where c is the confidence decay coefficient, pm is the normalized phase matching deviation value, vc is the normalized constraint violation count, and α and β are preset stage sensitivity coefficients. The calculated confidence decay coefficient is then proportionally adjusted against the initial weight values ​​of the corresponding samples in the original training sample library using the following formula:

[0199] Where w is the initial weight value and w′ is the corrected weight value. The w′ values ​​of all samples are rearranged according to their sample indices to form a dynamic sample weight matrix. Matrix sparsity processing is then performed to reduce the influence of irrelevant samples on the weights, ultimately outputting a dynamic sample weight matrix reflecting the degree of data distribution shift. Through the above-described inverse proportional correction method based on the confidence decay coefficient, the difficult sample identification results from the previous step are transformed into data weight indicators that can be directly used for model optimization, achieving dynamic weight adjustment for data distribution shift scenarios.

[0200] S8.4: Update the weighted terms of the loss function of the online evaluation module of the deep learning model according to the dynamic sample weight matrix, and input the set of difficult samples as a priority training batch into the deep learning model for gradient backpropagation fine-tuning, so as to complete the iterative optimization process of model parameters for the data distribution offset scenario.

[0201] Based on the numerical distribution characteristics of the dynamic sample weight matrix, the weighted terms of the loss function structure are updated in the online evaluation module of the prediction model. The weight value corresponding to each sample in the matrix is ​​embedded as the multiplication coefficient when inputting the loss function into the sample loss calculation path. Based on the updated loss function structure, the hard example sample set is prioritized for training batch arrangement. The samples in the set are sorted according to the degree of data distribution offset and loaded into the training data buffer in sequence. For the hard example sample data loaded in the buffer, the deep learning model is invoked to infer the forward path, calculate the prediction output, and solve for the error with the target label to obtain the sample loss value and gradient direction. The obtained gradient direction and the updated loss function weighted terms are input into the model's backpropagation engine, and the parameters are updated through layer-by-layer gradient calculation and weight correction. Numerical stability checks are performed on the weight correction values ​​of each layer. Correction values ​​exceeding the preset update magnitude threshold are proportionally limited to prevent large updates from causing a decrease in model convergence. Through the above processing, the dynamic sample weight matrix of the previous step is transformed into a loss function execution mechanism with data distribution offset adaptation capability, realizing iterative optimization of the prediction model parameters under offset scenarios.

[0202] S8.5: Based on the accuracy improvement rate of the validation set and the decrease in the constraint violation rate output by the iterative optimization process of the model parameters, a model convergence status flag is generated, and the model convergence status flag is fed back to the intelligent prediction module to trigger the deployment and switching of the new version of the speed curve prediction model, thereby forming a complete model iterative optimization closed loop.

[0203] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0204] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0205] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent speed control of a slide block drive motor in a forging press, characterized in that, Specifically, it includes: S1: Acquire multi-source sensor data during the operation of the forging press slide, and generate a slide motion phase label sequence containing several process stage identifiers based on the multi-source sensor data; S2: Based on the dynamic boundary characteristics of each process stage in the slider motion phase label sequence, a segmented kinematic envelope constraint set is constructed. The segmented kinematic envelope constraint set is defined as a set of engineering verifiable rules preset for each phase stage. S3: Based on the time segmentation points of the slider motion phase label sequence, the original predicted rotational speed time sequence output by the preset deep learning model is divided into sub-rotational speed curve segment sequences corresponding to the process stage; S4: Using the segmented kinematic envelope constraint set, perform symbolic logic comparison operation on the sub-speed curve segment sequence to determine whether each sub-speed curve segment satisfies the envelope boundary and qualitative characteristics of the corresponding process stage, and generate physical feasibility verification results; S5: For the abnormal sub-speed curve segments marked as constraint violations in the physical feasibility verification results, the nearest working condition template is retrieved based on the preset phase typical template library, and the abnormal segments are remapped to the corresponding positions of the nearest working condition template to generate reconstructed speed curve segments. S6: Replace the abnormal sub-speed curve segment that fails the verification with the corresponding reconstructed speed curve segment, and keep the verified sub-speed curve segment in its original state, and then splice the sub-speed curve segments in time order to generate the final physically consistent speed control command sequence.

2. The intelligent speed control method for a forging press slide drive motor according to claim 1, characterized in that, The slider motion phase label sequence includes downlink start, pressurized steady state, return acceleration, and braking stop.

3. The intelligent speed control method for a forging press slide drive motor according to claim 2, characterized in that, Step S3 specifically includes: Obtain the original predicted rotational speed time series and slider motion phase label sequence output by the deep learning model. Based on the timestamps contained in the slider motion phase label sequence, extract the start and end times of each of the four process stages: downlink start, pressurized steady state, return acceleration, and braking stop, and generate a process stage time window set containing four sets of time boundary parameters. Based on the start and end time points in the set of process stage time windows, a time-series indexing operation is performed on the original predicted speed time series to calculate the discrete data point index range corresponding to each process stage time window in the original predicted speed time series, and to generate a process stage data mapping table containing four sets of index intervals. Using the discrete data point index range in the process stage data mapping table, the speed value segments within the corresponding index interval are extracted from the original predicted speed time series, and the down-start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence are constructed respectively, generating an initial sub-speed curve segment set containing the down-start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence; Phase attribute binding processing is performed on the downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence in the initial sub-speed curve segment set to generate downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking stop speed subsequence with phase semantic tags; Based on the downlink start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking parking speed subsequence with phase semantic tags, the data structure is reorganized in chronological order and encapsulated to form the sub-speed curve segment sequence.

4. The intelligent speed control method for a slide block drive motor of a forging press according to claim 3, characterized in that, The step of performing phase attribute binding processing on the downhill start speed subsequence, pressurized steady-state speed subsequence, return acceleration speed subsequence, and braking parking speed subsequence in the initial sub-speed curve segment set includes: The process stage identifier for the downlink startup is appended as metadata to the header of the downlink startup speed subsequence; The process stage identifier of the pressurized steady state is appended as metadata to the header of the pressurized steady state speed subsequence; The process stage identifier corresponding to the return acceleration is appended as metadata to the header of the return acceleration speed subsequence. The process stage identifier corresponding to the braking stop is appended as metadata to the head of the braking stop speed subsequence.

5. The intelligent speed control method for a slide drive motor of a forging press according to claim 1, characterized in that, The set of verifiable rules for the project includes rotational speed monotonicity, acceleration non-negativity, fluctuation amplitude threshold, and convergence continuity requirements.

6. The intelligent speed control method for a forging press slide drive motor according to claim 5, characterized in that, Step S4 specifically includes: Based on the segmented kinematic envelope constraint set, the dynamic boundary threshold set of the current process stage is extracted, and the discrete speed sampling points in the sub-speed curve segment sequence are used as input objects. The instantaneous acceleration numerical sequence between adjacent discrete speed sampling points is calculated through differential operation to obtain the instantaneous acceleration numerical sequence. The sign polarity of the instantaneous acceleration numerical sequence is compared with the acceleration non-negativity requirement in the dynamic boundary threshold set. Monotonicity determination logic is then performed on the sub-speed curve segments of the downlink start-up phase and the return acceleration phase to generate phase monotonicity determination results. Based on the sub-speed curve segments marked as compliant in the monotonicity determination results of the aforementioned stage, the extreme value difference calculation is performed on the sub-speed curve segments of the pressurized steady-state stage by calling the fluctuation amplitude threshold rule. By comparing the magnitude relationship between the maximum speed deviation value and the preset fluctuation amplitude threshold, the fluctuation amplitude compliance flag is obtained. Based on the aforementioned fluctuation amplitude compliance flag and convergence continuity requirements, the final slope decay trend analysis is performed on the sub-speed curve segment during the braking and parking phase. By fitting the tangent slope parameter of the final speed trajectory and determining whether it strictly converges to zero, a convergence continuity verification index is generated. Based on the combined results of the stage monotonicity determination, the fluctuation amplitude compliance flag, and the convergence continuity verification index, a weighted aggregation process is used to calculate the degree of deviation between each sub-rotation speed curve segment and the ideal physical behavior model, and the physical feasibility verification result is output.

7. The intelligent speed control method for a forging press slide drive motor according to claim 1, characterized in that, The physical feasibility verification results include: phase matching degree index and constraint violation flag.

8. The intelligent speed control method for a slide block drive motor of a forging press according to claim 5, characterized in that, Step S5 specifically includes: Obtain the constraint violation flag and the corresponding abnormal sub-speed curve segment from the physical feasibility verification result. Based on the constraint violation flag, extract the speed data block to be corrected containing phase semantic tags, and perform feature vector standardization processing on the speed data block to be corrected to generate a normalized abnormal operating condition feature vector. Based on the normalized abnormal working condition feature vector, a multidimensional Euclidean distance similarity matching process is performed by calling a preset phase typical template library. The distance metric between the normalized abnormal working condition feature vector and the feature vector of each standard working condition template in the phase typical template library is calculated, and the nearest working condition template with the smallest distance metric is selected as the benchmark reference object. Using the standard speed trajectory data in the nearest working condition template and the time span parameter of the abnormal sub-speed curve segment, a linear affine transformation matrix is ​​constructed. The scaling factor and time axis translation offset are solved by the least squares method to generate a remapping control parameter set. Based on the remapping control parameter group, a point-by-point numerical reconstruction operation is performed on the abnormal sub-speed curve segment. The speed amplitude is adjusted by applying the scaling factor and the time axis translation offset is applied to align the phase start point. At the same time, a truncation operation is performed on redundant data segments that exceed the process stage boundary to generate candidate reconstructed speed curve segments. The candidate reconstructed rotational speed curve segment is subjected to a dynamic consistency final verification process. The final data stream that satisfies the requirements of the piecewise kinematic envelope constraint set for the acceleration continuity and rotational speed monotonicity of the candidate reconstructed rotational speed curve segment is encapsulated into the reconstructed rotational speed curve segment with a phase verification qualified identifier.

9. The intelligent speed control method for a forging press slide drive motor according to claim 1, characterized in that, After step S6, the following is included: S7: Input the final physical consistency speed control command sequence into the speed loop controller of the servo driver to drive the forging press slider to perform a smooth speed switching action. At the same time, during the execution process, the phase matching degree, remapping amplitude and constraint violation number are collected in real time as runtime governance indicators. S8: Update the sample weights of the online evaluation module of the deep learning model based on the runtime governance metrics.

10. The intelligent speed control method for a forging press slide drive motor according to claim 1, characterized in that, The multi-source sensor data includes: real-time load torque signal, displacement slope signal, and pressure platform duration signal.

Images