Multi-press line energy consumption model self-calibration method and system under dynamic load

Abstract

The invention discloses a self-calibration method and system for an energy consumption model of a multi-press line under a dynamic load. According to the method, mechanical, motion and electrical signals are synchronously collected on the edge side of each punching machine, periodic feature vectors are extracted, and quality filtering is carried out through an online robust random cutting forest model; utilizing effective data to locally train an energy consumption prediction model composed of a global physical shared core and a local personalized adapter head, wherein training loss comprises two physical constraints including instantaneous power balance and periodic energy conservation; and the global kernel gradient encryption is uploaded to a central collaboration device, the central collaboration device performs weighted security aggregation in combination with the data quality evaluation result of each device, and global parameters are updated and issued. By introducing the physical constraint and federated learning mechanism, the adaptability and prediction consistency of the model under the dynamic working condition are improved, manual intervention is not needed, and the method is suitable for stamping production line energy efficiency management under the complex dynamic working condition.

Description

Technical Field

[0001] This invention belongs to the field of xxxx technical technology and relates to a self-calibration method and system for energy consumption model of multi-machine stamping line under dynamic load. Background Technology

[0002] In manufacturing industries such as automobiles and home appliances, stamping production lines are crucial production equipment and also major energy consumers. To achieve energy conservation, emission reduction, and refined energy management, companies urgently need accurate predictions of energy consumption in the stamping process. However, existing technologies still have significant shortcomings in constructing practical and reliable energy consumption models.

[0003] Traditional methods mainly rely on static mechanistic models, which calculate energy consumption based on equipment nameplate parameters (such as motor efficiency and transmission ratio) and theoretical formulas. These models assume that equipment performance remains constant over a long period of time. However, in actual operation, the efficiency of a stamping press will continuously drift due to factors such as lubrication aging, component wear, and temperature changes. This causes the model accuracy to decline rapidly over time, and the error usually increases significantly after a few weeks of operation, making it difficult to meet the requirements of long-term online applications.

[0004] Another approach employs centralized data-driven models (such as neural networks), training prediction models with a large amount of historical data. While this method offers high accuracy under specific operating conditions, it suffers from two significant drawbacks: First, data from various stamping presses is typically scattered across different workshops or systems, making it impossible to centrally upload the original datasets due to data security and commercial confidentiality concerns. Second, purely data-driven models lack physical constraints, and when encountering unfamiliar operating conditions such as new molds or processes, they may output unreasonable results that violate fundamental physical laws such as the conservation of energy, rendering them unreliable and unusable on-site.

[0005] Furthermore, current industrial systems generally lack automatic calibration capabilities. Most energy consumption monitoring relies on fixed formulas or simple table lookups, failing to detect changes in equipment status (such as mold replacement or performance degradation). Once the model becomes inaccurate, manual intervention for recalibration is necessary, resulting in slow response and high costs.

[0006] Therefore, a new energy consumption modeling and self-calibration method is needed that can integrate physical laws, support multi-machine collaborative learning, protect local data privacy, and automatically adapt to dynamic operating conditions, in order to address the shortcomings of existing technologies in terms of accuracy, robustness, reliability, and automation. Summary of the Invention

[0007] To address the problems existing in the background technology, this invention proposes a self-calibration method and system for the energy consumption model of a multi-machine stamping line under dynamic load.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a self-calibration method for energy consumption model of a multi-machine stamping line under dynamic load, comprising: At each stamping press, the local model training step is performed: based on the synchronously acquired stamping process signals of the press, multi-dimensional periodic feature vectors are extracted; The periodic feature vector is quality filtered by unsupervised anomaly detection and dynamic threshold comparison. Using filtered periodic feature vectors, a local personalized adapter head for a local energy consumption prediction model is trained, and the gradient of the global physical shared kernel is calculated. The total loss function for model training includes a physical constraint loss term, which includes an instantaneous power balance constraint term and a periodic energy conservation constraint term. In the central coordination device, the model coordination aggregation step is executed: the gradient of the global physical shared kernel is received by each press in encrypted form; The received gradients are weighted and aggregated based on the quality assessment results of the data uploaded by each stamping press to update the parameters of the global physical shared kernel, and the updated parameters are sent to each stamping press. The local energy consumption prediction model consists of the globally shared physical core and the local personalized adapter head connected in series.

[0009] Specifically, the instantaneous power balance constraint term is constructed by making the energy consumption change rate predicted by the model approximate the reference power determined based on the measured force signal, speed signal and equipment reference efficiency function; The periodic energy conservation constraint is constructed by making the model-predicted total energy consumption per cycle approximate the reference total energy consumption determined based on the measured deformation work and the inherent energy loss of the equipment.

[0010] Specifically, the equipment reference efficiency function is obtained by performing efficiency-speed characteristic tests on the drive system of the corresponding stamping press under no-load and load conditions, and by performing function fitting on the discrete efficiency points obtained from the tests. The inherent energy loss of the equipment is obtained by conducting continuous no-load operation tests on the corresponding stamping press and statistically analyzing its average cycle energy consumption.

[0011] Specifically, the periodic feature vector is input into an online-updated robust random cutting forest anomaly detection model to calculate the anomaly score; The abnormal score is compared with a dynamic threshold, which is adaptively determined based on the statistical distribution characteristics of the abnormal scores within a historical time period.

[0012] Specifically, based on the average anomaly score corresponding to the batch data uploaded by each stamping machine, the weight of the gradient contributed by that batch data in the aggregation calculation is dynamically adjusted so that the batch with the lower anomaly score contributes more to the global model update.

[0013] Specifically, the local training step of the model is triggered by an event or condition; Triggering events include receiving a mold change command; Triggering conditions include the model prediction error continuously exceeding the allowable range for a predetermined time, or the number of valid periodic feature vectors accumulated in the local training buffer reaching a preset threshold.

[0014] Specifically, the global physical shared kernel is a multi-layer feedforward neural network, whose parameters are jointly updated by all stamping presses participating in collaborative training; The local personalized adapter is a neural network, and its parameters are updated only locally using local data on the corresponding stamping press.

[0015] Based on the above-mentioned self-calibration method for energy consumption model of multi-machine stamping line under dynamic load, the present invention further proposes a self-calibration system for energy consumption model of multi-machine stamping line under dynamic load, including: a central coordination device and multiple edge processing devices communicatively connected to the central coordination device; The edge processing devices are installed one-to-one with each stamping machine and are used for: In response to stamping cycle events, the machine's mechanical signals, motion signals, and electrical signals are simultaneously acquired and extracted into multi-dimensional periodic feature vectors. The multidimensional periodic feature vector is quality filtered using an unsupervised anomaly detection model. A local energy consumption prediction model is trained locally using filtered periodic feature vectors. The local energy consumption prediction model consists of a series of global physical shared kernels and a local personalized adapter head. The local training includes: updating the parameters of the local personalized adapter head with a total loss function that includes a physical constraint loss term, and calculating the gradient of the global physical shared kernel; The calculated gradient of the global physical shared core is encrypted and uploaded to the central collaborative device; And receive global parameters issued by the central coordination device, and update the local global physical shared core; The central coordination device is used for: Receive the encrypted gradient uploaded by each edge processing device; Based on the quality assessment information of the data uploaded by each edge processing device, the received gradients are weighted and securely aggregated to update the parameters of the global physical shared kernel; The updated global parameters are then distributed to each edge processing device.

[0016] Specifically, the physical constraint loss term includes: an instantaneous power balance constraint term, which is used to constrain the energy consumption change rate predicted by the model to be consistent with the reference instantaneous power calculated based on the measured force signal, the slider speed signal and the equipment reference efficiency function; In addition, there is a periodic energy conservation constraint term, which is used to constrain the model prediction of the total energy consumption per cycle, which is consistent with the sum of the measured deformation work and the inherent energy loss of the equipment.

[0017] Specifically, the unsupervised anomaly detection model is an online-updated robust random cutting forest model; The quality assessment information is the periodic feature vector anomaly score calculated by the data quality filtering module. The security aggregation module is configured to dynamically adjust the weight of its gradient in the weighted aggregation based on the average anomaly score corresponding to the data uploaded by each edge processing device.

[0018] Compared with the prior art, the present invention has the following advantages: by introducing physical constraints such as instantaneous power balance and periodic energy conservation, it ensures that the model output conforms to the law of energy conservation, avoids prediction results that do not conform to physical laws, and improves engineering practicality.

[0019] It adopts a global physical shared core and local personalized adapter head architecture, which not only shares the general rules across devices, but also accurately adapts to individual differences caused by wear, lubrication and other factors in a single machine.

[0020] It supports fully automatic online self-calibration. The system can automatically detect mold changes, performance drift, or abnormal operating conditions, trigger local training, and filter low-quality data, enabling long-term stable operation without manual intervention.

[0021] Ensuring data privacy and security. Raw sensor data remains locally transmitted; only encrypted gradients and lightweight quality scores are uploaded, enabling collaborative learning while preventing the leakage of sensitive information.

[0022] The aggregation process is more robust. The central coordination device dynamically weights the data based on the average anomaly score of each device, resulting in a greater contribution from high-quality data and effectively suppressing the interference of noisy or faulty devices on the global model. Attached Figure Description

[0023] Figure 1 This is a flowchart of a self-calibration method for energy consumption model of a multi-machine stamping line under dynamic load according to the present invention; Figure 2 This is a block diagram of a self-calibration system for energy consumption model of a multi-machine stamping line under dynamic load according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figures 1-2 As shown, the technical solution adopted in this invention is as follows: A self-calibration method for energy consumption model of a multi-machine stamping line under dynamic load, comprising: S1: On each stamping press, perform the local model training step: extract multi-dimensional periodic feature vectors based on synchronously acquired stamping process signals. This step is performed by the edge processing device deployed on each stamping press, completing data acquisition, feature engineering, and model training locally, without relying on the central coordinating device's access to the raw data.

[0026] Synchronous acquisition utilizes the Z-phase signal of the incremental encoder installed on the main motor as a global hardware trigger source. Through a synchronization mechanism consisting of an industrial Ethernet switch that supports the IEEE 1588v2 precision clock protocol and Time Sensitive Network (TSN) features, it ensures that the three types of sensor signals—force sensor, grating ruler, and power measurement unit—are strictly aligned on the time axis, with the sampling rate uniformly set to 10kHz.

[0027] The stamping process signals of this machine include: the real-time stamping force on the slide block. Slider displacement Motor input power The data are collected by a high-precision force sensor, a grating ruler, and a power measurement unit, respectively.

[0028] The multidimensional periodic eigenvector is a 12-dimensional real-valued vector:

[0029] ; in: This represents the maximum impact force within a single cycle. This represents the minimum impact force within a single cycle. This represents the average pressure time within a single cycle. ,like If , then set it as the maximum value; The average velocity of the slider during the downward phase is calculated by dividing the displacement difference by the downward phase time. This represents the average speed of the slider during the return stroke. This represents the maximum instantaneous velocity of the slider within a single cycle. The deformation work is determined by analyzing the force-displacement curves acquired synchronously. The Simpson numerical integration method is used to calculate the effective stroke range: ; The duration of a complete stamping cycle; This refers to the duration of the downward phase. This refers to the duration of the return journey phase. This represents the predicted energy consumption value output by the local energy consumption prediction model for the previous stamping cycle. Feature extraction is completed by edge intelligent computing nodes within 50ms after the end of each stamping cycle to ensure real-time performance.

[0030] Specifically, the local training step is triggered by events or conditions. Local training is not continuous; instead, it employs a hybrid triggering mechanism combining event-driven and periodic accumulation to balance computational load with the timeliness of model updates. The triggering logic is implemented by a state monitoring module within the edge processing device, initiating the local training process only when any preset triggering condition is met.

[0031] Triggering events include receiving a die change command. The die change command is a structured process switch signal from the Manufacturing Execution System (MES), containing a new die identifier and process parameter change information. This event indicates a structural change in the stamping load characteristics (such as changes in forming force distribution or stroke curve), rendering the original local personalized adapter parameters inapplicable and requiring immediate model fine-tuning.

[0032] The edge processing device listens to the event bus published by MES via OPC UA or MQTT protocol. Once it parses the mold change instruction, it immediately clears the current training buffer and marks the next cycle data as the starting point of the new working condition, and then triggers local training.

[0033] Triggering conditions include the model prediction error continuously exceeding the allowable range for a predetermined time, or the number of valid periodic feature vectors accumulated in the local training buffer reaching a preset threshold.

[0034] The first type of trigger condition is when the prediction error continues to exceed the limit: the model prediction error is defined as the measured energy consumption of the current cycle. With predicted energy consumption Relative deviation: ; Calculate its exponentially weighted moving average (EWMA): ; like And the inequality is satisfied continuously. If a cycle is completed, it is determined that the equipment status has drifted (such as lubrication deterioration or transmission efficiency reduction), triggering local training.

[0035] The second type of trigger condition is a full buffer: the local training buffer is a first-in-first-out (FIFO) queue, and its capacity limit is set to... There are 10 valid periodic feature vectors; the counter is incremented by one each time a new periodic feature vector passes the quality filter. When the counter... When the time comes, local training is triggered and the counter is reset.

[0036] The two types of triggering conditions mentioned above complement each other: the former is for sudden changes in operating conditions, and the latter is for slow performance degradation, together ensuring the model's adaptive capability.

[0037] S11: The periodic feature vector is quality filtered by unsupervised anomaly detection and dynamic threshold comparison.

[0038] This operation is performed by the data quality filtering module in the edge processing device, and its input is the periodic feature vector extracted in the previous step. .

[0039] Quality filtering refers to determining whether the characteristic vector of a cycle originates from the normal stamping process. If it originates from non-steady-state conditions such as sensor momentary failure, double feeding, slider jamming, or hydraulic shock, it is considered invalid data and is removed.

[0040] Unsupervised anomaly detection means that the judgment process does not rely on manually labeled normal / abnormal data, but is automatically modeled based on the statistical patterns implied in historical normal production data.

[0041] Dynamic threshold comparison means that the judgment criterion is not a fixed value, but an adaptive threshold that evolves over time to cope with normal data distribution drift caused by equipment aging, mold switching, and environmental temperature changes. This step is a key preprocessing step before local model training to ensure the process consistency and physical rationality of the training data.

[0042] Specifically, the periodic feature vector is input into an online-updated robust random-cut forest anomaly detection model to calculate the anomaly score.

[0043] The robust random cut forest anomaly detection model is a Robust RandomCut Forest (RRCF) model deployed on an edge processing device, and its structure consists of... It consists of randomly cut trees.

[0044] Online updates refer to the ability of a model to learn incrementally using streaming data and update its internal structure in real time.

[0045] During construction, each random cut tree is constructed by randomly selecting a dimension from the 12-dimensional feature space and randomly selecting a cut point within the current data range of that dimension, recursively splitting until a leaf node contains a single data point or the minimum depth limit is reached.

[0046] The RRCF model supports streaming online updates: whenever a new periodic feature vector is introduced... Upon arrival, the model first calculates its anomaly score. Then insert it into all tree structures to achieve continuous tracking of data flow distribution.

[0047] Abnormal scores Defined as the change in average cut depth caused by this vector across all trees, mathematically expressed as: ; in Indicates the first Insert into a tree The resulting conformal fraction reflects the degree of disturbance that point causes to the existing structure.

[0048] The larger, the more it indicates The more it deviates from the historical normal pattern, the more likely it is to be abnormal.

[0049] The RRCF model maintains a length of The sequence context window is used to prioritize the structural representativeness of recent data when inserting new points, thereby enhancing sensitivity to short-term operating condition changes.

[0050] The abnormal score is compared with a dynamic threshold, which is adaptively determined based on the statistical distribution characteristics of the abnormal scores within a historical time period.

[0051] The anomaly score is the output from the previous step. The dynamic threshold is denoted as... Its value is updated in real time at the end of each stamping cycle.

[0052] The system maintains a sliding statistics window in the edge processing device. , where window length It stores the anomaly scores for the most recent 1000 valid periods.

[0053] Defined as The 99.5th percentile of all elements in the set, i.e.: ; in For indicator functions, This represents the time step number within the sliding window, with values ​​ranging from 1 to... , with global time index Correlation. If the current period satisfies If this happens, the periodic feature vector is determined to be abnormal, will not be stored in the local training buffer, and may trigger a local alarm log recording. If the data is valid, it will be processed in the next step.

[0054] This mechanism allows the threshold to be automatically adjusted as the device state slowly drifts (such as increased friction causing the force signal to shift upwards), avoiding false filtering or missed filtering caused by a fixed threshold.

[0055] Specifically, based on the average anomaly score corresponding to the batch data uploaded by each stamping machine, the weight of the gradient contributed by that batch data in the aggregation calculation is dynamically adjusted so that the batch with the lower anomaly score contributes more to the global model update.

[0056] After local training is complete, the edge processing device will assign anomaly scores to all valid periodic feature vectors used in this training round. The arithmetic mean was calculated to obtain the value of the stamping machine. Average anomaly score in this round ,Right now: ; in This represents the number of valid samples that passed the quality filter in this batch. For the first The serial number of the valid samples from this round of stamping presses.

[0057] The edge processing device will use the gradient of the encrypted global physical shared kernel. With metadata They are uploaded together to the central collaborative device.

[0058] Before performing gradient aggregation, the central coordination unit provides each stamping press with... Calculate aggregate weights Its definition is: ; in This is the anomaly penalty coefficient, with a value range of [0.1, 5.0], and a typical value of [value missing]. This parameter can be determined through a grid search on a small validation set.

[0059] The aggregated global gradient, i.e., the weighted aggregated gradient Calculated as a weighted average: ; in This refers to the collection of stamping presses participating in this round of aggregation.

[0060] Because the exponential function is monotonically decreasing The smaller the value, the higher the data quality. The larger the gradient, the stronger its impact on the global model update. Conversely, for stamping presses with high outlier scores, if noise is continuously generated due to sensor loosening, the gradient contribution is suppressed, thereby achieving robust federated aggregation and preventing dirty data from polluting shared knowledge.

[0061] This mechanism directly feeds back the data quality assessment results from the edge to the central collaboration process, forming a closed-loop adaptive architecture of perception, assessment, weighting, and evolution.

[0062] S12: Using the filtered periodic feature vectors, train the local personalized adapter head of the local energy consumption prediction model and calculate the gradient of the global physical shared kernel. The total loss function of the model training includes a physical constraint loss term, which includes an instantaneous power balance constraint term and a periodic energy conservation constraint term.

[0063] The filtered periodic feature vectors refer to the periodic feature vectors that are determined to be valid by unsupervised anomaly detection and dynamic threshold comparison, and have been stored in the local training buffer of the edge processing device.

[0064] The local training buffer is a first-in-first-out (FIFO) queue with a maximum capacity of [missing information]. It only stores the effective periodic feature vector and its corresponding measured single-cycle energy consumption value. .

[0065] When local training is triggered, the training module randomly and uniformly samples from this buffer. These records constitute the dataset for the current training batch: ; superscript Indicates the first Taiwanese stamping machine For the first The periodic feature vector of each sample, The corresponding measured energy consumption is obtained by integrating the electrical power signal within the stamping cycle using the power measurement unit.

[0066] The local energy consumption prediction model consists of a cascaded global physical shared kernel and a local personalized adapter head, and its overall mapping relationship is as follows: ; in Represents a globally shared physical kernel (multilayer feedforward neural network), with parameters as follows: ; This indicates a local personalized adapter (lightweight feedforward network), with the following parameters: .

[0067] During local training, the parameters of the globally shared physical kernel are... Frozen and not involved in parameter updates, but the computation graph is preserved for backpropagation of gradients; only parameters of the locally customized adapter head are updated. Optimized.

[0068] Total loss function of model training Defined as: ; in , These are preset physical constraint weighting coefficients used to balance data fitting and physical consistency.

[0069] Training uses the AdamW optimizer (learning rate) Weight decay ), minimize To update .

[0070] Simultaneously, the system calculates the global physical shared kernel parameters. right gradient This gradient will be used in subsequent federated aggregation steps, but It is not updated locally.

[0071] The data fitting loss term The mean square error (MSE) characterizes the approximation of the local energy consumption prediction model to the measured energy consumption data. It is defined as the mean square error between the predicted total energy consumption per cycle and the measured total energy consumption per cycle for all samples in the current training batch. ; in: This refers to the training batch size; For the local energy consumption prediction model for the first 100 filtered periodic feature vectors The predicted energy consumption output; This corresponds to the actual energy consumption measured and integrated by the power measurement unit during the stamping cycle. This loss term serves as a monitoring signal, driving the local personalized adapter head parameters. The update ensures that the model output is consistent with actual energy consumption observations.

[0072] Specifically, the instantaneous power balance constraint is constructed by making the energy consumption change rate predicted by the model approximate the reference power determined based on the measured force signal, speed signal and equipment reference efficiency function.

[0073] The energy consumption change rate predicted by the model is not directly output, but is reconstructed through an intermediate representation within the global physical shared kernel.

[0074] Specifically, in the penultimate layer of the globally shared physical kernel, the model outputs a value of length [missing information]. Time series hidden states The sequence is aligned with the time axis of the stamping cycle.

[0075] Through a differentiable linear projection layer ,Will Mapped to predict instantaneous power : ; in, , For learnable parameters, For the hidden state dimension.

[0076] The reference power Calculated from the synchronously acquired raw signal: ; in, The real-time punching force of the slider (unit: N) is obtained by a high-precision force sensor sampling at 10kHz. The instantaneous velocity of the slider (unit: m / s) is obtained by measuring the displacement signal of the grating ruler. Perform central difference calculation: , ; The equipment's baseline efficiency function is the speed. The scalar function characterizes the electromechanical conversion efficiency of the drive system.

[0077] Instantaneous power balance constraint Defined as the mean square error between the predicted power and the reference power over the entire period: ; This constraint forces the model's internal dynamic behavior to conform to the principle of power transfer in mechanics: the input electrical power must be equal to the mechanical output power divided by the efficiency, thereby preventing the model from predicting unreasonable high energy consumption under low load and high speed conditions.

[0078] The periodic energy conservation constraint is constructed by making the model-predicted total energy consumption per cycle approximate the reference total energy consumption determined based on the measured deformation work and the inherent energy loss of the equipment.

[0079] The total energy consumption per cycle predicted by the model is the final output scalar of the model. The reference total energy consumption is denoted as... Its physical composition is the sum of useful work and unavoidable losses: ; in: This represents the energy consumed in the plastic forming of the material, as measured in actual deformation work. This represents the inherent energy loss of the equipment, indicating fixed energy consumption unrelated to forming. Measured deformation work. By analyzing the synchronously acquired force-displacement curves Within the effective stroke range Numerical integration is performed within the sample. Simpson's rule is used, and the number of sampling points is assumed to be even. Step length ,but: ; in , used to define the force values ​​at these nodes. The integer index represents the index of the first integer. _ sampling points, used in Simpson's integral formula, with a range of values ​​of _ . ,common One point. Indicates the first Displacement nodes The corresponding impact force value. Since the original data is a time series, it needs to be obtained through interpolation (such as linear or spline interpolation) from... Obtain the exact correspondence from of .

[0080] Periodic energy conservation constraint Defined as the model predicts total energy consumption Compared with reference total energy consumption Squared error: ; This constraint ensures that the model output satisfies the first law of thermodynamics: the single-cycle electrical energy input must be equal to the energy consumed by material deformation plus the inherent losses of the equipment, thus preventing energy predictions from appearing out of thin air or disappearing out of thin air, thereby improving the physical credibility and engineering interpretability of the model.

[0081] Specifically, the equipment reference efficiency function is obtained by performing efficiency-speed characteristic tests on the drive system of the corresponding stamping press under no-load and load conditions, and by performing function fitting on the discrete efficiency points obtained from the tests.

[0082] The equipment's baseline efficiency function is denoted as... It is defined in the slider speed A continuous scalar function on , used to characterize the . Electromechanical conversion efficiency of the stamping press drive system.

[0083] This function was obtained through offline calibration experiments. The specific steps are as follows: 1. Under no-load conditions (the mold is completely disengaged, and only the slider moves), the main motor is controlled by the frequency converter to run stably at six set points: 10%, 30%, 50%, 70%, 90%, and 100% of the rated speed.

[0084] 2. At each set point, the rotational speed fluctuation is less than And continue Seconds later, synchronously record: motor input power Measured by a 0.05-class power analyzer.

[0085] Output shaft mechanical power The torque was measured by a ±0.5%FS torque sensor. and angular velocity calculate.

[0086] 3. Calculate the no-load efficiency point: Corresponding speed .

[0087] 4. Repeat the above process to obtain another set of efficiency points under rated load conditions (using a standard load module to simulate typical forming forces). .

[0088] 5. Curve fitting was performed on both sets of data using cubic spline interpolation with endpoint smoothing constraints, yielding the results. and ; The final equipment baseline efficiency function is the arithmetic mean of the two: ; This function is embedded as prior knowledge into the instantaneous power balance constraint, enabling the model to reasonably infer energy consumption in unseen speed ranges and improve generalization ability.

[0089] The inherent energy loss of the equipment is obtained by conducting continuous no-load operation tests on the corresponding stamping press and statistically analyzing its average cycle energy consumption.

[0090] The inherent energy loss of the device is denoted as . The unit is joules (J), representing the fixed-cycle energy consumption of a stamping press when it is not performing forming tasks, due to friction, hydraulic leakage, cooling fans, control systems, etc. The method for obtaining this energy is as follows: 1. Place the stamping machine in a thermally stable state, with an ambient temperature of 20±5℃, and run it continuously for ≥30 minutes.

[0091] 2. The mold is completely detached, ensuring there is no forming load.

[0092] 3. Run continuously for at least 15 no-load cycles in standard single-cycle mode.

[0093] 4. Use a high-precision energy meter to record the actual energy consumption for each cycle. ,in .

[0094] 5. Calculate the sample mean with standard deviation .

[0095] 6. If the coefficient of variation If the data is stable, then let: Otherwise, extend the testing period until stability is achieved.

[0096] This value is embedded as a constant term in the periodic energy conservation constraint term, enabling the model to distinguish between useful work and unavoidable losses, and maintaining high accuracy in predictions even under low load or trial conditions.

[0097] S2: In the central coordination device, the model coordination aggregation step is executed: the gradient of the global physical shared core is received by each press in encrypted form.

[0098] The central coordination device refers to the central coordination device deployed at the workshop server layer. Its hardware platform is an industrial server with secure computing capabilities, running a federated learning coordination service program.

[0099] The model collaborative aggregation step is initiated periodically by the central collaborative device, and the aggregation period is set to every [number] days. Once per hour, or dynamically configured by the system administrator.

[0100] At the start of each aggregation round, the central coordinating unit broadcasts an aggregation start signal to all online presses; after the selected presses (usually 80% of the total number of online devices) respond, they perform local training and prepare to upload data.

[0101] The gradient of the global physical shared kernel is denoted as... superscript Indicates the first Taiwanese stamping machine The dimension of the globally shared physics kernel parameter vector (e.g.) ).

[0102] The gradient The parameters of the global physical shared kernel are calculated using the total loss function that includes the physical constraint loss term. The vector of partial derivatives obtained by backpropagation: ; To protect data privacy and model security, the edge processing device performs dual protection on the gradient before uploading: Differential privacy noise enhancement: Adding noise following a Gaussian distribution The noise, where the noise scale satisfy Differential privacy budget; Homomorphic encryption: Employs a public-key encryption algorithm that supports additive homomorphism (such as Paillier or CKKS), using the public key of the central coordinating device to encrypt the noisy gradient, generating ciphertext. The central coordination unit receives encrypted gradients from each stamping press via industrial Ethernet. The text is then decrypted in a secure environment to restore the plaintext gradient after adding noise, which is then used for subsequent aggregation.

[0103] The received gradients are weighted and aggregated based on the quality assessment results of the data uploaded by each stamping press to update the parameters of the global physical shared kernel, and the updated parameters are then sent to each stamping press.

[0104] The quality assessment result refers to the average anomaly score synchronously uploaded by each stamping machine after local training. and the number of valid samples Both were generated by a robust random cut forest anomaly detection model.

[0105] Based on this quality assessment information, the central coordination unit assigns each stamping press participating in the aggregation process a corresponding quality assessment. Calculate aggregate weights Its definition is: ; in: For the first The number of effective periodic feature vectors used for training in this round by the stamping press; This is the arithmetic mean of the outlier scores corresponding to these vectors; This is a preset anomaly penalty coefficient, with a value range of [0.1, 5.0], used to adjust the intensity of the impact of data quality on the weights.

[0106] Let the set of stamping presses participating in this round of aggregation be . (For example =4, corresponding to 4 out of 5 online devices), then the weighted aggregation gradient The calculation is as follows: ; This operation improves data quality (i.e.) Smaller stamping presses contribute more to global knowledge updates.

[0107] After aggregation is complete, the central coordinating device uses this gradient to update the parameters of the global physical shared kernel. The update rules are as follows: ; in This is the server-side learning rate. Updated parameters. It is encapsulated as a model parameter package and distributed to all stamping machines (including those not involved in this round of aggregation) via industrial Ethernet to ensure global knowledge synchronization.

[0108] The local energy consumption prediction model consists of the globally shared physical core and the local personalized adapter head connected in series.

[0109] The local energy consumption prediction model is a complete energy consumption prediction unit deployed in the edge processing device of each stamping machine, and its input is a single-cycle feature vector. The output is the predicted total energy consumption per cycle. .

[0110] This model adopts a serial connection architecture: firstly, ... Inputting a global physical shared kernel yields intermediate feature representations. Then Input a local personalized adapter, and the final output is... The mathematical expression is: ; in: A mapping function representing the globally shared physical kernel; Indicates the first Mapping functions for a local, customized adapter head specific to the stamping machine; These are parameters shared by all stamping machines; For only belonging to the first Private parameters of the stamping machine.

[0111] This architecture decouples general physical knowledge (encoded by a global kernel) from device-specific characteristics (encoded by a local header), supporting cross-device knowledge transfer while preserving the ability to model individual differences.

[0112] Specifically, the global physical shared kernel is a multi-layer feedforward neural network, whose parameters are jointly updated by all stamping presses participating in collaborative training.

[0113] The global physical shared kernel is a fully connected feedforward neural network with a depth of 4 layers. Its layer structure is as follows: input layer (12 dimensions) → hidden layer 1 (64 dimensions) → hidden layer 2 (32 dimensions) → hidden layer 3 (16 dimensions) → output layer (8 dimensions).

[0114] The activation function uses the Swish function: , in It is the Sigmoid function, which balances nonlinear expressiveness with gradient stability.

[0115] The parameter vector consists of all trainable parameters of the network (including the weight matrix and bias vector). ,in .

[0116] The initial values ​​are derived from the pre-training phase during the early stages of system deployment, obtained through joint training using historical normal data and physical constraints.

[0117] During the online operation phase, Instead of updating locally on any individual stamping press, gradients uploaded by each stamping press are collected through a central coordination device. After weighted aggregation, the data is updated uniformly, thereby realizing a collaborative evolution mechanism in which all stamping machines participating in collaborative training contribute and benefit together.

[0118] The local personalized adapter is a neural network, and its parameters are updated only locally using local data on the corresponding stamping press.

[0119] The local personalized adapter is a lightweight fully connected feedforward neural network with the following layer structure: input layer (8-dimensional) → hidden layer (4-dimensional) → output layer (1-dimensional).

[0120] The network receives the 8-dimensional output of the global physical shared core as input and finally regresses the predicted value of total energy consumption per cycle.

[0121] All its trainable parameters constitute a parameter vector. ,in .

[0122] Never leave the first The edge processing device of the stamping press uses training data entirely from the periodic feature vectors collected locally and the measured energy consumption. The training process is completed locally, gradients are not uploaded, and parameters are not shared.

[0123] This design ensures that the process characteristics of each stamping press (such as die wear, hydraulic system response delay, lubrication conditions, etc.) are accurately modeled, while preventing the leakage of sensitive production data and meeting the company's requirements for protecting trade secrets.

[0124] Based on the above-mentioned self-calibration method for energy consumption model of multi-machine stamping line under dynamic load, the present invention further proposes a self-calibration system for energy consumption model of multi-machine stamping line under dynamic load, including: a central coordination device and multiple edge processing devices.

[0125] The edge processing devices are deployed one-to-one with each stamping machine, and their functions are executed entirely locally without relying on the uploading of raw data.

[0126] The edge processing device first responds to the stamping cycle event, and synchronously collects the machine's mechanical signals (slider stamping force), motion signals (slider displacement), and electrical signals (motor input power) through high-precision sensors, and extracts a 12-dimensional multi-dimensional periodic feature vector based on these signals.

[0127] Subsequently, the edge processing device invokes the built-in unsupervised anomaly detection model to perform quality filtering on the multidimensional periodic feature vector, eliminating unreliable samples caused by abnormal operating conditions or sensor interference.

[0128] For the filtered periodic feature vectors, the edge processing device uses them to perform local training on a local energy consumption prediction model, which consists of a cascaded global physical shared kernel and a local personalized adapter head.

[0129] The local training specifically includes: taking the total loss function, which includes the physical constraint loss term, as the optimization objective, updating only the parameters of the local personalized adapter head, and simultaneously calculating the gradient of the global physical shared kernel.

[0130] The gradient of the calculated global physical shared kernel is uploaded to the central coordination device after differential privacy noise addition and homomorphic encryption. At the same time, the device receives the updated global parameters issued by the central coordination device to replace the locally stored global physical shared kernel parameters, thus completing the model synchronization.

[0131] The central coordination device is deployed at the workshop server layer and is responsible for coordinating the collaborative learning process of the entire system.

[0132] It first receives encrypted gradient data uploaded from each edge processing device.

[0133] Next, based on the quality assessment information attached to each device, namely the periodic feature vector anomaly score calculated by the data quality filtering module, the gradient is weighted and securely aggregated.

[0134] Specifically, the security aggregation module dynamically adjusts the weight of the gradient in the aggregation based on the average anomaly score of each batch uploaded by the edge processing device: the lower the anomaly score, the higher the weight, thereby ensuring that high-quality data plays a leading role in the global model update.

[0135] After aggregation is complete, the central coordinating device updates the parameters of the global physical shared core and securely distributes the new parameters to all edge processing devices.

[0136] Specifically, the physical constraint loss term includes two items: The first is the instantaneous power balance constraint term, which is used to force the energy consumption change rate predicted within the model to be consistent with the reference instantaneous power calculated based on the measured force signal, slider speed signal and equipment reference efficiency function.

[0137] Second, there is the periodic energy conservation constraint, which forces the model to predict that the total energy consumption per cycle is equal to the sum of the measured deformation work and the inherent energy loss of the equipment.

[0138] The unsupervised anomaly detection model is specifically implemented as an online-updated robust random cut forest model (RRCF), which supports streaming data insertion and real-time anomaly scoring.

[0139] In one specific embodiment, an automobile manufacturing company has a stamping production line consisting of five large mechanical presses (numbered P1 to P5) for producing automotive body panels such as doors and hoods. To achieve refined energy management, optimized scheduling, and predictive maintenance, the company has deployed the self-calibration system for the energy consumption model of a multi-machine stamping line under dynamic load, as described in this invention.

[0140] 1. System deployment and initialization.

[0141] Each press (taking P1 as an example) is locally deployed with an industrial-grade edge computing box as an edge processing device E1, which integrates a high-precision data acquisition card, an industrial Ethernet interface supporting the IEEE 1588v2 synchronization protocol, and an embedded AI acceleration module. An industrial server is deployed in the workshop control room as a central coordination device, responsible for federated coordination.

[0142] E1 connects to three types of sensors: A high-precision force sensor (range 2000kN, accuracy 0.5%FS) is mounted on a slider to acquire mechanical signals. ; A grating ruler (1μm resolution) is mounted on the machine body and used to acquire motion signals. ; The power measurement unit (0.5 accuracy) is installed on the main motor input side to collect electrical signals. ; The incremental encoder Z-phase signal at the main motor shaft end serves as a global hardware trigger source, and a time-sensitive network (TSN) ensures strict time alignment of the three types of signals at a 10kHz sampling rate.

[0143] Before the system is officially launched, perform offline calibration on P1: Equipment baseline efficiency function calibration: Under no-load and rated load conditions, the main motor is controlled to operate stably at six fixed speed points, and the input electrical power and output mechanical power are measured simultaneously to calculate the efficiency at each point. A cubic spline interpolation method with endpoint smoothing constraints is used for fitting to obtain the equipment baseline efficiency function specific to P1. This function reflects the nonlinear characteristics of the drive system efficiency as a function of slider speed.

[0144] Equipment inherent energy loss calibration: With the mold completely detached and the engine in a stable thermal state, P1 was run continuously under no-load for 20 cycles, and the measured energy consumption of each cycle was recorded; the calculated average value was 7500J, and the coefficient of variation was 0.8% < 1%, therefore the inherent energy loss of P1 was determined. .

[0145] Model initialization: The central coordinating device pre-trains an initial global physical shared kernel (4-layer fully connected neural network) using historical normal production data and distributes it to E1; E1 simultaneously initializes a local personalized adapter head (2-layer lightweight network) dedicated to P1, forming a complete local energy consumption prediction model.

[0146] 2. Online self-calibration process during normal operation.

[0147] Currently, P1 is stamping the inner panel of the left front door (mold A).

[0148] (a) Single-cycle data processing.

[0149] At the start of the stamping cycle, the encoder's Z phase triggers synchronous acquisition, and E1 acquires a 6-second sequence. , , sequence.

[0150] Within 50 ms after the period ends, E1 extracts a 12-dimensional periodic feature vector. ,Include: , , , , , , , , , , , .

[0151] (b) Quality filtration.

[0152] E1 will Input an online-updated Robust Random Cut Forest (RRCF) model to compute anomaly scores. .

[0153] Dynamic threshold The 99.5th percentile of the most recent 1000 outlier scores ( ).

[0154] because If the data is deemed valid, it is stored in the local training buffer (capacity 300). When the buffer is full, local training is triggered.

[0155] (c) Local training.

[0156] E1 samples 32 data points from the buffer and performs local training: Forward propagation: Input The local energy consumption prediction model (global physical shared core + local personalized adapter) outputs predicted energy consumption. And reconstruct the predicted instantaneous power sequence .

[0157] Loss calculation: Data fitting loss: compared with measured energy consumption Calculate the mean square error; Instantaneous power balance constraint: Calculate the reference power ,constraint ; Periodic energy conservation constraint: Calculate the reference total energy consumption ,constraint .

[0158] Parameter update: Freeze the global physical shared kernel parameters and update only the local personalized adapter header parameters; at the same time, calculate the gradient of the global physical shared kernel. .

[0159] E1 calculates the average anomaly score for this batch. After adding noise and encrypting the gradient, along with the metadata Uploaded to the central coordination device.

[0160] 3. Handle sudden changes in working conditions and replace molds.

[0161] When the MES issues a mold change command (switch to the outer hood panel, mold B): E1 detected the triggering event via the OPC UA protocol; Immediately clear the current local training buffer (originally containing template A data); The new cycle data is marked as the starting point of the new operating condition and stored in the buffer after quality filtering. Subsequently, due to buffer accumulation or prediction error exceeding the limit, a new round of local training will be triggered to quickly adapt to the load characteristics of mold B.

[0162] 4. Central coordination and knowledge sharing.

[0163] The central coordination device initiates a convergence round every 4 hours. Assume P1, P3, and P4 participate in this round: Receive encryption gradients and metadata: P1: ; P3: ; P4: .

[0164] Weighted safe aggregation: Decrypting the gradient; Calculate the weights: ,Pick ; ; Aggregate gradient: ; Update global physical shared kernel parameters: ; The new parameters were sent to all five edge processing devices.

[0165] 5. System self-calibration effect.

[0166] Individual Adaptability: P1 continuously tracks its own state drift (such as lubrication deterioration) through a local personalized adapter. When the prediction error EWMA exceeds 2% for 20 consecutive cycles, it automatically triggers local training to achieve self-healing.

[0167] Knowledge generalization: Global physics sharing and multi-machine fusion experience, with the ability to understand universal physical laws, so that when a new model is launched, only the local head needs to be fine-tuned to achieve rapid convergence.

[0168] Physical reliability: In high-speed, low-load model testing, the instantaneous power balance constraint term forces the model to conform to... This avoids unreasonable high energy consumption predictions caused by data extrapolation.

[0169] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-calibration method for the energy consumption model of a multi-machine stamping line under dynamic load, characterized in that, include: At each stamping press, the local model training step is performed: based on the synchronously acquired stamping process signals of the press, multi-dimensional periodic feature vectors are extracted; The periodic feature vector is quality filtered by unsupervised anomaly detection and dynamic threshold comparison. Using filtered periodic feature vectors, a local personalized adapter head for a local energy consumption prediction model is trained, and the gradient of the global physical shared kernel is calculated. The total loss function for model training includes a physical constraint loss term, which includes an instantaneous power balance constraint term and a periodic energy conservation constraint term. In the central coordination device, the model coordination aggregation step is executed: the gradient of the global physical shared kernel is received by each press in encrypted form; The received gradients are weighted and aggregated based on the quality assessment results of the data uploaded by each stamping press to update the parameters of the global physical shared kernel, and the updated parameters are sent to each stamping press. The local energy consumption prediction model consists of the globally shared physical core and the local personalized adapter head connected in series.

2. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 1, characterized in that, The instantaneous power balance constraint term is constructed by making the energy consumption change rate predicted by the model approximate the reference power determined based on the measured force signal, velocity signal and equipment reference efficiency function; The periodic energy conservation constraint is constructed by making the model-predicted total energy consumption per cycle approximate the reference total energy consumption determined based on the measured deformation work and the inherent energy loss of the equipment.

3. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 2, characterized in that, The equipment reference efficiency function is obtained by performing efficiency-speed characteristic tests on the drive system of the corresponding stamping press under no-load and load conditions, and by fitting the discrete efficiency points obtained from the tests to a function. The inherent energy loss of the equipment is obtained by conducting continuous no-load operation tests on the corresponding stamping press and statistically analyzing its average cycle energy consumption.

4. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 1, characterized in that, The periodic feature vector is input into an online-updated robust random-cut forest anomaly detection model to calculate the anomaly score; The abnormal score is compared with a dynamic threshold, which is adaptively determined based on the statistical distribution characteristics of the abnormal scores within a historical time period.

5. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 4, characterized in that, Based on the average anomaly score corresponding to the batch data uploaded by each stamping machine, the weight of the gradient contributed by that batch data in the aggregation calculation is dynamically adjusted so that the batch with the lower anomaly score contributes more to the global model update.

6. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 1, characterized in that, The local training step of the model is executed by being triggered by an event or condition; Triggering events include receiving a mold change command; Triggering conditions include the model prediction error continuously exceeding the allowable range for a predetermined time, or the number of valid periodic feature vectors accumulated in the local training buffer reaching a preset threshold.

7. The self-calibration method for energy consumption model of multi-machine stamping line under dynamic load according to claim 1, characterized in that, The global physical shared kernel is a multi-layer feedforward neural network, whose parameters are updated jointly by all stamping presses participating in collaborative training; The local personalized adapter is a neural network, and its parameters are updated only locally using local data on the corresponding stamping press.

8. A self-calibration system for energy consumption model of a multi-machine stamping line under dynamic load, characterized in that, include: A central coordinating device and multiple edge processing devices communicatively connected to the central coordinating device; The edge processing devices are installed one-to-one with each stamping machine and are used for: In response to stamping cycle events, the machine's mechanical signals, motion signals, and electrical signals are simultaneously acquired and extracted into multi-dimensional periodic feature vectors. The multidimensional periodic feature vector is quality filtered using an unsupervised anomaly detection model. A local energy consumption prediction model is trained locally using filtered periodic feature vectors. The local energy consumption prediction model consists of a series of global physical shared kernels and a local personalized adapter head. The local training includes: updating the parameters of the local personalized adapter head with a total loss function that includes a physical constraint loss term, and calculating the gradient of the global physical shared kernel; The calculated gradient of the global physical shared core is encrypted and uploaded to the central collaborative device; And receive global parameters issued by the central coordination device, and update the local global physical shared core; The central coordination device is used for: Receive the encrypted gradient uploaded by each edge processing device; Based on the quality assessment information of the data uploaded by each edge processing device, the received gradients are weighted and securely aggregated to update the parameters of the global physical shared kernel; The updated global parameters are then distributed to each edge processing device.

9. A self-calibration system for energy consumption model of a multi-machine stamping line under dynamic load as described in claim 8, characterized in that, The physical constraint loss term includes: an instantaneous power balance constraint term, which is used to constrain the energy consumption change rate predicted by the model to be consistent with the reference instantaneous power calculated based on the measured force signal, the slider speed signal and the equipment reference efficiency function; In addition, there is a periodic energy conservation constraint term, which is used to constrain the model prediction of the total energy consumption per cycle, which is consistent with the sum of the measured deformation work and the inherent energy loss of the equipment.

10. A self-calibration system for energy consumption model of a multi-machine stamping line under dynamic load as described in claim 8, characterized in that, The unsupervised anomaly detection model is an online-updated robust random cutting forest model; The quality assessment information is the periodic feature vector anomaly score calculated by the data quality filtering module. The security aggregation module is configured to dynamically adjust the weight of its gradient in the weighted aggregation based on the average anomaly score corresponding to the data uploaded by each edge processing device.

Images