A mold die set deformation monitoring method, system, intelligent terminal and storage medium

Abstract

This application relates to the technical field of condition monitoring for forging equipment, and in particular to a method, system, intelligent terminal, and storage medium for monitoring the deformation of a die holder. The method includes acquiring contact sensor signals; performing dynamic baseline calibration and impact separation processing on the contact sensor signals to filter out transient interference and compensate for signal drift, extracting deformation signals characterizing plastic accumulation; analyzing the deformation development dynamics and quantifying its accumulation process based on the temporal changes of the deformation signals; inputting the quantified deformation process information into a degradation assessment model trained based on historical service data to obtain assessment results; and triggering corresponding graded early warning or deformation compensation control commands based on the assessment results. This application improves upon the high false alarm rate and insufficient intelligence of existing die holder deformation monitoring methods, enhancing the real-time performance, accuracy, and reliability of monitoring.

Description

Technical Field

[0001] This application relates to the technical field of condition monitoring of forging equipment, and in particular to a method, system, intelligent terminal and storage medium for monitoring mold base deformation. Background Technology

[0002] During the forging process, the die holder will undergo plastic deformation due to accumulated stress over a long period of use. This deformation is usually hidden and difficult to observe with the naked eye, but it directly leads to a decrease in the precision of the forged parts and an increase in the scrap rate. Scrap caused by hidden die holder deformation accounts for approximately 17% of the total scrap rate for forging parts, and each unplanned downtime for inspection results in an average loss of about 8 hours of production capacity. Therefore, timely and accurate monitoring of die holder deformation is necessary to ensure product quality and improve production efficiency.

[0003] In related technologies, the following methods are mainly used to monitor the deformation of the mold base: First, periodic shutdown and disassembly, followed by manual inspection using a three-dimensional coordinate measuring machine, but this method is lagging and affects production continuity; second, online sensing technologies such as laser ranging or strain gauges are used, but the former is easily interfered with in oily environments, and the latter requires damage to the mold base structure during installation. In terms of circuit control monitoring, the following solutions are mainly used: First, a switching circuit composed of mechanical relays is used, which has a simple structure but slow response, easy oxidation of contacts, and poor reliability; second, a threshold judgment circuit based on voltage comparators is used, which has a low cost but can only set a fixed threshold, cannot effectively distinguish between instantaneous mechanical impact and real continuous deformation, and has a high false judgment rate.

[0004] Regarding the aforementioned technologies, existing circuit monitoring solutions struggle to balance reliability and intelligent judgment, often exhibiting key flaws such as a lack of signal drift compensation mechanisms, inability to dynamically adjust alarm thresholds, and a lack of multi-level early warning functions. These shortcomings result in persistently high false alarm rates in practical applications. Summary of the Invention

[0005] To improve the high false alarm rate and insufficient intelligence of existing mold base deformation monitoring methods, and to enhance the real-time performance, accuracy, and reliability of monitoring, this application provides a mold base deformation monitoring method, system, intelligent terminal, and storage medium.

[0006] Firstly, this application provides a method for monitoring the deformation of a mold base, employing the following technical solution:

[0007] A method for monitoring mold base deformation includes:

[0008] The contact sensing signal triggered by the plastic deformation of the mold base is obtained by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element.

[0009] Dynamic baseline calibration and impact separation are performed on the contact sensing signal to filter out transient interference and compensate for signal drift, and to extract the deformation signal characterizing plastic accumulation.

[0010] Based on the temporal variation of deformation signals, we analyze the dynamics of deformation development and quantify its cumulative process;

[0011] The quantified deformation process information is input into a degradation assessment model trained based on historical service data to obtain assessment results that include risk level and trend prediction.

[0012] Based on the assessment results, corresponding graded early warnings or deformation compensation control commands are triggered.

[0013] By adopting the above technical solutions, a complete monitoring process from signal acquisition, intelligent processing, trend analysis to proactive decision-making is constructed, improving the accuracy and intelligence level of mold base deformation monitoring. Contact signals originating from plastic deformation are obtained by monitoring the closure of preset mechanical gaps, ensuring the relevance and reliability of the monitoring signals. Dynamic baseline calibration and impact separation processing filter out signal fluctuations caused by environmental interference and instantaneous mechanical impacts, solving the problem of high false alarm rates caused by signal drift and interference in traditional fixed threshold methods. Analyzing the temporal changes of deformation signals and using a degradation assessment model trained on historical data for prediction enables quantitative assessment of the deformation accumulation process and judgment of risk trends, overcoming the limitations of traditional methods that only provide delayed and discrete alarm states. Based on the intelligent assessment results, graded early warning or compensation commands are triggered, making response measures more precise and timely. While ensuring forging precision, this maximizes the maintenance of production continuity, achieving real-time monitoring, accuracy, and reliability.

[0014] Optionally, the steps of performing dynamic baseline calibration and impact separation processing on the contact sensing signal include:

[0015] The sensor signals of the mold base under non-plastic deformation state or specific operating stage are acquired, and the reference signal sequence is collected based on the moving time window and its statistical characteristics are established as a dynamic baseline.

[0016] The real-time acquired contact sensing signal is compared with the dynamic baseline to calculate the signal offset. The dynamic baseline is then adaptively updated using a recursive filtering algorithm to compensate for slow signal drift caused by environmental temperature drift or component aging.

[0017] By combining time-domain peak detection with frequency-domain high-pass filtering, high-frequency, high-amplitude signal components generated by instantaneous mechanical shock or vibration are identified and separated.

[0018] The impact interference component is removed from the baseline-calibrated signal, and a low-frequency deformation signal that mainly reflects the cumulative plastic deformation of the mold base is output.

[0019] By adopting the above technical solutions, the combined signal processing strategy enhances the system's stability and anti-interference capability in complex industrial environments. Utilizing a moving-time window to establish and adaptively update the dynamic baseline continuously tracks and compensates for slow signal drift caused by factors such as ambient temperature changes and sensor component aging, ensuring the long-term stability of the monitoring benchmark and overcoming the baseline drift problem commonly found in traditional circuit solutions. The combination of time-domain peak detection and frequency-domain high-pass filtering accurately identifies and separates the inherent high-frequency, high-amplitude instantaneous impact and vibration components during the forging process. The joint analysis method effectively distinguishes between real, slowly accumulating plastic deformation signals and transient mechanical interference, avoiding false triggering caused by normal production impacts and reducing the system's false alarm rate.

[0020] Optionally, the step of outputting the deformation signal includes:

[0021] A signal separation model is constructed based on the mechanical constitutive relationship of the mold base material;

[0022] Using a signal separation model, the signal after baseline calibration and removal of impact interference is analyzed into a first component corresponding to instantaneous elastic deformation and a second component corresponding to cumulative plastic deformation.

[0023] The first component is not used; the second component is separated and output as an incremental signal characterizing the development process of plastic deformation.

[0024] By adopting the above technical solution and introducing a signal separation model based on the constitutive relationship of material mechanics, the physical essence of the sensing signal can be analyzed, improving the accuracy and theoretical reliability of deformation monitoring. The model can decompose the pre-processed mixed signal into components corresponding to recoverable instantaneous elastic deformation and unrecoverable cumulative plastic deformation, based on the mechanical behavior of the material. The separation process removes the inherent elastic response in each forging cycle from the physical mechanism level, allowing the final output signal to purely characterize the permanent plastic accumulation caused by factors such as fatigue and creep. This more sensitively captures the intrinsic correlation between microscopic damage and macroscopic deformation of the die material, enabling earlier detection and quantification of the plastic deformation process.

[0025] Optionally, the steps for constructing a signal separation model include:

[0026] After the mold is first installed or overhauled, a series of known loads are applied and high-precision deformation data is acquired simultaneously to calibrate the initial key parameters of the model.

[0027] In the early stages of normal service of the mold, the initial key parameters are fine-tuned online using continuous forging cycle data and a recursive estimation algorithm to match the model output with the actual transient response of the mold base.

[0028] After detecting a maintenance or replacement event in the mold, the drift compensation parameters in the model are reset or recalibrated based on the baseline offset of the deformation signals before and after the event.

[0029] The model prediction residuals are continuously calculated. When the statistical characteristics of the residuals exceed the allowable range, a model failure warning is generated and an offline high-precision calibration process is triggered.

[0030] By adopting the above technical solutions, a full lifecycle, adaptive model building and maintenance mechanism is provided, ensuring the accuracy and reliability of the signal separation model in long-term operation. Initial high-precision calibration establishes a precise starting point for the model that matches the actual model base materials and structure. Online fine-tuning during the initial service phase enables the model to quickly adapt to the dynamic response characteristics of specific equipment and operating conditions, improving personalized matching. Parameter reset and recalibration processes for maintenance, replacement, and other events continuously monitor model prediction residuals, providing timely warnings and triggering recalibration when performance degrades. This ensures that the monitoring model does not gradually fail due to equipment maintenance, component replacement, or its own performance drift, maintaining high-precision separation capabilities and guaranteeing the continuous reliability and accuracy of the entire monitoring system throughout its service life.

[0031] Optionally, the steps for analyzing the dynamics of deformation development and quantifying its cumulative process include:

[0032] Time-frequency analysis is performed on the incremental signal to extract its frequency domain distribution characteristics in order to assess the stability of the deformation process;

[0033] Based on the stability assessment results, the deformation evolution stages are divided;

[0034] The incremental signal is subjected to time-series accumulation operation to quantify the cumulative total amount of plastic deformation.

[0035] By employing the above technical solutions, time-frequency analysis focuses on the cumulative magnitude of deformation, and can assess the stability and pattern of deformation development by extracting frequency domain distribution characteristics, thus identifying potential risks of deformation acceleration or instability at an early stage. Based on stability assessment, deformation evolution stages are divided, and the continuous deformation process is divided into intervals with different characteristics, helping operators understand the current damage state and development trend of the mold base. Specialized cumulative calculations are performed on the plastic deformation components to quantify the total amount of irreversible plastic deformation, which is directly related to the structural integrity and remaining life of the mold base. This provides a richer information hierarchy than a single threshold alarm, offering multi-dimensional and in-depth quantitative basis for predictive maintenance decisions.

[0036] Optionally, the degradation assessment model is a time-series prediction model built on a recurrent neural network. The training process of the degradation assessment model includes:

[0037] The temporal feature sequence representing the deformation process is obtained as input;

[0038] It learns long-term dependencies and evolution patterns in temporal feature sequences through its network structure;

[0039] The output includes assessment results that include discrete risk level classification and continuous remaining life prediction.

[0040] By adopting the above technical solution, the time-series prediction model based on recurrent neural networks can receive continuous time-series feature sequences characterizing the deformation process. Utilizing the network structure's memory capacity, it learns the complex long-term dependencies, nonlinear evolution patterns, and historical degradation laws within the deformation signals. Through learning, the model can not only output a discrete risk level classification for the current moment but also continuously predict the remaining service life of the mold base or the deformation development trend. This enables the system to provide early warnings of potential failure risks, arrange planned maintenance, prepare spare parts, or adjust production plans, elevating the maintenance mode from reactive remediation and periodic overhauls to state-based predictive maintenance, thereby improving the proactiveness of production support.

[0041] Optionally, the steps to trigger the corresponding tiered warning include:

[0042] When the assessment result indicates that the deformation risk is low, a status alert message is generated and sent.

[0043] When the assessment results indicate an increased risk of deformation, control commands are generated to limit or adjust the operating parameters of the forging equipment.

[0044] When the assessment results indicate that the deformation risk has reached the highest level, a control command is generated to initiate deformation compensation and interrupt the forging operation.

[0045] By adopting the above technical solutions, a multi-level and differentiated early warning and response mechanism has been implemented, achieving refined and intelligent risk management and balancing production safety and efficiency. When the risk assessment is low, only status alerts are provided to avoid unnecessary interference with the production process and maintain production continuity. When the risk increases, the system proactively generates control commands, limiting or adjusting the working parameters of the forging equipment (such as reducing forging speed and adjusting pressure) to continue production within permissible limits while actively suppressing further acceleration of deformation through mitigation strategies. When the risk reaches the highest level, the system decisively triggers deformation compensation or interrupts operations to prevent catastrophic failure. The tiered response system changes the one-size-fits-all alarm approach, ensuring that response measures match the risk level. This avoids capacity losses caused by over-maintenance and eliminates equipment damage and safety accidents that may result from insufficient response, achieving a balance between safety and efficiency.

[0046] Secondly, this application provides a mold base deformation monitoring system, which adopts the following technical solution:

[0047] A mold base deformation monitoring system, including

[0048] The acquisition module is used to acquire contact sensing signals triggered by the plastic deformation of the mold base, which are generated by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element.

[0049] A memory for storing a program for a mold base deformation monitoring method as described above;

[0050] The processor and the program in the memory can be loaded and executed by the processor to implement the mold base deformation monitoring method as described above.

[0051] Thirdly, this application provides a smart terminal, which adopts the following technical solution:

[0052] A smart terminal includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described above.

[0053] Fourthly, this application provides a computer storage medium capable of storing corresponding programs, employing the following technical solution:

[0054] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above.

[0055] In summary, this application includes at least one of the following beneficial technical effects:

[0056] This scheme employs dynamic baseline calibration and impact separation technology to filter out environmental drift and instantaneous mechanical interference. It also introduces a signal separation model based on the constitutive relations of material mechanics. This model can physically resolve the sensing signal into elastic and plastic components, thereby directly extracting the plastic accumulation signal characterizing irreversible damage. By combining anti-interference preprocessing with mechanistic model analysis, the scheme ensures the authenticity and accuracy of the monitoring signal from the source and in principle, overcoming the high false alarm rate caused by signal drift and interference confusion in traditional methods.

[0057] By employing models such as recurrent neural networks, deep learning is performed on the quantified deformation process time series data. The model captures the long-term dependencies and nonlinear patterns of deformation evolution. While outputting the current risk level, it continuously predicts the remaining life of the mold base or the deformation trend, enabling the monitoring system to have forward-looking judgment capabilities. This provides early warning and quantitative basis for predictive maintenance decisions, such as planned shutdowns and spare parts preparation, and elevates the maintenance mode from post-processing to pre-planning, thereby improving the initiative of production management and equipment reliability.

[0058] This solution is deeply integrated with the production control system. Based on the risk level generated by intelligent assessment, it automatically triggers multi-level and differentiated response commands, ranging from risk alerts and adjustments to process parameters for mitigation to initiation compensation or shutdown. The tiered response mechanism overcomes the shortcomings of a one-size-fits-all alarm system, ensuring that control measures are precisely matched to the severity of the risk. It can maintain production continuity in the early stages of risk and intervene decisively when the risk is critical, maximizing production efficiency while ensuring equipment safety and product accuracy, thus balancing safety and profitability. Attached Figure Description

[0059] Figure 1 This is a flowchart of a mold base deformation monitoring method according to an embodiment of this application.

[0060] Figure 2 This is a flowchart of the steps for performing dynamic baseline calibration and impact separation processing on contact sensing signals according to an embodiment of this application.

[0061] Figure 3 This is a flowchart of the steps for outputting deformation signals in an embodiment of this application.

[0062] Figure 4 This is a flowchart of the steps for constructing a signal separation model according to an embodiment of this application.

[0063] Figure 5 This is a flowchart illustrating the steps of analyzing deformation development dynamics and quantifying its cumulative process in an embodiment of this application.

[0064] Figure 6 This application embodiment shows that the degradation assessment model is a time-series prediction model built based on a recurrent neural network, and the flowchart shows the training process of the degradation assessment model.

[0065] Figure 7 This is a flowchart of the steps for triggering corresponding graded warnings in an embodiment of this application.

[0066] Figure 8 This is a block diagram of a mold base deformation monitoring system according to an embodiment of this application. Detailed Implementation

[0067] The present application will be further described in detail below with reference to the accompanying drawings.

[0068] This specific embodiment is merely an explanation of this application and is not intended to limit it. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they fall within the scope of the claims of this application.

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the appendices in the embodiments of this application will be described below. Figure 1-8The technical solutions in the embodiments of this application are clearly and completely described. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] This application discloses a mold base deformation monitoring device. The device can directly detect the sinking displacement of the mold base caused by cumulative plastic deformation and convert it into a quantifiable and processable electrical signal, providing a hardware foundation for monitoring methods. The mold base deformation monitoring device includes a mold base assembly, a sensing assembly, and a circuit control assembly. The three work together to convert the mechanical deformation of the mold base into an electrical signal.

[0071] The die base assembly is the main body for bearing and transmitting deformation, and includes a die base body and positioning blocks. The die base body is the lower die base body of the forging die, and a fixing groove is formed at the bottom of the die base body. There can be multiple positioning blocks, preferably four in this embodiment, which are fixedly installed around the inner wall of the fixing groove.

[0072] The positioning block is made of a material with certain damping characteristics, such as nylon or polyurethane, which mainly provides lateral positioning and buffering, absorbs some high-frequency vibration impacts, and avoids interference with the vertical force measurement of the core.

[0073] The sensing assembly is the core component that senses the sinking of the mold base and generates the initial mechanical action. It is installed in the fixing groove of the mold base body. The sensing assembly includes a fixing post and a spring steel washer. The fixing post is vertically fixed at the bottom center of the fixing groove and must have sufficient rigidity to provide a stable support reference.

[0074] The spring steel washer is fixedly installed on the top of the fixed column. The washer is elastic and its edge is fitted with the positioning blocks around it. When the spring steel washer is subjected to vertical pressure, it can undergo slight elastic deformation along the axis of the fixed column. Under the limit of the positioning blocks, the edge can only wobble horizontally within a very small range, thereby ensuring the stability of the force direction and isolating horizontal disturbances.

[0075] The circuit control component is responsible for completing the force-to-electricity conversion and preliminary signal processing. This component includes a piezoresistive pressure sensor and a signal conditioning circuit. The piezoresistive pressure sensor is the core sensing element, fixedly mounted at the bottom of the mounting slot, directly below the spring steel pad. After initial installation and calibration, a 1-2 mm vertical air gap (preset mechanical clearance) is maintained between the upper surface of the piezoresistive pressure sensor (the pressure-sensing surface) and the bottom surface of the spring steel pad.

[0076] The signal conditioning circuit is electrically connected to the output terminal of the piezoresistive pressure sensor. The signal conditioning circuit includes an excitation power supply, a Wheatstone bridge, an instrumentation amplifier, a filter circuit, etc., which are used to power the sensor and amplify and filter the weak millivolt-level differential signal output by the sensor, converting it into a stable and reliable analog voltage or digital signal output for subsequent processing system acquisition.

[0077] This application discloses a method for monitoring the deformation of a mold base.

[0078] Reference Figure 1 Methods for monitoring mold base deformation include:

[0079] Step S100: Acquire a contact sensing signal triggered by the plastic deformation of the mold base, generated by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element.

[0080] The preset mechanical clearance refers to a 1-2 mm vertical air gap between the sensing surface of the piezoresistive pressure sensor and the bottom surface of the spring steel gasket, corresponding to the critical physical distance from the initial intact state of the mold base to the point where measurable plastic deformation begins to occur. The contact sensing signal refers to the electrical signal generated by the spring steel gasket pressing on the piezoresistive pressure sensor after the gap closes due to plastic deformation of the mold base. This signal is proportional to the pressure. The contact sensing signal linearly reflects the force exerted on the sensing component by the plastic deformation of the mold base.

[0081] The signal acquisition process is as follows: The system acquires the output of the piezoresistive pressure sensor in real time through a signal conditioning circuit. Before the mold base undergoes sufficient plastic deformation to close the gap, the sensor outputs a small signal corresponding to the zero point or ambient noise. When the accumulated plastic deformation causes the mold base body to sink to the preset gap value, the spring steel gasket begins to contact the sensor, and the signal conditioning circuit outputs a significantly increased base voltage. The system continuously acquires this voltage signal at a fixed sampling rate (e.g., 1 kHz), and this signal sequence is the original contact sensing signal.

[0082] Step S101: Perform dynamic baseline calibration and impact separation processing on the contact sensing signal, filter out transient interference and compensate for signal drift, and extract deformation signals that characterize plastic accumulation.

[0083] Dynamic baseline calibration is used to eliminate slow signal changes caused by sensor zero-point temperature drift, circuit drift, and changes in initial contact stress of gaskets. Impact separation processing is used to filter out instantaneous pressure peaks caused by working load impacts within a single forging cycle. These peaks are short-term, high-frequency elastic responses and do not represent permanent deformation.

[0084] The specific processing steps are as follows: 1. Baseline establishment and updating: During periods of mold idleness or when there is known no increase in plastic deformation, a signal is acquired and its mean is calculated as the initial baseline B0. During operation, a suitable moving-time window (e.g., covering the most recent 100 forging cycles) is used. Within the window, signal values ​​at non-forging moments in each cycle (e.g., when the slider is at the top dead center) are selected, and the sliding average is calculated as the observed baseline B0. obs Use a recursive filtering algorithm (such as a first-order low-pass filter) to filter B. obs Baseline B at the previous time point k-1 The fusion yields the updated dynamic baseline B. k B k =α·B obs +(1-α)·B k-1 1. Where α is the filter coefficient. 2. Impulse separation: The original signal S acquired in real time is separated. raw Subtract the current dynamic baseline B k The offset signal ΔS, after removing slow drift, is obtained. Time-domain analysis is performed on ΔS, an amplitude threshold is set, and transient spikes (corresponding to forging impact) with amplitudes far exceeding the long-term trend are identified and marked. Simultaneously, frequency-domain analysis (Fast Fourier Transform) is performed on ΔS, and a digital high-pass filter is designed to filter out high-frequency components related to the forging impact frequency. The impact components identified in both the time and frequency domains are removed from ΔS. 3. Signal extraction: The offset signal after impact removal is low-pass filtered to retain low-frequency variation components, mainly reflecting the slowly increasing or changing contact pressure caused by accumulated plastic deformation. The processed signal is then output, which can be used for further separation of the deformation signal Sdef(t).

[0085] Step S102: Based on the temporal changes of the deformation signal, analyze the dynamics of deformation development and quantify its cumulative process.

[0086] Timing variation refers to the deformation signal S def A sequence formed over time. The cumulative process refers to the overall manifestation of plastic deformation. The deformation signal refers to the processed signal Sdef that can be further analyzed.

[0087] The specific analysis and quantification process is as follows: 1. Dynamic analysis, for S def Time-frequency analysis (e.g., using short-time Fourier transform) is performed on the sequence to observe the distribution of signal energy across different frequency bands over time. Statistical characteristics (e.g., sliding variance, trend line slope) of the signal's energy or amplitude in key characteristic frequency bands (e.g., near DC or extremely low frequencies) are calculated to assess the stability of deformation growth. If the variance increases dramatically or the trend line slope increases significantly, it is determined that the deformation is at risk of acceleration or instability. 2. Stage division, based on stability assessment results and S... defThe long-term trend divides the deformation process into multiple stages, such as the "gap unclosed stage" (signal fluctuates near zero), the "linear slow growth stage," and the "nonlinear acceleration stage." 3. Quantization accumulation, for S def The signal can be directly integrated numerically or the portion exceeding the initial threshold can be accumulated to calculate the total cumulative plastic deformation A. total This physical quantity (or a quantity proportional to it) is the core indicator for quantifying the deformation process. For example, the signal S def Corresponding to pressure, then A total In terms of dimensions, it approximates "impulse" and indirectly reflects the work done by deformation or the accumulated plastic strain energy.

[0088] Step S103: Input the quantified deformation process information into the degradation assessment model trained based on historical service data to obtain assessment results including risk level and trend prediction.

[0089] The quantized deformation process information is a feature vector, including the current deformation signal value S. def Cumulative total A total The data includes the current growth rate, signal spectrum characteristics (such as the proportion of low-frequency energy), and the current deformation stage. The degradation assessment model is a time-series prediction model based on deep learning. The training data comes from the deformation process information sequences monitored by multiple sets of similar molds throughout their complete life cycle, as well as the real status labels corresponding to the end points of the sequences, such as "normal", "warning", "requires maintenance", and "failure".

[0090] The specific evaluation process is as follows: The system extracts the deformation process feature vectors in real time and organizes them into a fixed-length time window (e.g., containing the feature sequence of the most recent 1000 samples). This time window data is input into a pre-trained recurrent neural network model (LSTM or GRU network). The model learns the long-term dependencies in historical data and outputs two core evaluation results: 1. Risk level classification: Outputs the probability distribution of each predefined risk level, such as "low," "medium," and "high," and takes the level corresponding to the highest probability as the current judgment result. 2. Trend prediction: Outputs a prediction of deformation development in the future (such as within the next maintenance cycle), usually reflected as an estimate of the remaining useful life (e.g., predicting the number of times or days that can still be safely forged), or a prediction curve of the future cumulative deformation.

[0091] Step S104: Based on the evaluation results, trigger the corresponding graded early warning or deformation compensation control command.

[0092] Based on the evaluation results output in step S103, the system executes a preset hierarchical response strategy:

[0093] If the risk level is "low," a status alert is generated and sent, such as displaying a green status and cumulative deformation value on the local HMI interface, or uploading the status log to the management system, and production proceeds normally. If the risk level is "medium," a control command is generated to adjust the working parameters of the forging equipment through the equipment control network, such as reducing the forging speed by 10%-20% or reducing the closing height. Simultaneously, an audible and visual alarm is triggered to notify the operator. If the risk level is "high" or the predicted remaining lifespan is below the emergency threshold, the highest level alarm (such as SMS, audible and visual alarm) is immediately generated, and an interruption command is generated to stop production. If the equipment is equipped with a deformation compensation mechanism (such as an electric adjusting pad under the die holder), the system can simultaneously calculate the compensation amount based on the deformation signal characteristics and generate a deformation compensation control command to send to the compensation mechanism actuator.

[0094] Reference Figure 2 The steps for dynamic baseline calibration and impact separation processing of contact sensor signals include:

[0095] Step S200: Acquire the sensing signals of the mold base in a state of no plastic deformation or a specific operating stage, collect the reference signal sequence based on the moving time window, and establish its statistical characteristics as a dynamic baseline.

[0096] The state without plastic deformation refers to the initial state after mold installation and debugging and before formal production, or the state after major repair confirming that the mold base deformation has been repaired. A specific operating phase refers to a stage in the production process where it can be clearly determined that no new plastic deformation has occurred in the mold base, such as the initial stage of continuous production or during equipment no-load cycles. A moving time window is a virtual window that slides on the time axis to select a segment of continuous signal data; its length (e.g., covering the most recent N forging cycles or a fixed duration) can be set according to signal stability and update frequency requirements. The dynamic baseline is a reference signal value that updates slowly over time, representing the "baseline" or "zero point" level of the sensor signal due to environmental, circuit, and other factors under conditions without new plastic deformation interference.

[0097] The process of establishing a dynamic baseline is as follows: Under conditions of no plastic deformation or a specific operating phase, the system acquires a segment of raw contact sensor signal at a fixed sampling rate. This signal segment is preliminarily analyzed, and after removing obvious occasional interference pulses, its statistical characteristics are calculated as the initial dynamic baseline. The most commonly used statistical characteristic is the arithmetic mean. For example, when the mold is preheated but not yet forging, 10 seconds of signal data are acquired, and the average value is calculated as the initial baseline value B. initDuring subsequent continuous operation, the system maintains a moving time window, such as a window length set to the most recent 50 forging cycles. At each new processing cycle, the system slides the window forward (e.g., removing the oldest cycle data and adding the latest cycle data), and then calculates the average value of all signal points within the window that are in the "non-forging period" (e.g., signal segments where the slider is near the top dead center in each cycle). This average value is the observation baseline B for the current window. obs B obs Used for subsequent baseline updates.

[0098] Step S201: Compare the real-time acquired contact sensing signal with the dynamic baseline, calculate the signal offset, and use a recursive filtering algorithm to adaptively update the dynamic baseline to compensate for slow signal drift caused by environmental temperature drift or component aging.

[0099] This step includes: 1. Calculating the signal offset, the system at real-time sampling point t. k Obtain the raw contact sensor signal value S raw (t k Compare this with the current dynamic baseline value B(t). k-1 The instantaneous offset ΔS(t) is compared with the baseline updated at the previous time step. k )=S raw (t k )-B(t k-1 The offset can initially remove the influence of slow drift and directly reflect instantaneous changes. 2. Baseline adaptive update: To compensate for slow changes caused by environmental temperature drift, sensor zero-point drift, etc., the dynamic baseline itself needs to be able to track these slow disturbances, which is achieved through a recursive filtering algorithm. First-order low-pass filtering (exponential weighted moving average), the update formula is B(t). k )=α·B obs +(1-α)·B(t k -1), where B(t) k ) is the updated dynamic baseline, B obs The current moving window observation baseline obtained in step S200 is α, which is a smoothing factor (0 < α << 1) that determines the speed at which the baseline tracks slowly varying interference. The smaller α is, the smoother the baseline and the less sensitive it is to instantaneous changes, representing a long-term slowly varying trend. Through updates, the dynamic baseline B(t) can adaptively follow the background drift of the signal.

[0100] Step S202 involves identifying and separating high-frequency, high-amplitude signal components generated by instantaneous mechanical impact or vibration by combining time-domain peak detection with frequency-domain high-pass filtering.

[0101] Instantaneous mechanical shock or vibration refers to the huge and short force impact generated at the moment the die closes during each forging cycle. In terms of signal, it is a spike with a short duration (usually on the order of milliseconds) and an amplitude much higher than the baseline level.

[0102] The identification and separation operations are as follows: 1. Time-domain peak detection: Perform time-domain analysis on the signal offset sequence ΔS(t) calculated in step S201. Set a dynamic or static amplitude threshold V. th When |ΔS(t)| exceeds V th At this point, it is determined that the point may be an impact spike. To further confirm, the slope or amplitude changes of adjacent points before and after this point can be checked to see if they conform to the typical characteristics of an impact pulse (rapid rise, rapid fall). The signal components at the point marked as impact are considered interference. 2. Frequency domain high-pass filtering: Frequency domain analysis is performed on the ΔS(t) signal. Forging impact contains rich high-frequency components. ΔS(t) is filtered by a digital high-pass filter (such as a Butterworth high-pass filter or an FIR high-pass filter) to filter out components below the cutoff frequency f. c Element. f c It needs to be set to a frequency higher than the signal change frequency caused by plastic deformation in order to separate the high-frequency signal component ΔS, which is mainly contributed by impact. high (t). 3. Component combination and separation: The impact time period marked by time-domain detection is combined with the high-frequency component ΔS obtained by high-pass filtering in the frequency domain. high (t) is compared and combined. Finally, a comprehensive impact disturbance component I(t) is determined, which includes the main signal components generated by the instantaneous mechanical impact.

[0103] Step S203: Remove the impact interference component from the baseline-calibrated signal and output a low-frequency deformation signal that mainly reflects the cumulative plastic deformation of the mold base.

[0104] The final signal extraction is as follows: 1. Remove interference: Subtract the impulse interference component I(t) identified in step S202 from the signal offset ΔS(t) obtained in step S201 to obtain the preliminarily purified signal ΔS. clean (t) = ΔS(t) - I(t). 2. Low-frequency extraction: Since plastic deformation is a slow accumulation process, its corresponding signal changes are concentrated in the low-frequency range. For ΔS clean (t) Apply a digital low-pass filter (such as a Butterworth low-pass filter) to filter out any remaining high-frequency noise, preserving and smoothing its low-frequency trend components. Further slight smoothing (such as a moving average) can be applied to the filtered signal. 3. Signal output: The final output signal is the deformation signal S, which characterizes the accumulation of plasticity. def(t). This signal has been filtered out to the greatest extent possible from environmental drift and instantaneous impact interference. Its slowly changing trend and cumulative growth directly and reliably reflect the generation and development process of the mold base's plastic deformation. For example, S def (t) may be represented as a curve that starts near zero and slowly rises in a step-like or sloping manner as production time or number of forgings increases.

[0105] Reference Figure 3 The steps for outputting the deformation signal include:

[0106] Step S300: Based on the mechanical constitutive relationship of the mold base material, construct a signal separation model.

[0107] The mechanical constitutive relation of a die holder material refers to the physical law describing the relationship between stress and strain in a material (usually die steel) under load, particularly its elastoplastic behavior. The mechanical constitutive relation clarifies that material deformation occurs in two stages: the elastic deformation stage, where stress and strain are directly proportional, and the deformation is fully recovered after unloading; and the plastic deformation stage, where irreversible permanent deformation occurs when the stress exceeds the yield limit. The signal separation model is a mathematical or algorithmic model based on the aforementioned physical laws. Its core function is to inversely separate the components contributed by elastic deformation and plastic deformation from the input sensor signal (which has been freed from drift and impact interference, essentially reflecting the comprehensive deformation response of the die holder under forging pressure).

[0108] The process of constructing the signal separation model is as follows: First, key parameters such as the elastic modulus E, yield strength σy, and hardening modulus in the plastic stage of the mold material are obtained through material handbooks or experimental tests. Based on elastoplastic theory (such as the bilinear kinematic hardening model), a physical model reflecting the stress-strain response of the mold under cyclic loading is established. By combining this physical model with the generation mechanism of the sensing signal (i.e., the approximate relationship between the pressure signal and the overall compressive deformation of the mold), a signal separation model can be derived or trained. This model takes the sensing signal S as input, and its internal parameters reflect the constitutive relationship of the material. One feasible implementation is to use a state-space model (such as a Kalman filter), setting elastic deformation and plastic deformation as two state variables of the system, and estimating these two states by observing the signal (i.e., S). Another implementation is to use a data-driven method, generating a large amount of paired data of "signal-elastic component-plastic component" under known material parameters and simulated loads, to train a neural network model (such as a multilayer perceptron) as the signal separation model.

[0109] Step S301: Using the signal separation model, the signal after baseline calibration and removal of impact interference is analyzed into a first component corresponding to instantaneous elastic deformation and a second component corresponding to cumulative plastic deformation.

[0110] The first component (instantaneous elastic deformation component) refers to the portion of the signal corresponding to recoverable deformation that is instantaneously generated with the application of load in each forging cycle and instantaneously disappears with the removal of load. In the signal, this is represented by high-frequency fluctuations synchronized with the forging cycle. The second component (cumulative plastic deformation component) refers to the portion of the signal generated by irrecoverable plastic deformation that accumulates and increases with each forging cycle. In the signal, this is represented by a slowly rising "step" or "slope" trend.

[0111] The analysis process is as follows: The signal S output in step S203, which has undergone dynamic baseline calibration and impact separation processing, is... def (t) is used as input and fed into the pre-constructed signal separation model. The model processes the signal at each sampling point or in each forging cycle based on embedded physical laws or learned mapping relationships. For each input data point, the model outputs two values: one is the first component S. e (t) represents the magnitude of the purely elastic response at that moment; the other is the second component S. p (t) represents the accumulated signal substrate generated by pure plastic deformation up to that moment. def (t)≈S e (t)+S p (t). S e (t) fluctuates around zero, and the time integral is zero; S p (t) is a monotonically non-decreasing function whose change reflects the process of plastic deformation.

[0112] In step S302, the first component is not used, and the second component is separated and output as an incremental signal characterizing the development process of plastic deformation.

[0113] The decision not to adopt this approach refers to the analysis of the first component S in the signal separation model. e (t) and the second component S p After (t), the system intentionally ignores, discards, or does not include the first component S. e (t) is used in the subsequent deformation state analysis and evaluation process, because the first component S e (t) represents the instantaneous, recoverable elastic response, with a mean or cumulative value of zero, and does not reflect permanent damage accumulation in the model base structure. Separate output refers to explicitly selecting and transmitting the second component S from the two outputs of the signal separation model. p (t) to the subsequent data processing link. The incremental signal characterizing the development process of plastic deformation specifically refers to the second component S. p S(t) is a signal characterizing the monotonic accumulation process of plastic deformation, which can be directly used for time-series analysis. At each processing moment (e.g., each forging cycle), S... pThe value of (t) or the change relative to the previous moment directly characterizes the signal increment contributed by the newly generated, irreversible plastic deformation within that time interval. The temporal variation of this signal fully depicts the dynamic development process of plastic deformation gradually occurring, accumulating, and possibly accelerating.

[0114] The specific process is as follows: 1. Signal separation and selection. In step S301, the signal separation model has synchronously output the first component S. e (t) and the second component S p (t). The system sets a selector or logical decision in the data stream to explicitly select only the second component S. p (t) is passed to downstream modules as a valid result, while S is not stored or processed. e (t). 2. Output signal, the selected second component S p (t) is directly output as the final deformation signal characterizing plastic accumulation and is identified as an incremental signal. p (t) itself is a non-negative, monotonically non-decreasing (or step-like increasing) sequence with time. Each upward step or increase in slope in the sequence corresponds to the actual occurrence of plastic deformation within one period or time. 3. Signal meaning and examples, output S p (t) signal, the magnitude of which is proportional to the amount of plastic deformation. For example, assuming the signal separation model has been calibrated, S p (t) is in units of equivalent plastic micro-strain. During the forging process, S... p (t) may slowly and linearly increase from 0 to 50 units in the first 100,000 forging cycles, and then accelerate to 120 units in the next 50,000 forging cycles. Then, S from 0 to 120... p The (t) sequence is a complete incremental signal record of the plastic deformation process. Its current instantaneous value of 120 reflects the cumulative degree so far, while the trend over time depicts the entire process of deformation from slow initiation to accelerated development.

[0115] Reference Figure 4 The steps for constructing a signal separation model include:

[0116] Step S400: After the mold is first installed or overhauled, the initial key parameters of the model are calibrated by applying a series of known loads and simultaneously acquiring high-precision deformation data.

[0117] The series of known loads refers to a series of controllable and precisely measured static or quasi-static forces applied to the mold base using a press or a dedicated loading device during the mold debugging phase. Loads of 50kN, 100kN, 150kN... up to close to the mold's rated tonnage can be applied sequentially.

[0118] High-precision deformation data refers to the deformation (such as displacement or micro-strain) of key points of the mold base directly measured by a high-precision displacement measuring instrument (such as a laser displacement sensor, a capacitive micrometer, or a strain gauge attached to the mold base) when a known load is applied. For signal separation models based on physical constitutive relations, initial key parameters mainly refer to parameters in the model related to the mold base structural stiffness, material yield point, and hardening behavior, such as the equivalent elastic stiffness K mentioned in step S300. e Yield load F y wait.

[0119] The calibration process is as follows: After mold installation or major overhaul and before formal production, an offline calibration procedure is performed. The system-controlled loading device applies a first-level known load F1 to the mold base, simultaneously recording the signal value S1 output by the pressure sensor and the deformation D1 measured by the high-precision displacement sensor. Multiple load levels F are then applied sequentially. i Obtain the corresponding signal S i With deformation D i This forms a set of calibration data pairs {(F i ,S i D i These data are then input into a model parameter identification algorithm (such as least squares). The algorithm is based on the theoretical framework of the signal separation model (e.g., assuming that in the elastic phase S...). i With D i The relationship is linear, and the slope implies K. e When S i With D i When a clear inflection point appears in the relationship, the corresponding load can be considered as the yield load (Fy). A set of optimal model parameters is calculated to ensure that the model's predicted output (deformation) matches the measured high-precision deformation data D. i The parameter set with the smallest error is set as the initial key parameter of the model and serves as the benchmark for subsequent online monitoring.

[0120] Step S401: In the early stage of normal service of the mold, the initial key parameters are fine-tuned online using continuous forging cycle data and a recursive estimation algorithm to make the model output match the actual transient response of the mold base.

[0121] The initial stage of normal service refers to a period of time after the mold begins continuous production (such as the first few thousand forging cycles). During this stage, the macroscopic plastic deformation of the mold base is not significant, but microscopic contact and break-in may cause slight changes in characteristics. Recursive estimation algorithms are algorithms that can recursively update model parameters as new data arrives. Common examples include recursive least squares and Kalman filters.

[0122] The online fine-tuning process is as follows: After the mold begins normal production, the system continuously collects data for each forging cycle, including pressure sensor signals and the corresponding forging event timing. Since high-precision deformation measurement equipment cannot be installed on the production line in real time, the fine-tuning is based on the consistency within the model: that is, the dynamic characteristics of the elastic components analyzed by the model based on the current parameters and input signal (pressure) should match the transient response of the actual forging cycle (loading-unloading). For example, within a forging cycle, after the load is unloaded, the elastic components analyzed by the model should quickly return to zero, while the plastic components should remain constant. The system uses the recursive least squares method to continuously fine-tune the model's stiffness parameter K with the goal of minimizing the sum of the squares of the residuals of the elastic components output by the model (i.e., the deviations from the ideal unloading and zeroing state). e The process is gradual and adaptive, allowing the model to quickly "learn" and adapt to the actual dynamic mechanical properties of the specific mold.

[0123] Step S402: After detecting that the mold has experienced a maintenance event or replacement event, the drift compensation parameters in the model are reset or recalibrated according to the baseline offset of the deformation signals before and after the event.

[0124] Maintenance or replacement events include disassembly, grinding, welding repair, and reheat treatment of the mold base, or replacement of sensing elements or sensors. These events may directly change the mechanical boundary conditions of the mold base or the characteristics of the sensing system. Drift compensation parameters mainly refer to the internal parameters in the signal separation model used to handle systematic offsets such as sensor zero-point drift and small changes in mounting gaps, affecting the model's judgment of the signal "zero point" or "baseline".

[0125] The reset or recalibration process is as follows: The system has a maintenance event recording interface. This step is triggered when an operator records a maintenance event through the HMI or maintenance system, or when the system automatically detects a step change in sensor characteristics (considered a potential replacement event) by analyzing signal features. The system extracts and compares the deformation signal baseline (i.e., the stable value of the signal after processing in step S101 without forging) for a period of time before and after the event. The offset ΔB is calculated. If the offset ΔB exceeds a preset threshold (e.g., exceeding 5% of full scale), it is determined that the event has had a significant impact on the monitoring basis. The system will automatically reset the drift compensation parameter in the signal separation model, that is, clear it to zero or restore it to a default state. Then, based on the newly acquired short-term data after the event, a simplified online fine-tuning process (similar to S401, but with a shorter cycle) is re-executed to quickly recalibrate the parameter, ensuring that the model baseline is aligned with the new physical state.

[0126] Step S403: Continuously calculate the model prediction residuals. When the statistical characteristics of the residuals exceed the allowable range, generate a model failure warning and trigger the offline high-precision calibration process.

[0127] Model prediction residuals refer to a self-validation metric for the model. In this embodiment, it can be defined as the residual rk, which is the remaining value (theoretically zero) of the elastic component analyzed by the model after the load is completely unloaded in each forging cycle. This is because an ideal model should be able to completely separate the recoverable elastic component. Statistical characteristics refer to the feature values ​​calculated from the residual sequence {rk} over a continuous period of time, such as the mean, standard deviation, or root mean square value. When the model is working normally, these values ​​should be close to zero and remain stable.

[0128] The health monitoring and early warning process is as follows: The system continuously calculates the residual rk for each cycle in the background and maintains a residual sequence for a sliding window (e.g., the most recent 500 cycles). Periodically (e.g., every 100 cycles), the mean μr and standard deviation σr of the residuals within this window are calculated. An allowable range is set, for example, |μr| < ϵ1 and σr < ϵ2 (ϵ1, ϵ2 are thresholds set according to system accuracy requirements). If multiple consecutive calculations show that μr or σr continuously exceeds the allowable range, it indicates that the model's separation performance has degraded, possibly due to material property changes exceeding the model's applicable range, sensor characteristic drift not being fully compensated, or model parameter mismatch. At this time, the system generates a model failure early warning, indicating "signal separation model confidence has decreased," and, under safe conditions (such as during the next planned shutdown), prompts the operator to trigger the offline high-precision calibration process, requiring, as in step S400, to re-perform synchronous measurement and calibration under known loads, and comprehensively refresh the model's key parameters.

[0129] Reference Figure 5 The steps for analyzing the dynamics of deformation development and quantifying its cumulative process include:

[0130] Step S500: Perform time-frequency analysis on the incremental signal and extract its frequency domain distribution characteristics to evaluate the stability of the deformation process.

[0131] The incremental signal is the plastic deformation component S output in step S302. p (t). Time-frequency analysis is used to observe the changes in the frequency components of a signal over time; in this embodiment, short-time Fourier transform can be used. Frequency domain distribution characteristics include indicators such as dominant frequency, spectral centroid, and spectral entropy, which are used to quantify the distribution pattern of signal energy.

[0132] The specific evaluation process is as follows: The system evaluates the incremental signal S... p (t) or its rate of change signal is subjected to time-frequency analysis to generate a time-frequency spectrum. Frequency domain features of the current time period are extracted from the time-frequency spectrum at fixed intervals (e.g., every 1000 forging cycles), such as calculating the spectral centroid f. centroid And spectral entropy H. Stability is assessed by observing the trends of these eigenvalues ​​over time: if f centroidThe fact that H and H remain low and stable over a long period indicates that the deformation process is smooth; if f centroid A significant increase in H or H indicates potential instability in the deformation. The system can calculate the moving standard deviation of these characteristics as a quantitative indicator of stability.

[0133] Step S501: Based on the stability assessment results, divide the deformation evolution stages.

[0134] The partitioning process is based on the stability index calculated in step S500 and S p The growth trend of (t). The system predefines the following evolution stages: 1. Unclosed / emergent stage, S p The value of (t) remains close to zero, indicating extremely low stability; 2. In the linear stable growth phase, S p (t) linear and slow growth, with stability indicators remaining at a low level; 3. Nonlinear acceleration stage, S p (t) The growth rate accelerates, and / or the stability index continues to exceed the first threshold; 4. Instability warning stage, S p (t) The data exhibits drastic fluctuations or rapid increases, while stability indicators are severely exceeded, and the spectrum shows diffuse high-frequency components. The system determines the characteristics satisfied by the current data in real time, assigning a stage label to the current moment. This label serves as an important feature input into the subsequent evaluation model.

[0135] Step S502: Perform time-series accumulation operation on the incremental signal to quantify the total cumulative amount of plastic deformation.

[0136] Due to the incremental signal S p (t) itself is a monotonically non-decreasing sequence, and the current value represents the degree of accumulation of plastic deformation. Therefore, the time-series accumulation operation specifically refers to: directly using S p The current value P of (t) current As an instantaneous representation of the total accumulated plastic deformation; or its value relative to the initial value P at the monitoring starting point. initial The net increment ΔP=P current -P initial This is the cumulative net total.

[0137] The specific quantification process is as follows: The system continuously records S p (t). Its current value P current Or the net increment ΔP is the total cumulative amount of plastic deformation directly output. For example, if S p (t) is the unit of “equivalent plastic microstrain”. When its value increases from 0 to 120, the cumulative total ΔP is 120 microstrain. This value is the core quantitative indicator for assessing the degree of damage to the mold base.

[0138] Reference Figure 6 The degradation assessment model is a time-series prediction model built on a recurrent neural network. The training process of the degradation assessment model includes:

[0139] Step S600: Obtain the temporal feature sequence representing the deformation process as input.

[0140] The temporal feature sequence characterizing the deformation process refers to the sequence data formed by arranging the multidimensional features reflecting the deformation state of the mold base monitored over a continuous time period in chronological order. The feature vector of each time point (such as each forging cycle or fixed time interval) contains key indicators extracted and calculated from the original signal.

[0141] The process of constructing the input sequence is as follows: 1. Feature extraction: For each processing cycle (e.g., after each forging operation), the system extracts a set of predefined features from the results of step S102 (analyzing deformation development dynamics) and step S302 (outputting plastic accumulation signal) to form a feature vector x. t Features typically include fundamental time-domain characteristics: the current plastic deformation signal value S. p (t), the moving average and standard deviation of the signal within a time window; cumulative characteristics: the cumulative total of plastic deformation ΔP total The current value and its recent increment; stability characteristics: the current spectral centroid, dominant frequency, specific frequency band energy ratio, spectral entropy, or sliding variance of recent stability indicators extracted from time-frequency analysis; stage identifier: the code of the current deformation evolution stage (such as the stage divided in step S501) (e.g., using 0, 1, 2, 3 to represent different stages); working condition auxiliary characteristics: current forging times, average forging tonnage, production shift code, etc. 2. Sequence construction: the system maintains a fixed-length T first-in-first-out queue (e.g., T=500 or T=1000 cycles). Each new feature vector x t During generation, the vector is added to the end of the queue, while the old vector at the front of the queue is removed. In this way, a latest temporal feature sequence X=[x] of length T is always maintained. t−T+1 ,x t−T+2 ,...,x t As input to the model, the length T must be chosen to be sufficient to cover a complete trend cycle of deformation development.

[0142] Step S601: Learn the long-term dependencies and evolution patterns in the temporal feature sequences through its network structure.

[0143] Recurrent Neural Networks (RNNs) are deep learning models suitable for processing sequential data and possessing internal state memory. Long Short-Term Memory (LSTM) networks or gated recurrent units are preferred, as they effectively overcome the gradient vanishing / exploding problem of traditional RNNs and better capture long-term dependencies. Long-term dependencies and evolutionary patterns refer to the causal relationships, trends, and state transition characteristics exhibited by deformation features over a longer timescale (spanning hundreds or thousands of forging cycles). For example, how small changes in early spectral characteristics predict later accelerated growth; the transition patterns of risk levels after the cumulative total reaches a certain threshold, etc.

[0144] The model construction and training process is as follows: 1. Network architecture: Construct a multi-layer recurrent neural network. The input layer dimension is equal to the feature vector x. t The network has several LSTM or GRU hidden layers in the middle, which are used to process the input sequence step by step and extract high-level abstract features. The final output layer of the network is divided into two parallel "heads": a classification head, which is usually a fully connected layer followed by a Softmax activation function, and the output dimension is equal to the preset number of risk levels (e.g., 3 classes: low, medium, high), and the output value represents the probability of belonging to each level; and a regression head, which is usually a fully connected layer (which can use linear activation or ReLU, etc.), and outputs a scalar representing the predicted remaining useful life (e.g., in terms of the remaining forging times or the remaining days). 2. Model training: Collect a large amount of complete monitoring data of historical molds from their use to failure or overhaul (i.e., the end of their lifespan). For each mold, the monitoring data throughout its entire lifespan is cut into multiple overlapping time-series feature sequence samples of length T according to the method in step S600. Each sample is labeled with a true value: a classification label, assigned a risk level based on the actual historical state of the mold at the last time point (e.g., maintenance records, precision inspection reports); and a regression label, calculated as the number of forging operations or days elapsed from the last time point to the actual end of the mold's lifespan, representing the remaining lifespan as the true value. A multi-task learning framework is employed, with the total loss function L... total The cross-entropy loss L is designed for classification tasks. cls Mean squared error loss L for regression task re Weighted sum of g: L total =λL cls +(1-λ)L re g, where λ is a hyperparameter used to balance the two tasks. The backpropagation algorithm and optimizer (such as Adam) are used to minimize the total loss on the training dataset, iteratively updating the network weights until the model converges. After training, the model learns the ability to infer the current risk level and predict remaining lifespan from time-series feature sequences.

[0145] Step S602: Output the assessment results, which include discrete risk level classification and continuous remaining life prediction.

[0146] In real-time monitoring, the system inputs the latest time-series feature sequence X, which is currently being maintained, into the trained degradation assessment model. After forward propagation, the model outputs two results: 1. Discrete risk level classification result: The model's "classification head" outputs a probability vector, such as P=[plow, pmed, phigh]=[0.15, 0.35, 0.50]. The system takes the category with the highest probability (here, "high") as the current risk level and can also output this probability vector as a confidence level reference. 2. Continuous remaining lifetime prediction result: The model's "regression head" outputs a numerical value RUL. pred For example, 8500 times. This value is the number of forging cycles predicted by the model that the die holder can still operate safely under the current conditions. The system compares this predicted value with preset warning thresholds at various levels (e.g., warning threshold = 2000 times, severe threshold = 500 times). Finally, a structured assessment result such as "high risk, remaining life predicted to be approximately 8500 times" is output and transmitted to the warning and control module.

[0147] Reference Figure 7 The steps to trigger the corresponding tiered warning include:

[0148] Step S700: When the evaluation result indicates that the deformation risk is low, generate and send a status prompt message.

[0149] Status alerts are a type of low-intrusive, notification-based information output designed to inform relevant personnel that the current status of the module is good and does not require immediate intervention, but should be monitored.

[0150] The generation and transmission process is as follows: The system receives the evaluation results from the degradation assessment model. If the current risk level is determined to be "low", this step is triggered. The status prompt information typically includes: 1. Core status indicators, such as the current cumulative total plastic deformation (ΔPtotal), deformation signal value, and predicted remaining life; 2. Visual indicators, on the local human-machine interface (such as a touch screen or industrial control computer monitor), setting the corresponding mold status indicator light to solid green, and dynamically updating the above indicators in a dedicated area in the form of numbers or trend charts; 3. Log records, generating a record containing a timestamp, equipment identifier, and status summary (such as "Status normal, cumulative deformation XX micro-strain, predicted remaining life YY times"), and sending it to the local database or uploading it to a remote manufacturing execution system / IoT platform for long-term traceability and statistical analysis.

[0151] This step only involves information dissemination and does not change any operating parameters of the production equipment, ensuring the continuity and high efficiency of the production process.

[0152] Step S701: When the evaluation results indicate an increased risk of deformation, control commands are generated to limit or adjust the operating parameters of the forging equipment.

[0153] Control commands are a type of digital commands that directly act on the control system of forging equipment. They are used to slow down the further development of plastic deformation of the die holder by actively reducing the equipment load or changing the working mode, thus buying time for planned maintenance.

[0154] The generation and execution process is as follows: When the assessment result is "medium" risk level, the system immediately triggers this step. Control commands are sent to the programmable logic controller or CNC system of the forging equipment through a preset communication interface (such as industrial Ethernet or fieldbus). The specific content of the commands is based on a preset optimization strategy library and usually includes adjustments to one or more of the following operating parameters: 1. Reduce forging speed: Reduce the slide running speed by a percentage (e.g., 15%-25%). Reducing the speed can reduce the peak impact load and dynamic effects, thereby mitigating the instantaneous impact force on the die holder; 2. Adjust closing height / pressure: Fine-tune the final closing position of the slide to make it slightly higher than the nominal position, or reduce the system working pressure setting value to directly reduce the maximum forming force acting on the die holder; 3. Within the allowable process range, shorten the holding time to reduce the continuous stress time of the die holder under high pressure; 4. Enable degraded production mode. For multi-station dies, some non-critical or high-load stations may be skipped.

[0155] Simultaneously with the generation of the instruction, the system triggers a flashing yellow alarm and text prompt on the HMI (such as "Deformation acceleration detected, automatic speed reduction has been implemented, inspection recommended"). After receiving the instruction, the equipment controller smoothly transitions to the new parameter settings in the next one or more production cycles, thereby proactively controlling production while maintaining output.

[0156] Step S702: When the evaluation results indicate that the deformation risk has reached the highest level, a control command is generated to initiate deformation compensation and interrupt the forging operation.

[0157] Initiating deformation compensation refers to calling upon the micro-displacement compensation mechanism integrated into the mold base or externally attached to attempt to physically correct mold closing accuracy deviations caused by deformation. Interrupting forging operations means immediately stopping the automatic production cycle and placing the equipment in a safe standby or shutdown state to prevent the malfunction from escalating or producing scrap.

[0158] The generation and execution process is as follows: When the assessment result is a "high" risk level, or the predicted remaining lifespan is lower than the set emergency threshold, the system triggers the highest level response and performs the following operations sequentially or in parallel: 1. Complete and send an interrupt command, immediately sending an emergency stop or cyclic interrupt command to the forging equipment control system. After receiving the command, the equipment stops automatic operation after completing the current forging cycle (or immediately enters the emergency stop sequence, depending on the safety strategy), the slider returns to the top dead center and locks. 2. Trigger the highest level alarm, displaying a flashing red alarm and a critical warning message on the HMI, and sending alarm information including the equipment number and risk details (such as "severe deformation, cumulative amount ZZ, predicted lifespan only WW cycles remaining") to the maintenance engineer or production supervisor via the factory broadcast system, SMS, or application push. 3. Calculate and generate deformation compensation commands (if supported by the equipment). Based on the current deformation signal characteristics (such as signal differences from multiple sensors), the system uses a built-in compensation algorithm to calculate the compensation amount required to offset the main deformation error (such as the micron-level displacement adjustment required for each of the four corner compensation pads under the mold base). Then, it generates control commands for the deformation compensation mechanism (such as piezoelectric ceramic actuators or servo-electric adjusting wedges) to drive it to perform compensation actions during equipment downtime or maintenance windows. 4. Record and report: Generate detailed fault warning reports, recording all key data, evaluation results, and executed operations at the time of the event, and mandatorily report to the management system.

[0159] Based on the same inventive concept, embodiments of this application provide a mold base deformation monitoring system, including:

[0160] The acquisition module is used to acquire contact sensing signals triggered by the plastic deformation of the mold base, which are generated by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element.

[0161] A memory for storing programs for the mold base deformation monitoring method as described above;

[0162] The processor and memory can load and execute programs to implement the mold base deformation monitoring method described above.

[0163] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0164] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a method for monitoring mold base deformation.

[0165] Computer storage media include, for example, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media that can store program code.

[0166] Based on the same inventive concept, embodiments of this application provide a smart terminal, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as a mold base deformation monitoring method.

[0167] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0168] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for monitoring the deformation of a mold base, characterized in that, include: The contact sensing signal triggered by the plastic deformation of the mold base is obtained by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element. Dynamic baseline calibration and impact separation processing are performed on contact sensing signals to filter out transient interference and compensate for signal drift, and to extract deformation signals that characterize plastic accumulation. Based on the temporal variation of deformation signals, we analyze the dynamics of deformation development and quantify its cumulative process; The quantified deformation process information is input into a degradation assessment model trained based on historical service data to obtain assessment results that include risk level and trend prediction. Based on the assessment results, corresponding graded early warnings or deformation compensation control commands are triggered.

2. The method for monitoring mold base deformation according to claim 1, characterized in that, The steps for dynamic baseline calibration and impact separation processing of contact sensor signals include: The sensor signals of the mold base under non-plastic deformation state or specific operating stage are acquired, and the reference signal sequence is collected based on the moving time window and its statistical characteristics are established as a dynamic baseline. The real-time acquired contact sensing signal is compared with the dynamic baseline to calculate the signal offset. The dynamic baseline is then adaptively updated using a recursive filtering algorithm to compensate for slow signal drift caused by environmental temperature drift or component aging. By combining time-domain peak detection with frequency-domain high-pass filtering, high-frequency, high-amplitude signal components generated by instantaneous mechanical shock or vibration are identified and separated. After removing the impact interference component from the baseline-calibrated signal, the output is a low-frequency deformation signal that mainly reflects the cumulative plastic deformation of the mold base.

3. The method for monitoring mold base deformation according to claim 2, characterized in that, The steps for outputting the deformation signal include: A signal separation model is constructed based on the mechanical constitutive relationship of the mold base material; Using a signal separation model, the signal after baseline calibration and removal of impact interference is analyzed into a first component corresponding to instantaneous elastic deformation and a second component corresponding to cumulative plastic deformation. The first component is not used; the second component is separated and output as an incremental signal characterizing the development process of plastic deformation.

4. The method for monitoring mold base deformation according to claim 3, characterized in that, The steps for constructing a signal separation model include: After the mold is first installed or overhauled, a series of known loads are applied and high-precision deformation data is acquired simultaneously to calibrate the initial key parameters of the model. In the early stages of normal service of the mold, the initial key parameters are fine-tuned online using continuous forging cycle data and a recursive estimation algorithm to match the model output with the actual transient response of the mold base. After detecting a maintenance or replacement event in the mold, the drift compensation parameters in the model are reset or recalibrated based on the baseline offset of the deformation signals before and after the event. The model prediction residuals are continuously calculated. When the statistical characteristics of the residuals exceed the allowable range, a model failure warning is generated and an offline high-precision calibration process is triggered.

5. The method for monitoring mold base deformation according to claim 3, characterized in that, The steps for analyzing the dynamics of deformation development and quantifying its cumulative process include: Time-frequency analysis is performed on the incremental signal to extract its frequency domain distribution characteristics in order to assess the stability of the deformation process; Based on the stability assessment results, the deformation evolution stages are divided; The incremental signal is subjected to time-series accumulation operation to quantify the cumulative total amount of plastic deformation.

6. The method for monitoring mold base deformation according to claim 1, characterized in that, The degradation assessment model is a time-series prediction model built on a recurrent neural network. The training process of the degradation assessment model includes: The temporal feature sequence representing the deformation process is obtained as input; It learns long-term dependencies and evolution patterns in temporal feature sequences through its network structure; The output includes assessment results that include discrete risk level classification and continuous remaining life prediction.

7. The method for monitoring mold base deformation according to claim 1, characterized in that, The steps to trigger the corresponding tiered warning include: When the assessment result indicates that the deformation risk is low, a status alert message is generated and sent. When the assessment results indicate an increased risk of deformation, control commands are generated to limit or adjust the operating parameters of the forging equipment. When the assessment results indicate that the deformation risk has reached the highest level, a control command is generated to initiate deformation compensation and interrupt the forging operation.

8. A mold base deformation monitoring system, characterized in that, include: The acquisition module is used to acquire contact sensing signals triggered by the plastic deformation of the mold base, which are generated by monitoring the closed state of the preset mechanical gap between the mold base and the sensing element. A memory for storing the program of the mold base deformation monitoring method as described in any one of claims 1 to 7; The processor and the program in the memory can be loaded and executed by the processor to implement the mold base deformation monitoring method as described in any one of claims 1 to 7.

9. A smart terminal, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7 for monitoring the deformation of the mold base.

10. A computer-readable storage medium, characterized in that, The computer program is stored and can be loaded by a processor and executed as described in any one of claims 1 to 7 for monitoring mold base deformation.

Images