Kinetic recovery optimization method, system, and medium for an automated production line

Abstract

The application provides a kinetic energy recovery optimization method, system and medium for an automated production line, relates to the technical field of automation production, and comprises the following steps: performing multi-source working condition sensing to construct a plurality of energy fingerprint characteristic vectors; dynamically dividing a plurality of kinetic energy recovery nodes into K homogenized recovery groups; performing main node election to obtain K group representative fingerprints and match K group exclusive optimization base models; combining real-time working condition data flow to perform group internal small sample meta-learning adaptive parameter adjustment and construct K group node level optimization models; and tracking working condition deviation states during implementation of the adaptive kinetic energy recovery parameter adjustment optimization process and performing model smoothing parameter adjustment. The application solves the technical problem in the prior art that kinetic energy recovery control modes mostly adopt fixed parameter control strategies or unified model control modes, different kinetic energy recovery nodes are difficult to realize targeted control under complex working conditions, and thus part of the nodes have low energy recovery efficiency and poor output stability.

Description

Technical Field

[0001] This invention relates to the field of automated production technology, and more specifically to a method, system, and medium for optimizing kinetic energy recovery in automated production lines. Background Technology

[0002] In existing technologies, piezoelectric energy recovery devices, electromagnetic energy recovery devices, or energy storage and regulation modules are typically used to recover mechanical vibration and impact energy in production lines, thereby reducing equipment operating energy consumption and improving energy utilization efficiency. However, due to significant differences in operating load, vibration frequency, installation structure, and working environment among different production line equipment, the energy recovery state between various kinetic energy recovery nodes exhibits strong dynamism and dispersion. Therefore, how to dynamically optimize the kinetic energy recovery process based on actual operating condition changes has gradually become an important research direction in the field of automated production lines.

[0003] However, most existing kinetic energy recovery control methods adopt fixed parameter control strategies or unified model control methods. They usually adjust energy recovery based on a single vibration parameter or static empirical parameter, which makes it difficult to achieve targeted control of different kinetic energy recovery nodes under complex working conditions. This results in some nodes having problems such as low energy recovery efficiency, poor output stability, and high energy loss. Summary of the Invention

[0004] This application provides a method, system, and medium for optimizing kinetic energy recovery in automated production lines. It aims to solve the technical problem that most existing kinetic energy recovery control methods adopt fixed parameter control strategies or unified model control methods, making it difficult to achieve targeted control of different kinetic energy recovery nodes under complex working conditions, which leads to low energy recovery efficiency and poor output stability of some nodes.

[0005] The first aspect disclosed in this application provides a method for optimizing kinetic energy recovery in an automated production line. The method includes: performing multi-source operating condition sensing during a baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in the automated production line to construct multiple energy fingerprint feature vectors; dynamically dividing the multiple kinetic energy recovery nodes into K homogeneous recovery groups based on the energy similarity of the multiple energy fingerprint feature vectors; electing a master node for each of the K homogeneous recovery groups to obtain K group representative fingerprints, and matching K group-specific optimization base models in a pre-trained meta-model library; combining the real-time operating condition data streams of the K homogeneous recovery groups to perform intra-group small-sample meta-learning adaptation and parameter tuning on the K group-specific optimization base models to construct K group node-level optimization models; and driving K edge execution modules to implement adaptive kinetic energy recovery parameter tuning optimization for the K homogeneous recovery groups using the K group node-level optimization models, while tracking the operating condition offset state of the K homogeneous recovery groups and performing model smoothing parameter tuning on the K group node-level optimization models.

[0006] The second aspect of this application discloses a kinetic energy recovery optimization system for an automated production line. The system is used in the aforementioned kinetic energy recovery optimization method for an automated production line. The system includes: a multi-source operating condition sensing unit, used to perform multi-source operating condition sensing during a baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in the automated production line to construct multiple energy fingerprint feature vectors; a dynamic partitioning unit, used to dynamically partition the multiple kinetic energy recovery nodes into K homogeneous recovery groups based on the energy similarity of the multiple energy fingerprint feature vectors; and a master node election unit, used to elect a master node from the K homogeneous recovery groups. Node election yields K group representative fingerprints, which are then matched with K group-specific optimized base models in a pre-trained meta-model library. An adaptation and parameter tuning unit combines the real-time operating data streams of the K homogeneous recycling groups to perform intra-group small-sample meta-learning adaptation and parameter tuning of the K group-specific optimized base models, constructing K group-level optimized models. A model smoothing and parameter tuning unit drives K edge execution modules to implement adaptive kinetic energy recovery parameter tuning optimization for the K homogeneous recycling groups using the K group-level optimized models, tracking the operating condition offset states of the K homogeneous recycling groups and performing model smoothing and parameter tuning of the K group-level optimized models.

[0007] The third aspect disclosed in this application provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the energy recovery optimization method for automated production lines in the first aspect.

[0008] One or more technical solutions provided in this application have at least the following beneficial effects: By performing multi-source operating condition sensing at multiple kinetic energy recovery nodes in an automated production line and constructing corresponding energy fingerprint feature vectors, an accurate data foundation is provided for subsequent group optimization and model adaptation. Dynamically homogenizing and grouping multiple kinetic energy recovery nodes based on the energy similarity of multiple energy fingerprint feature vectors improves control consistency and optimization targeting within groups, avoiding the insufficient adaptability and energy recovery efficiency fluctuations inherent in traditional globally unified optimization methods. By electing a master node for each homogenized recovery group and matching the corresponding group-specific optimization base model from a pre-trained meta-model library based on the group representative fingerprint, rapid initialization and targeted construction of the optimization model are achieved, improving model deployment efficiency and optimization under complex operating conditions. The system exhibits enhanced responsiveness. By combining real-time operating data streams from various homogeneous energy recovery groups, it performs small-sample meta-learning and parameter tuning on the group-specific optimized base model. This allows for rapid and personalized adjustments to the model based on the real-time operating status of each energy recovery node, thereby improving the control accuracy and energy conversion efficiency of energy recovery and reducing control mismatch issues caused by changes in operating conditions. Furthermore, by driving the edge execution module to implement adaptive energy recovery parameter tuning control using a node-level optimized model and tracking the operating condition offset in real time, the system performs smooth parameter tuning on the node-level optimized model. This enables progressive updates of model parameters during continuous changes in operating conditions, thereby improving the operational stability, adaptive control capability, and long-term operational reliability of the energy recovery system.

[0009] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of a kinetic energy recovery optimization method for automated production lines provided in an embodiment of this application.

[0011] Figure 2 This is a schematic diagram of the kinetic energy recovery optimization system for an automated production line, provided as an embodiment of this application.

[0012] Figure labeling: Multi-source working condition sensing unit 10, dynamic partitioning unit 20, master node election unit 30, adaptation parameter tuning unit 40, model smoothing parameter tuning unit 50. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, a method for optimizing kinetic energy recovery in automated production lines is provided, the method comprising: In the process of benchmark kinetic energy recovery at multiple kinetic energy recovery nodes in an automated production line, multi-source operating condition sensing is performed to construct multiple energy fingerprint feature vectors.

[0015] During the baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in an automated production line, mechanical vibration parameters, electrical output parameters, load variation parameters, and environmental condition parameters are collected for each node under operating conditions. Mechanical vibration parameters include vibration frequency, vibration amplitude, and impact period; electrical output parameters include output voltage, output current, and energy storage change; load variation parameters include equipment operating cycle time and load fluctuation amplitude; and environmental condition parameters include temperature, humidity, and equipment installation location. The collected multi-source operating condition data undergoes time synchronization, anomaly removal, and normalization processing, and is then fused and encoded according to preset feature combination rules to construct an energy fingerprint feature vector corresponding to each kinetic energy recovery node, characterizing the comprehensive energy recovery characteristics of each node under the current operating conditions.

[0016] Based on the energy similarity of the multiple energy fingerprint feature vectors, the multiple kinetic energy recovery nodes are dynamically divided into K homogeneous recovery groups.

[0017] Multiple energy fingerprint feature vectors are mapped to a preset high-dimensional feature space, and the feature distance between nodes is calculated based on the similarity between energy recovery efficiency, vibration response characteristics, and load fluctuation characteristics. Density clustering is performed on multiple feature particles according to a preset energy similarity diameter, aggregating kinetic energy recovery nodes with similar energy characteristics into corresponding homogeneous particle groups. Then, based on the clustering results, multiple kinetic energy recovery nodes in the automated production line are dynamically grouped and mapped to generate K homogeneous recovery groups. After grouping, the local control permissions of each homogeneous recovery group are bound to the corresponding edge execution module to facilitate subsequent group-level collaborative optimization control.

[0018] Master node election is performed on the K homogeneous recycling groups to obtain K group representative fingerprints, and K group-specific optimized base models are matched in the pre-trained meta-model library.

[0019] Energy efficiency indicators (EEIs) of kinetic energy recovery nodes within each homogeneous recovery group are statistically analyzed. These EEIs include mechanical vibration capture rate, piezoelectric conversion efficiency, circuit conversion loss rate, and node operation failure rate. A comprehensive score is calculated based on a preset dynamic weighting strategy. The kinetic energy recovery node with the highest score in each homogeneous recovery group is selected as the group representative recovery node, and the energy condition characteristics corresponding to this group representative recovery node are extracted as the group representative fingerprint. The group representative fingerprint is input into a pre-trained meta-model library, and matching is performed at the energy mode layer, operating condition label layer, and topology location layer to obtain the corresponding basic physical optimization model, electromechanical coupling optimization model, and structural resonance suppression model. The parameters of multiple models are superimposed and fused to generate a dedicated optimization base model for the corresponding homogeneous recovery group.

[0020] By combining the real-time operating data streams of the K homogeneous recycling groups, the group-specific optimization base models are adapted and tuned using small-sample meta-learning within each group, thus constructing a K-group node-level optimization model.

[0021] The system continuously receives real-time operating condition data streams from multiple kinetic energy recovery nodes within each homogeneous recovery group, extracts corresponding real-time operating condition features, and compares them with the energy operating condition features in the group representative fingerprint to obtain the operating condition feature offset of each node. Based on the operating condition feature offset, a progressive matching process is performed in the dynamic compensation sub-model library to select the corresponding compensation sub-model parameters, which are then injected into the basic control framework of the group-specific optimization base model. Historical operating condition data of the corresponding nodes is retrieved as a small-sample training set. With maximizing piezoelectric conversion efficiency as the optimization objective, the model parameters are fine-tuned through gradient backpropagation to generate a node-level optimization model adapted to the current operating conditions of each kinetic energy recovery node.

[0022] During the process of adaptive kinetic energy recovery parameter tuning optimization of the K homogeneous recycling groups using the K sets of node-level optimization models, the operating condition offset status of the K homogeneous recycling groups is tracked, and the K sets of node-level optimization models are smoothly tuned.

[0023] The edge execution module calls the corresponding node-level optimization model to perform adaptive kinetic energy recovery control on the kinetic energy recovery nodes in the homogeneous recovery group, and outputs the corresponding electrical control parameters in real time to dynamically adjust the energy storage circuit, voltage regulation strategy, and energy recovery cycle. During operation, the operating condition offset status of each kinetic energy recovery node is continuously tracked. When the increment of the operating condition offset exceeds the preset dynamic offset threshold, a sliding window is constructed to collect the recent incremental feature vector, and constrained gradient descent is performed on the current node-level optimization model based on the incremental feature vector to extract the corresponding parameter increment vector. The parameter increment vector is then updated to the current node-level optimization model in a smooth fusion manner to obtain the updated node-level parameter tuning model, which replaces the original node-level optimization model to continue executing subsequent adaptive kinetic energy recovery control.

[0024] Furthermore, based on the energy similarity of the multiple energy fingerprint feature vectors, the multiple kinetic energy recovery nodes are dynamically divided into K homogeneous recovery groups. The method includes: Based on the characteristic dimensions of energy fingerprints, a high-dimensional feature space is constructed; the multiple energy fingerprint feature vectors are mapped to the high-dimensional feature space to form multiple feature particles; based on a preset energy similarity diameter, density clustering of the multiple feature particles is performed in the feature space to obtain K homogeneous particle groups; based on the K homogeneous particle groups, dynamic grouping mapping of the multiple kinetic energy recovery nodes is performed to generate the K homogeneous recovery groups; the local control permissions of the K homogeneous recovery groups are revoked and bound to the K edge execution modules to complete the allocation of local control permissions.

[0025] A corresponding high-dimensional feature space is established based on the characteristic dimensions of the energy fingerprint. These dimensions include vibration frequency characteristics, energy output characteristics, load fluctuation characteristics, operational stability characteristics, and environmental condition characteristics. Each feature dimension is normalized to a unified dimension, and corresponding feature axes are assigned according to preset feature coordinate rules to construct a high-dimensional feature space characterizing the operational characteristics of multiple kinetic energy recovery nodes. This allows for differentiated expression of the energy recovery status of different nodes within a unified feature space.

[0026] Multiple energy fingerprint feature vectors are input into the high-dimensional feature space, and spatial coordinate mapping is completed based on the feature values ​​corresponding to each feature dimension, so that each kinetic energy recovery node forms a feature particle. The positions of different feature particles in the high-dimensional feature space are used to represent the comprehensive operating characteristics of the corresponding kinetic energy recovery node in terms of energy recovery efficiency, vibration response, and load changes, thereby forming multiple feature particle sets that can be used for cluster analysis.

[0027] Based on a preset energy similarity diameter, density clustering analysis is performed on multiple feature particles in a high-dimensional feature space. First, the spatial distance and local neighborhood density between each feature particle are calculated, and feature particles whose distance is less than the preset energy similarity diameter are grouped into the same density neighborhood. Multiple feature particles with continuous density and similar feature distributions are aggregated to form corresponding homogeneous particle groups, thereby achieving automatic classification of the operational characteristics of multiple kinetic energy recovery nodes into similar groups, ultimately resulting in K homogeneous particle groups.

[0028] Based on the correspondence between the K homogeneous particle groups and kinetic energy recovery nodes, dynamic grouping mapping is performed on multiple kinetic energy recovery nodes in the automated production line, assigning kinetic energy recovery nodes belonging to the same homogeneous particle group to the same homogeneous recovery group. Combined with real-time operating condition changes, the group affiliation status of each kinetic energy recovery node is dynamically adjusted to ensure high consistency in energy recovery characteristics and operating conditions among nodes within the homogeneous recovery group.

[0029] After the dynamic division of the K homogeneous energy recovery groups is completed, the local control permissions of the corresponding kinetic energy recovery nodes in each homogeneous energy recovery group are uniformly reclaimed, and the control permissions of each homogeneous energy recovery group are bound to the corresponding edge execution modules. The edge execution modules, according to the unified optimization strategy of their respective homogeneous energy recovery groups, perform coordinated parameter control and energy recovery adjustment on multiple kinetic energy recovery nodes within the group, thereby completing the allocation of local control permissions and the deployment of group-level coordinated control.

[0030] Furthermore, a master node election is performed on the K homogeneous recycling groups to obtain K group representative fingerprints, and K group-specific optimized base models are matched in the pre-trained meta-model library. The method includes: Energy recovery process indicators are extracted from W kinetic energy recovery nodes in the first homogeneous recovery group to obtain W node energy efficiency feature vectors. The node energy efficiency feature vectors include mechanical vibration capture rate, piezoelectric conversion efficiency, circuit conversion loss rate, and node operation failure rate. Based on a preset dynamic weighting strategy, a comprehensive score is calculated for the W node energy efficiency feature vectors. After obtaining the comprehensive score of W nodes, the highest score is located in descending order to determine the first group of representative recovery nodes. The energy condition features of the first group of representative recovery nodes are used as the first group of representative fingerprints and matched with the first group of dedicated optimization base models in the pre-trained meta-model library.

[0031] Energy recovery process indicators were extracted for W kinetic energy recovery nodes in the first homogeneous recovery group. During the benchmark kinetic energy recovery process at each node, mechanical vibration response data, piezoelectric energy conversion data, circuit energy storage data, and operating status data were collected. Specifically, the mechanical vibration capture rate was calculated based on the correspondence between the mechanical vibration input energy and the actual captured energy; the piezoelectric conversion efficiency was calculated based on the conversion relationship between the piezoelectric output energy and the input mechanical energy; the circuit conversion loss rate was calculated based on the energy attenuation during circuit transmission; and the node failure rate was calculated by combining the number of abnormal shutdowns, the number of fault alarms, and the stable operating cycle. These indicators were normalized and feature fused to generate node energy efficiency feature vectors for the corresponding W kinetic energy recovery nodes.

[0032] Based on a preset dynamic weighting strategy, a comprehensive score is calculated for the energy efficiency feature vectors of W nodes. Positive weights are assigned to mechanical vibration capture rate and piezoelectric conversion efficiency, while negative weights are assigned to circuit conversion loss rate and node failure rate. The weight parameters for each indicator are dynamically adjusted based on the current production line's operating load status. The comprehensive score for each of the W nodes is obtained from the weighted calculation results, and these scores are sorted in descending order. The kinetic energy recovery node with the highest comprehensive score is identified as the representative node of the first group, serving as a typical representative node for the first homogeneous recovery group.

[0033] The energy condition characteristics of the first group of representative recycling nodes during operation are extracted as the first group of representative fingerprints. These energy condition characteristics include vibration frequency distribution characteristics, piezoelectric output variation characteristics, energy storage response characteristics, and load fluctuation characteristics. The first group of representative fingerprints is input into a pre-trained meta-model library. Energy pattern matching, operating condition label matching, and topology location matching are performed in the meta-model library to select the set of optimized models that are closest to the current operating conditions of the first homogeneous recycling group. The corresponding model parameters are then combined and fused to generate the first set of dedicated optimized base models suitable for the first homogeneous recycling group.

[0034] Furthermore, the method involves using the energy condition features of the first set of representative recycling nodes as the first set of representative fingerprints, and matching them with the first set of dedicated optimized base models in the pre-trained meta-model library. The first set of representative fingerprints is used to perform energy pattern matching, operating condition label matching, and topological position matching in the energy pattern layer, operating condition label layer, and topological position layer of the pre-trained meta-model library, respectively, to obtain the basic physical optimization model, electromechanical coupling optimization model, and structural resonance suppression model. With the basic physical optimization model as the core framework, the parameters of the electromechanical coupling optimization model and the structural resonance suppression model are superimposed to complete the dynamic generation of the first set of dedicated optimization base models.

[0035] The first set of representative fingerprints is input into the pre-trained meta-model library, and hierarchical matching is performed at the energy mode layer, operating condition label layer, and topology location layer. Specifically, in the energy mode layer, based on the vibration frequency characteristics, energy output characteristics, and energy storage change characteristics in the first set of representative fingerprints, a basic physical optimization model corresponding to the current energy recovery mode is matched. In the operating condition label layer, based on the load change state, equipment operating cycle time, and environmental operating condition characteristics, an electromechanical coupling optimization model suitable for the current operating state is matched. In the topology location layer, based on the installation position of the kinetic energy recovery node in the automated production line, structural connection relationships, and vibration propagation path, a corresponding structural resonance suppression model is matched. After completing the hierarchical matching, the basic physical optimization model, electromechanical coupling optimization model, and structural resonance suppression model are obtained.

[0036] Using the aforementioned fundamental physical optimization model as the overall control framework, the dynamic coupling parameters in the electromechanical coupling optimization model and the vibration suppression parameters in the structural resonance suppression model are superimposed and fused. Specifically, the load compensation parameters, output adjustment parameters, and response correction parameters from the electromechanical coupling optimization model are injected into the fundamental physical optimization model, while the resonance attenuation parameters, frequency avoidance parameters, and vibration smoothing parameters from the structural resonance suppression model are superimposed into the fundamental control process. The fused multi-model parameters are uniformly coordinated and constrained to ensure that the parameters meet the requirements of energy recovery stability and control real-time performance, thereby dynamically generating a first set of dedicated optimization base models suitable for the first homogeneous recovery group.

[0037] Furthermore, by combining the real-time operating data streams of the K homogeneous recycling groups, the K group-specific optimization base models are subjected to in-group small-sample meta-learning adaptation and parameter tuning to construct K group node-level optimization models. The method includes: Extract the basic control framework of the first set of dedicated optimization base models; extract real-time operating condition data stream features of the W kinetic energy recovery nodes in the first homogeneous recovery group to generate W node real-time operating condition features, compare them with the energy operating condition features of the first set of representative recovery nodes, and calculate the W operating condition feature offsets; perform progressive matching and selection in a dynamic compensation sub-model library built based on a preset compensation strategy according to the W operating condition feature offsets to obtain W compensation sub-model parameters; after injecting the W compensation sub-model parameters into the basic control framework, perform small sample learning gradient fine-tuning to construct W node-level optimization models, forming the first set of node-level optimization models.

[0038] The corresponding basic control framework is extracted from the first set of dedicated optimization base models. The basic control framework is preferably a kinetic energy recovery control model constructed based on a model predictive control algorithm and a feedforward / feedback joint control algorithm. The basic control framework includes a vibration input prediction layer, an energy conversion regulation layer, an energy storage control layer, and a feedback constraint layer. Specifically, the vibration input prediction layer is constructed using a time-series operating condition prediction model to predict the current mechanical kinetic energy input state based on the historical vibration frequency, vibration amplitude, and load change trend of the kinetic energy recovery node; the energy conversion regulation layer is constructed using a piezoelectric impedance matching control model to dynamically adjust the equivalent impedance parameters, voltage regulation parameters, and rectification control parameters in the piezoelectric conversion circuit based on the predicted mechanical kinetic energy input state; the energy storage control layer is constructed using an energy storage power allocation model to adjust the energy storage charging and discharging parameters based on the current remaining capacity of the energy storage unit, output load demand, and instantaneous recovered power; and the feedback constraint layer is constructed using a closed-loop feedback constraint model to dynamically constrain and correct the overall model parameters based on output voltage stability, circuit temperature rise, and recovery efficiency deviation.

[0039] Furthermore, the construction process of the basic control framework includes: firstly, collecting historical mechanical vibration data, electrical output data, energy storage status data, and equipment load data of multiple kinetic energy recovery nodes under different operating conditions, and performing time synchronization, anomaly removal, and normalization on the historical data; using vibration frequency, vibration amplitude, collision period, piezoelectric output voltage, output current, circuit temperature rise rate, and load change rate as model input features, and using piezoelectric conversion efficiency, output stability, and energy storage recovery efficiency as model optimization objectives, constructing a basic training sample set; and using a gradient-based iterative model parameter training method to jointly train the control parameters in the vibration input prediction layer, energy conversion regulation layer, energy storage control layer, and feedback constraint layer to generate the basic control framework.

[0040] Furthermore, the inputs of the basic control framework include a real-time operating condition feature vector, which includes vibration frequency variation characteristics, peak collision pressure, piezoelectric output fluctuation rate, energy storage voltage change rate, circuit temperature rise rate, and equipment load change rate. The outputs of the basic control framework include a set of electrical control parameters, which includes piezoelectric impedance adjustment parameters, rectification control parameters, energy storage charging and discharging parameters, and output voltage adjustment parameters. During the operation of the basic control framework, the vibration input prediction layer first predicts the current mechanical kinetic energy input trend, then the energy conversion adjustment layer generates corresponding piezoelectric adjustment parameters based on the prediction results, then the energy storage control layer generates corresponding energy storage control parameters based on the current energy storage state, and finally the feedback constraint layer constrains and corrects the above control parameters based on the real-time output state, and outputs the final set of electrical control parameters.

[0041] Real-time operating condition data streams are continuously collected from W kinetic energy recovery nodes in the first homogeneous recovery group, and corresponding real-time operating condition features are extracted. These features include vibration frequency variation characteristics, piezoelectric output fluctuation characteristics, load variation characteristics, and circuit temperature rise characteristics. The real-time operating condition features of each node are compared with the energy operating condition features corresponding to the representative recovery nodes in the first group, and the degree of difference in the corresponding features is calculated to generate W operating condition feature offsets. These offsets characterize the degree of deviation of the current operating state of each kinetic energy recovery node from the representative operating condition of the group.

[0042] Based on the W operating condition feature offsets, a progressive matching selection is performed in a dynamic compensation sub-model library constructed based on a preset compensation strategy. Preferably, the dynamic compensation sub-model library is a set of incremental compensation models constructed based on operating condition offset classification. According to the offset direction, offset magnitude, and rate of change of the operating condition feature offsets, multiple compensation sub-models in the dynamic compensation sub-model library are filtered level by level, and corresponding dynamic compensation parameters are matched. The W compensation sub-model parameters corresponding to the current operating state of each kinetic energy recovery node are output for personalized adaptation of subsequent node-level optimization models.

[0043] The parameters of the W compensation sub-models are injected into the basic control framework, and the control parameters in the basic control framework are dynamically corrected to form the initial optimization model for the corresponding node. Historical operating data and recent operational data of the corresponding kinetic energy recovery node are retrieved as a small sample training set, and gradient backpropagation fine-tuning training is performed on the initial optimization model with the optimization objectives of improving piezoelectric conversion efficiency and enhancing energy recovery stability. After training, W node-level optimization models adapted to the current operating state of each kinetic energy recovery node are output, forming the first set of node-level optimization models.

[0044] Furthermore, after injecting the W compensation sub-model parameters into the basic control framework, small-sample numeric learning gradient fine-tuning is performed to construct W node-level optimization models, forming the first set of node-level optimization models. The method includes: The parameters of the first compensation sub-model are injected into the basic control framework to construct the first initial optimization model. M historical multi-dimensional operating condition data of the first kinetic energy recovery node in M ​​adjacent collision events are backtracked and extracted to construct input features, resulting in M ​​training feature vectors. These training feature vectors include the peak collision pressure, piezoelectric output voltage fluctuation rate, and circuit temperature rise rate. A multi-process joint loss function is constructed with maximizing piezoelectric conversion efficiency as the optimization objective. The M training feature vectors are used as a fine-tuning training set, and gradient backpropagation is performed on the first initial optimization model based on the multi-process joint loss function to update it, outputting the first node-level optimization model. The input of the first node-level optimization model is the real-time operating condition feature vector, and the output is a set of electrical control parameters.

[0045] The parameters of the first compensation sub-model are injected into the basic control framework, and the energy harvesting parameters, energy storage regulation parameters, and output control parameters within the basic control framework are dynamically corrected to form a first initial optimization model suitable for the current operating state of the first kinetic energy recovery node. This first initial optimization model retains the core control logic of the basic control framework and, in conjunction with the parameters of the first compensation sub-model, achieves preliminary adaptation to the current node's operating condition deviation.

[0046] Historical operational data of the first kinetic energy recovery node in M ​​neighboring collision events were retrieved, and time alignment and feature extraction were performed on the mechanical response data, electrical output data, and thermal change data corresponding to each collision event. Multi-dimensional operating condition features such as collision pressure peak, piezoelectric output voltage fluctuation rate, and circuit temperature rise rate were combined and encoded to construct M training feature vectors, which were used as data inputs for subsequent model fine-tuning training.

[0047] With maximizing piezoelectric conversion efficiency and improving energy recovery stability as joint optimization objectives, a corresponding multi-process joint loss function is constructed. This multi-process joint loss function integrates piezoelectric output efficiency loss, circuit energy loss, and output fluctuation error. By jointly constraining multiple optimization objectives, the model maintains output stability and operational reliability while improving energy recovery efficiency.

[0048] The M training feature vectors are used as the fine-tuning training set and input into the first initial optimization model. Gradient backpropagation is performed based on the multi-process joint loss function to update the model's internal control parameters iteratively. After multiple rounds of gradient fine-tuning, a first node-level optimization model suitable for the current operating conditions of the first kinetic energy recovery node is output. This first node-level optimization model takes real-time operating condition feature vectors as input and electrical control parameter sets as output, and is used to generate corresponding energy storage regulation parameters, voltage control parameters, and energy recovery control parameters in real time.

[0049] Furthermore, during the adaptive kinetic energy recovery parameter tuning optimization process of the K homogeneous recovery groups using the K sets of node-level optimization models, the operating condition offset status of the K homogeneous recovery groups is tracked, and the K sets of node-level optimization models are subjected to model smoothing parameter tuning. The method includes: When the magnitudes of the first working condition offset increment and the first working condition feature offset of the first kinetic energy recovery node are greater than the preset dynamic offset threshold, a sliding window is constructed to collect M incremental feature vectors; the M incremental feature vectors are used to perform constrained gradient descent update on the first node-level optimization model to extract the first parameter increment vector; the first parameter increment vector is smoothly fused into the first node-level optimization model to obtain the first node-level parameter tuning model, which replaces the first node-level optimization model to perform subsequent adaptive kinetic energy recovery control.

[0050] During the adaptive kinetic energy recovery control process at the first kinetic energy recovery node, the changes in the first operating condition offset increment and the first operating condition characteristic offset are continuously monitored. When the magnitude of the first operating condition offset increment and the first operating condition characteristic offset exceeds the preset dynamic offset threshold, it is determined that a significant drift has occurred in the current operating condition. Based on a sliding window mechanism, the incremental operating data of the first kinetic energy recovery node within the current time period is continuously collected, and the corresponding vibration change features, piezoelectric output change features, and temperature rise change features are extracted to construct M incremental feature vectors, which are used to characterize the dynamic change trend during the current operating condition drift process.

[0051] The M incremental feature vectors are input into the first node-level optimization model, and constrained gradient descent updates are performed on the model while maintaining basic control stability. Specifically, based on the direction of change in operating conditions corresponding to the current incremental feature vector, the energy adjustment parameters, output control parameters, and dynamic compensation parameters within the model are adjusted using gradients, while limiting the parameter update magnitude to avoid control oscillations caused by abrupt changes in model parameters. After completing the constrained gradient descent update, the corresponding model parameter changes are extracted to obtain the first parameter increment vector.

[0052] The first parameter increment vector is injected into the first node-level optimization model according to a preset smooth fusion strategy, and the parameters of the original model are progressively updated to construct the first node-level parameter tuning model. Specifically, a dynamic weight fusion method is used for buffered updates of the newly added parameters, enabling the model to maintain its original control stability while adapting to new operating conditions. After parameter fusion is completed, the first node-level parameter tuning model replaces the original first node-level optimization model, and subsequent adaptive kinetic energy recovery control continues.

[0053] Furthermore, the K edge execution modules collect group-level operating condition-control incremental data based on a preset configurable synchronization period and upload it to the federated cloud for collaborative evolution of the global base model.

[0054] The K edge execution modules collect and encapsulate group-level operating condition data, model parameter incremental data, and control adjustment results generated by each homogeneous recycling group during operation, according to a preset configurable synchronization cycle. The corresponding group-level operating condition-control incremental data is uploaded to the federated cloud, where it aggregates, analyzes, and collaboratively trains the data uploaded by multiple edge execution modules. Joint optimization and parameter evolution are performed on the global base model to improve its general adaptability and global optimization capabilities across different automated production line scenarios.

[0055] Example 2, based on the same inventive concept as the kinetic energy recovery optimization method for automated production lines in the foregoing examples, such as... Figure 2 As shown in the figure, this application provides a kinetic energy recovery optimization system for automated production lines, the system comprising: The multi-source working condition sensing unit 10 is used to perform multi-source working condition sensing during the baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in an automated production line, in order to construct multiple energy fingerprint feature vectors. The dynamic partitioning unit 20 is used to dynamically divide the multiple kinetic energy recovery nodes into K homogeneous recovery groups based on the energy similarity of the multiple energy fingerprint feature vectors. The master node election unit 30 is used to elect master nodes for the K homogeneous recovery groups, obtain K group representative fingerprints, and match K group-specific optimization base models in a pre-trained meta-model library. The adaptation and parameter tuning unit 40 is used to combine the real-time working condition data stream of the K homogeneous recovery groups to perform intra-group small sample meta-learning adaptation and parameter tuning of the K group-specific optimization base models, and construct K group node-level optimization models. The model smoothing and parameter tuning unit 50 is used to drive K edge execution modules to implement adaptive kinetic energy recovery parameter tuning optimization for the K homogeneous recovery groups using the K group node-level optimization models, track the working condition offset state of the K homogeneous recovery groups, and perform model smoothing and parameter tuning of the K group node-level optimization models.

[0056] Furthermore, the dynamic partitioning unit 20 is used to perform the following operation steps: Based on the characteristic dimensions of energy fingerprints, a high-dimensional feature space is constructed; the multiple energy fingerprint feature vectors are mapped to the high-dimensional feature space to form multiple feature particles; based on a preset energy similarity diameter, density clustering of the multiple feature particles is performed in the feature space to obtain K homogeneous particle groups; based on the K homogeneous particle groups, dynamic grouping mapping of the multiple kinetic energy recovery nodes is performed to generate the K homogeneous recovery groups; the local control permissions of the K homogeneous recovery groups are revoked and bound to the K edge execution modules to complete the allocation of local control permissions.

[0057] Furthermore, the master node election unit 30 is used to perform the following operation steps: Energy recovery process indicators are extracted from W kinetic energy recovery nodes in the first homogeneous recovery group to obtain W node energy efficiency feature vectors. The node energy efficiency feature vectors include mechanical vibration capture rate, piezoelectric conversion efficiency, circuit conversion loss rate, and node operation failure rate. Based on a preset dynamic weighting strategy, a comprehensive score is calculated for the W node energy efficiency feature vectors. After obtaining the comprehensive score of W nodes, the highest score is located in descending order to determine the first group of representative recovery nodes. The energy condition features of the first group of representative recovery nodes are used as the first group of representative fingerprints and matched with the first group of dedicated optimization base models in the pre-trained meta-model library.

[0058] Furthermore, the master node election unit 30 is used to perform the following operation steps: The first set of representative fingerprints is used to perform energy pattern matching, operating condition label matching, and topological position matching in the energy pattern layer, operating condition label layer, and topological position layer of the pre-trained meta-model library, respectively, to obtain the basic physical optimization model, electromechanical coupling optimization model, and structural resonance suppression model. With the basic physical optimization model as the core framework, the parameters of the electromechanical coupling optimization model and the structural resonance suppression model are superimposed to complete the dynamic generation of the first set of dedicated optimization base models.

[0059] Furthermore, the adaptation parameter tuning unit 40 is used to perform the following operation steps: Extract the basic control framework of the first set of dedicated optimization base models; extract real-time operating condition data stream features of the W kinetic energy recovery nodes in the first homogeneous recovery group to generate W node real-time operating condition features, compare them with the energy operating condition features of the first set of representative recovery nodes, and calculate the W operating condition feature offsets; perform progressive matching and selection in a dynamic compensation sub-model library built based on a preset compensation strategy according to the W operating condition feature offsets to obtain W compensation sub-model parameters; after injecting the W compensation sub-model parameters into the basic control framework, perform small sample learning gradient fine-tuning to construct W node-level optimization models, forming the first set of node-level optimization models.

[0060] Furthermore, the adaptation parameter tuning unit 40 is used to perform the following operation steps: The parameters of the first compensation sub-model are injected into the basic control framework to construct the first initial optimization model. M historical multi-dimensional operating condition data of the first kinetic energy recovery node in M ​​adjacent collision events are backtracked and extracted to construct input features, resulting in M ​​training feature vectors. These training feature vectors include the peak collision pressure, piezoelectric output voltage fluctuation rate, and circuit temperature rise rate. A multi-process joint loss function is constructed with maximizing piezoelectric conversion efficiency as the optimization objective. The M training feature vectors are used as a fine-tuning training set, and gradient backpropagation is performed on the first initial optimization model based on the multi-process joint loss function to update it, outputting the first node-level optimization model. The input of the first node-level optimization model is the real-time operating condition feature vector, and the output is a set of electrical control parameters.

[0061] Furthermore, the model smoothing parameter tuning unit 50 is used to perform the following operation steps: When the magnitudes of the first working condition offset increment and the first working condition feature offset of the first kinetic energy recovery node are greater than the preset dynamic offset threshold, a sliding window is constructed to collect M incremental feature vectors; the M incremental feature vectors are used to perform constrained gradient descent update on the first node-level optimization model to extract the first parameter increment vector; the first parameter increment vector is smoothly fused into the first node-level optimization model to obtain the first node-level parameter tuning model, which replaces the first node-level optimization model to perform subsequent adaptive kinetic energy recovery control.

[0062] Furthermore, the K edge execution modules collect group-level operating condition-control incremental data based on a preset configurable synchronization period and upload it to the federated cloud for collaborative evolution of the global base model.

[0063] Through the foregoing detailed description of the kinetic energy recovery optimization method for automated production lines, those skilled in the art can clearly understand the kinetic energy recovery optimization system for automated production lines in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0064] Example 3 provides a storage medium on which a computer program is stored, which, when executed by a processor, implements any step of Example 1.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for optimizing kinetic energy recovery in automated production lines, characterized in that, The method includes: During the baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in an automated production line, multi-source operating condition sensing is performed to construct multiple energy fingerprint feature vectors. Based on the energy similarity of the multiple energy fingerprint feature vectors, the multiple kinetic energy recovery nodes are dynamically divided into K homogeneous recovery groups; Master node election is performed on the K homogeneous recycling groups to obtain K group representative fingerprints, and K group-specific optimized base models are matched in the pre-trained meta-model library. By combining the real-time operating data streams of the K homogeneous recycling groups, the K group-specific optimization base models are adapted and tuned using small sample meta-learning within the groups to construct K group node-level optimization models. During the process of adaptive kinetic energy recovery parameter tuning optimization of the K homogeneous recycling groups using the K sets of node-level optimization models, the operating condition offset status of the K homogeneous recycling groups is tracked, and the K sets of node-level optimization models are smoothed and tuned.

2. The energy recovery optimization method for automated production lines as described in claim 1, characterized in that, Based on the energy similarity of the multiple energy fingerprint feature vectors, the multiple kinetic energy recovery nodes are dynamically divided into K homogeneous recovery groups. The method includes: Based on the characteristic dimensions of energy fingerprints, a high-dimensional feature space is constructed; The multiple energy fingerprint feature vectors are mapped to the high-dimensional feature space to form multiple feature particles; Based on a preset energy similarity diameter, density clustering of the multiple feature particles is performed in the feature space to obtain K homogeneous particle groups. Based on the K homogeneous particle groups, the multiple kinetic energy recovery nodes are dynamically grouped and mapped to generate the K homogeneous recovery groups; The local control permissions of the K homogeneous recycling groups are revoked and bound to the K edge execution modules to complete the allocation of local control permissions.

3. The energy recovery optimization method for automated production lines as described in claim 1, characterized in that, The method involves electing a master node for each of the K homogeneous recycling groups to obtain K group representative fingerprints, and matching K group-specific optimized base models in a pre-trained meta-model library. Energy recovery process indicators are extracted from W kinetic energy recovery nodes in the first homogeneous recovery group to obtain W node energy efficiency feature vectors, wherein the node energy efficiency feature vectors include mechanical vibration capture rate, piezoelectric conversion efficiency, circuit conversion loss rate and node operation failure rate. Based on a preset dynamic weighting strategy, a comprehensive score is calculated for the energy efficiency feature vectors of the W nodes. After obtaining the comprehensive score of the W nodes, the highest score is located in descending order to determine the first group of representative recycling nodes. The energy condition features of the first group representing the recycling nodes are used as the first group of representative fingerprints, and matched with the first group of dedicated optimized base models in the pre-trained meta-model library.

4. The energy recovery optimization method for automated production lines as described in claim 3, characterized in that, The method involves using the energy condition features of the first group representing the recycling nodes as the first group of representative fingerprints, and matching them with the first group of dedicated optimized base models in the pre-trained meta-model library. The first set of representative fingerprints is used to perform energy pattern matching, working condition label matching and topological position matching in the energy pattern layer, working condition label layer and topological position layer of the pre-trained meta-model library, respectively, to obtain the basic physical optimization model, electromechanical coupling optimization model and structural resonance suppression model. Using the aforementioned basic physical optimization model as the core framework, the parameters of the electromechanical coupling optimization model and the structural resonance suppression model are superimposed to complete the dynamic generation of the first set of dedicated optimization base models.

5. The energy recovery optimization method for automated production lines as described in claim 4, characterized in that, Combining the real-time operating data streams of the K homogeneous recycling groups, the K group-specific optimization base models are adapted and tuned using small-sample meta-learning within each group to construct K group node-level optimization models. The method includes: Extract the basic control framework of the first set of dedicated optimization base models; Real-time operating condition data stream features are extracted from the W kinetic energy recovery nodes in the first homogeneous recovery group to generate W real-time operating condition features of the nodes. The W operating condition features are compared with the energy operating condition features of the first group representing the recovery nodes, and the offset of the W operating condition features is calculated. Based on the W working condition feature offsets, a progressive matching selection is performed in the dynamic compensation sub-model library constructed based on a preset compensation strategy to obtain W compensation sub-model parameters; After injecting the parameters of the W compensation sub-models into the basic control framework, small sample learning gradient fine-tuning is performed to construct W node-level optimization models, forming the first group of node-level optimization models.

6. The energy recovery optimization method for automated production lines as described in claim 5, characterized in that, After injecting the parameters of the W compensation sub-models into the basic control framework, small-sample meta-learning gradient fine-tuning is performed to construct W node-level optimization models, forming the first set of node-level optimization models. The method includes: Inject the parameters of the first compensation sub-model into the basic control framework to construct the first initial optimization model; M historical multi-dimensional operating condition data of the first kinetic energy recovery node in M ​​adjacent collision events are backtracked and extracted to construct input features and obtain M training feature vectors, wherein the training feature vectors include the peak collision pressure, the piezoelectric output voltage fluctuation rate and the circuit temperature rise rate. To maximize piezoelectric conversion efficiency, a joint loss function for multiple processes is constructed. Using the M training feature vectors as a fine-tuning training set, the first initial optimization model is updated by gradient backpropagation based on the multi-process joint loss function, and a first node-level optimization model is output. The input of the first node-level optimization model is the real-time operating condition feature vector, and the output is the set of electrical control parameters.

7. The energy recovery optimization method for automated production lines as described in claim 6, characterized in that, During the adaptive kinetic energy recovery parameter tuning optimization process of the K homogeneous recovery groups using the K sets of node-level optimization models, the operating condition offset status of the K homogeneous recovery groups is tracked, and the K sets of node-level optimization models are subjected to model smoothing parameter tuning. The method includes: When the magnitudes of the first working condition offset increment and the first working condition feature offset of the first kinetic energy recovery node are greater than the preset dynamic offset threshold, a sliding window is constructed to collect M incremental feature vectors. Using the M incremental feature vectors, the first node-level optimization model is updated by constrained gradient descent to extract the first parameter incremental vector; The first parameter increment vector is smoothly fused into the first node-level optimization model to obtain the first node-level parameter tuning model, which replaces the first node-level optimization model to perform subsequent adaptive kinetic energy recovery control.

8. The energy recovery optimization method for automated production lines as described in claim 1, characterized in that, The K edge execution modules collect group-level operating condition-control incremental data based on a preset configurable synchronization period and upload it to the federated cloud for collaborative evolution of the global base model.

9. A kinetic energy recovery optimization system for automated production lines, characterized in that, The system is used to implement the kinetic energy recovery optimization method for automated production lines according to any one of claims 1 to 8, the system comprising: The multi-source working condition sensing unit is used to perform multi-source working condition sensing during the baseline kinetic energy recovery process at multiple kinetic energy recovery nodes in an automated production line, in order to construct multiple energy fingerprint feature vectors. A dynamic partitioning unit is used to dynamically divide the multiple kinetic energy recovery nodes into K homogeneous recovery groups based on the energy similarity of the multiple energy fingerprint feature vectors. The master node election unit is used to elect master nodes for the K homogeneous recycling groups, obtain K group representative fingerprints, and match K group-specific optimized base models in the pre-trained meta-model library. The adaptation and parameter tuning unit is used to combine the real-time operating data streams of the K homogeneous recycling groups to perform intra-group small sample meta-learning adaptation and parameter tuning of the K group-specific optimization base models, and to construct the K group node-level optimization models. The model smoothing parameter tuning unit is used to drive K edge execution modules to implement adaptive kinetic energy recovery parameter tuning optimization for the K homogeneous recovery groups using the K sets of node-level optimization models, track the working condition offset status of the K homogeneous recovery groups, and perform model smoothing parameter tuning for the K sets of node-level optimization models.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the kinetic energy recovery optimization method for an automated production line as described in any one of claims 1 to 8.

Images