Injection molding machine key component remaining life distributed prediction system
By using an edge-cloud collaborative architecture and a self-attention mechanism for feature fusion, the data privacy and cost issues in predicting the remaining life of key components of injection molding machines are resolved. This enables efficient and secure cross-factory collaborative modeling, improving prediction accuracy and generalization ability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MASCH TECH DEV CO LTD
- Filing Date
- 2025-08-11
- Publication Date
- 2026-07-24
Smart Images
Figure CN121030601B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent maintenance technology for industrial equipment, and in particular to a distributed prediction system for the remaining life of key components of an injection molding machine. Background Technology
[0002] In the field of industrial equipment health management and predictive maintenance technology, predicting the remaining useful life (RUL) of key components of injection molding machines is crucial for ensuring production efficiency and reducing maintenance costs. Existing technologies primarily rely on collecting multimodal data such as equipment vibration and pressure, employing centralized data storage and traditional machine learning models (such as LSTM) for RUL prediction. However, this technology faces multiple technical bottlenecks in practical applications: Regarding data compliance, industrial production data, as a carrier of core corporate competitiveness, faces severe data privacy protection issues. Related industry reports indicate that a significant proportion of manufacturing companies refuse to share equipment data due to data privacy risks. Traditional centralized modeling relies on cross-domain transmission of raw data, resulting in significant legal compliance deficiencies. In a typical case, a company faced substantial economic penalties for violating data protection regulations due to the leakage of equipment operation data caused by centralized storage. Regarding system economics, a single injection molding machine generates a massive amount of multimodal data daily. The raw data transport model under a centralized architecture leads to high deployment costs for the Industrial Internet, placing a significant burden on small and medium-sized enterprises. The economic costs of large-scale data transmission have become a major obstacle to technology promotion. Regarding model effectiveness, injection molding machines in different factories exhibit significant differences in operating conditions, and the data shows non-independent and identically distributed characteristics. This causes traditional LSTM models trained on single-factory data to show a significant increase in prediction error when predicting across operating conditions, making it difficult to meet the actual needs of industrial sites for model generalization capabilities. In terms of feature utilization, existing methods use fixed weights to fuse multi-source signals, neglecting the dynamic differences in the importance of features such as vibration and pressure under different operating conditions. They also lack sufficient differential processing of early equipment fault characteristics (such as the high-frequency micro-amplitude vibration characteristics in the early stages of screw wear and the pressure signal characteristics when the hydraulic system is abnormal), resulting in poor key feature recognition and insufficient exploitation of the potential value of multimodal data.
[0003] In summary, existing technologies have systemic deficiencies in terms of legal compliance, system economy, model generalization ability, and efficiency in utilizing multimodal features. There is an urgent need to propose a prediction system for the remaining service life of key components of injection molding machines that takes into account data security, low-cost communication, strong generalization ability, and dynamic feature fusion. Summary of the Invention
[0004] In view of this, this application proposes a distributed prediction system for the remaining life of key components of injection molding machines. Through edge-cloud collaboration, it can improve prediction accuracy across operating conditions, optimize communication costs, and enhance model robustness while ensuring data privacy and compliance, thus meeting the real-time and multi-scenario generalization requirements of industry.
[0005] To achieve the above objectives, this application provides a distributed prediction system for the remaining life of key components of an injection molding machine, including multiple edge computing nodes and a cloud server;
[0006] The edge computing node is deployed in the control cabinet of the injection molding machine in the industrial site and includes a data preprocessing module, a multimodal feature fusion module, and a local model training module. The data preprocessing module is used to perform Savitzky-Golay filtering and Z-score normalization on the multimodal data of the key components of the injection molding machine. The multimodal feature fusion module uses a self-attention mechanism to dynamically adjust the weights of the features in the time domain, frequency domain, and statistical domain to generate fused features. The local model training module is used to train the local model based on the fused features to generate local model parameters.
[0007] The cloud server includes an adaptive federated aggregation module and a secure communication module. The adaptive federated aggregation module is used to receive local model parameters uploaded by each edge computing node, and dynamically weight and aggregate them based on data quality scores and local model errors to generate global model parameters. The secure communication module is used to receive local model parameters uploaded by each edge computing node through an encrypted tunnel, and to distribute the global model parameters to each edge computing node to perform predictions of the remaining lifespan of corresponding key components of the injection molding machine.
[0008] Therefore, this application adopts an edge-cloud collaborative architecture. By deploying edge computing nodes in the control cabinet of an injection molding machine in an industrial setting, and using Savitzky-Golay filtering and Z-score normalization to preprocess the collected multimodal data at the edge computing nodes, noise can be effectively removed and data distribution standardized. Then, a self-attention mechanism is used to dynamically adjust the weights of features in different domains to achieve the fusion of multiple feature vectors. The fused feature vectors are then used for local model training. The cloud server uses an adaptive federated aggregation framework to dynamically weight and aggregate the local model parameters of each edge computing node to generate global model parameters, which are then distributed to each edge computing node using a secure communication module to monitor and predict the remaining lifespan of key components of the injection molding machine. This application coordinates the data processing results of multiple edge computing nodes through the cloud server, realizing cross-factory collaborative data modeling, solving the legal compliance issues faced by traditional centralized modeling methods, and also improving the model's generalization ability under different operating conditions.
[0009] Optionally, the self-attention mechanism of the multimodal feature fusion module converts the preprocessed feature vector into Q, K, and V matrices through a shared linear layer, and calculates the weights in the time domain, frequency domain, and statistical domain using scaled dot product attention; the calculation formula for this scaled dot product attention is:
[0010]
[0011] Where Q represents the query, K represents the key, V represents the value, T represents the time point, and d represents the time point. k The dimension of the key;
[0012] Based on the weights in the time domain, frequency domain, and statistical domain, fused features are generated:
[0013] F fusion =αF time +βF freq +γF stat
[0014] Among them, F fusion For fusion features, F time For time-domain features, F freq For frequency domain characteristics, F stat Let α, β, and γ be the statistical domain features, and let α, β, and γ be the weights, with α + β + γ = 1.
[0015] As described above, the preprocessed feature vectors are converted into Q (Query), K (Key), and V (Value) matrices through a shared linear layer, and the weights in the time domain, frequency domain, and statistical domain are calculated respectively. The fused features are then generated using these weights, achieving efficient and accurate multimodal feature fusion.
[0016] Optionally, the multimodal feature fusion module also employs a condition-sensitive weight compensation mechanism. When a specific abnormal condition is detected, the weight compensation algorithm is triggered to forcibly increase the weight of the abnormal feature.
[0017] As shown above, by introducing a working condition sensitive weight compensation mechanism, the multimodal feature fusion module can increase the weight of relevant abnormal features when a specific abnormal working condition is detected, ensuring that these features are given sufficient attention during the fusion process, thereby enhancing the robustness of the entire feature fusion process and avoiding model performance degradation or decision-making errors due to the neglect of abnormal features.
[0018] Optionally, the edge computing node also employs a self-checking mechanism, which is used to detect abnormalities in the multimodal data of key components of the injection molding machine, and automatically switches to offline prediction mode when an abnormality occurs, and triggers device alarms through the ModbusRTU protocol;
[0019] The self-checking mechanism is also used to execute a completion strategy of cubic spline interpolation, feature dimensionality reduction prediction, and loading a historical weighted average model in sequence when the missing rate of the multimodal data exceeds a set value.
[0020] As described above, the self-checking mechanism can detect anomalies in the multimodal data of key components of the injection molding machine in real time. When an anomaly occurs, it automatically switches to offline prediction mode to avoid system misjudgment or failure due to data anomalies. Simultaneously, it triggers equipment alarms via the Modbus RTU protocol, promptly notifying maintenance personnel to handle the issue and improving the overall reliability and security of the system. When the multimodal data missing rate exceeds a set value, the self-checking mechanism effectively fills in the missing data by sequentially executing a completion strategy of cubic spline interpolation, feature dimensionality reduction prediction, and loading a historical weighted average model. This reduces the impact of missing data on prediction accuracy and ensures the continuous and stable operation of the system.
[0021] Optionally, the data quality scoring system of the adaptive federated aggregation module is based on a comprehensive calculation of signal-to-noise ratio, data integrity, time decay factor, and cosine similarity of operating condition labels. The calculation formula is as follows:
[0022] Q i =0.6Q base +0.2Q time +0.2Q match
[0023] Q base =(SNR) i / 20)·(1-MissingRate i )
[0024] Among them, SNR i MissingRate is the signal-to-noise ratio. i Q represents the data missing rate. time Q is the time decay factor. match The cosine similarity between the working condition label and the global distribution.
[0025] Based on the above, by comprehensively considering multiple key dimensions such as signal-to-noise ratio, data integrity, time decay factor, and cosine similarity of working condition labels, a comprehensive and scientific data quality scoring system has been constructed. The adaptive federated aggregation module can more accurately filter and aggregate high-quality data, optimize parameter updates and model aggregation in the federated learning process, thereby improving the performance and efficiency of the entire federated learning system and achieving more reliable and intelligent decision-making and prediction.
[0026] Optionally, the dynamic weighted aggregation process of the adaptive federated aggregation module introduces a Z-score normalization mechanism, and the weighted aggregation formula is:
[0027]
[0028] Among them, w i For aggregate weights, For standardized data quality scoring, For the standardized local model error, Q i E scores the data quality. i Let μQ be the local model error, μQ be the mean of all data quality scores, and σQ be the standard deviation of all data quality scores. E σ is the mean of all local model errors. E This represents the standard deviation of the errors for all local models.
[0029] As described above, by standardizing the data quality score and local model error through the Z-score standardization mechanism, the influence of differences in the units and numerical ranges of different indicators is eliminated. This allows the data quality score and local model error to be compared fairly on the same scale. The aggregation weight is calculated based on the standardized data quality score and local model error, which can more accurately reflect the actual contribution of each participant's data and model. Participants with high data quality and small model error will receive greater weight and play a more important role in the aggregation process, thereby optimizing the overall effect of federated aggregation and improving the performance of the final model.
[0030] Optionally, the adaptive federated aggregation module also employs an abnormal node elimination mechanism. When the error of a node exceeds a threshold and fails to improve for two consecutive rounds, the aggregation of the node's parameters is skipped.
[0031] As shown above, during the federated learning process, individual nodes may encounter abnormal conditions, such as data quality issues or computational errors. By skipping the parameter aggregation of these abnormal nodes in a timely manner, the interference of abnormal factors on the system can be reduced, and the negative impact of low-quality or erroneous parameters on the overall model can be prevented. This improves the quality of federated aggregation and makes the final trained model more accurate and reliable.
[0032] Optionally, the convergence condition of the adaptive federated aggregation module is that the standard deviation of the MAE of the global model on the validation set is less than a set value for multiple consecutive rounds and the mean is less than the industry threshold. At this time, training is terminated and a new version of the model is released. The validation set is a global dataset spanning all nodes, and a set proportion of historical data is randomly extracted in each round.
[0033] As described above, a global dataset spanning all nodes is used as the validation set, with a set proportion of historical data randomly sampled in each round. This design makes the validation data more representative and diverse, enabling a more comprehensive evaluation of the model's performance under various data distributions and scenarios. Then, the convergence condition is used when the standard deviation of the global model's MAE (mean absolute error) on the validation set is consistently less than a set value and the mean is less than an industry threshold for multiple rounds. This allows for precise judgment of whether the model has reached its ideal state from both statistical stability and practical performance perspectives. Training is only terminated when the model's performance is stable across different rounds (small standard deviation) and the overall error level is lower than the industry-recognized standard (small mean), ensuring that the newly released model version has reliable and high-quality performance.
[0034] Optionally, the secure communication module also integrates a privacy enhancement mechanism, specifically including:
[0035] Laplace noise is added to the local model parameters of the edge computing node to achieve differential privacy protection;
[0036] The global model parameters after weighted aggregation are compressed by block sparse pruning, while preserving the continuity of key weight matrices to adapt to GPU acceleration.
[0037] Therefore, when integrating privacy enhancement mechanisms into the secure communication module, combining differential privacy protection and block sparse pruning compression, and optimizing the continuity of the weight matrix for GPU acceleration, can significantly improve privacy, communication efficiency, and computational performance.
[0038] Optionally, the multimodal data of the key components of the injection molding machine includes at least one of vibration signals, pressure signals, and temperature signals;
[0039] The data preprocessing module extracts wavelet packet energy entropy of a specific frequency band as the vibration feature vector based on the vibration signal, calculates kurtosis index as the pressure feature vector based on the pressure signal, and calculates standard deviation as the temperature feature vector based on the temperature signal.
[0040] As described above, by collecting multimodal data such as vibration, pressure, and temperature from key components of an injection molding machine, and employing specific feature extraction methods such as wavelet packet energy entropy, kurtosis index, and standard deviation, the most representative and discriminative key features can be extracted from the original multimodal data. This allows for a comprehensive and accurate characterization of the component's operating state from multiple dimensions. Specifically, the wavelet packet energy entropy of the vibration signal reflects the energy distribution and complexity of the vibration; the kurtosis index of the pressure signal reflects the impact characteristics of the pressure signal; and the standard deviation of the temperature signal describes temperature fluctuations. Representing the feature vectors extracted from different types of data in a unified form facilitates subsequent multimodal feature fusion.
[0041] These and other aspects of this application will become more apparent in the description of the following embodiments(s). Attached Figure Description
[0042] Figure 1 A structural diagram of a distributed prediction system for the remaining life of key components of an injection molding machine, provided in an embodiment of this application;
[0043] Figure 2 A structural diagram of the multimodal feature fusion model provided in the embodiments of this application;
[0044] Figure 3 A flowchart of the adaptive federated aggregation process provided in this application embodiment;
[0045] Figure 4 This is a flowchart for removing abnormal nodes provided in an embodiment of this application.
[0046] It should be understood that the dimensions and shapes of the block diagrams in the above structural diagrams are for reference only and should not constitute an exclusive interpretation of the embodiments of this application. The relative positions and inclusion relationships between the block diagrams presented in the structural diagrams are only schematic representations of the structural relationships between the block diagrams, and are not intended to limit the physical connection methods of the embodiments of this application. Detailed Implementation
[0047] The technical solutions provided in this application will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the system architecture and business scenarios provided in the embodiments of this application are mainly for illustrating possible implementations of the technical solutions of this application and should not be construed as the sole limitation on the technical solutions of this application. Those skilled in the art will recognize that the technical solutions provided in this application are equally applicable to similar technical problems as system architectures evolve and new business scenarios emerge.
[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.
[0049] The solutions provided in this application will now be described in detail with reference to the accompanying drawings and embodiments.
[0050] This application provides a distributed prediction system for the remaining life of key components of injection molding machines. Through edge-cloud collaboration, it improves prediction accuracy across operating conditions, optimizes communication costs, and enhances model robustness while ensuring data privacy and compliance, thus meeting the real-time and multi-scenario generalization requirements of industry.
[0051] like Figure 1 As shown in the embodiment of this application, a distributed prediction system for the remaining life of key components of an injection molding machine is provided, including multiple edge computing nodes 100 and a cloud server 200.
[0052] The edge computing node 100 is deployed in the control cabinet of an injection molding machine in an industrial setting. It includes a data preprocessing module 110, a multimodal feature fusion module 120, and a local model training module 130. The data preprocessing module 110 is used to perform Savitzky-Golay filtering and Z-score normalization on the multimodal data of key components of the injection molding machine. The multimodal feature fusion module 120 uses a self-attention mechanism to dynamically adjust the weights of features in the time domain, frequency domain, and statistical domain to generate fused features. The local model training module 130 is used to train a local model based on the fused features to generate local model parameters.
[0053] The cloud server 200 includes an adaptive federated aggregation module 210 and a secure communication module 220. The adaptive federated aggregation module 210 is used to receive local model parameters uploaded by each edge computing node 100, and dynamically weight and aggregate them based on data quality scores and local model errors to generate global model parameters. The secure communication module 220 is used to receive local model parameters uploaded by each edge computing node 100 through an encrypted tunnel, and to distribute the global model parameters to each edge computing node 100 to perform prediction of the remaining life of corresponding key components of the injection molding machine.
[0054] In some embodiments, multimodal data of key components of the injection molding machine may include vibration signals, pressure signals, and temperature signals, etc., as shown in the reference. Figure 2 As shown, the data preprocessing module performs Savitzky-Golay filtering on the vibration signal and Z-score normalization on the temperature signal. Based on the filtered vibration signal, it extracts wavelet packet energy entropy of a specific frequency band as the vibration feature vector, which reflects the energy distribution and complexity of the vibration. It calculates the kurtosis index based on the pressure signal as the pressure feature vector, which reflects the impact characteristics of the pressure signal. It calculates the standard deviation based on the normalized temperature signal as the temperature feature vector, which describes the temperature fluctuation. By representing the feature vectors extracted from different types of data in a unified form, the operating status of the component can be comprehensively and accurately characterized from multiple dimensions.
[0055] Based on the aforementioned vibration feature vector, pressure feature vector, and temperature feature vector, the multimodal feature fusion module can dynamically adjust the weights of the time domain, frequency domain, and statistical domain features using a self-attention mechanism. Specifically, the self-attention mechanism of this multimodal feature fusion module converts the preprocessed feature vectors into Q, K, and V matrices through a shared linear layer, and calculates the weights of the time domain, frequency domain, and statistical domain using scaled dot product attention. These weights determine the contribution of each feature to the local model. The formula for calculating the scaled dot product attention is:
[0056]
[0057] Where Q represents the query, K represents the key, V represents the value, T represents the time point, and d represents the time point. k The dimension of the key.
[0058] Based on the weights in the time domain, frequency domain, and statistical domain, fused features are generated:
[0059] F fusion =αF time +βF freq +γF stat
[0060] Among them, F fusion For fusion features, F time For time-domain features, F freq For frequency domain characteristics, F stat Let α, β, and γ be the statistical domain features, and let α + β + γ = 1.
[0061] In some embodiments, the multimodal feature fusion module also supports a condition-sensitive weight compensation mechanism. When a specific abnormal condition is detected, the weight compensation algorithm is triggered to forcibly increase the weight of the abnormal feature, ensuring that these features receive sufficient attention during the fusion process. This enhances the robustness of the entire feature fusion process and avoids model performance degradation or decision-making errors due to the neglect of abnormal features. For example, when hydraulic cavitation (pressure signal kurtosis > 3) or screw wear (vibration energy entropy increase > 20%) is detected, the weight compensation algorithm is triggered: β′=β·(1+δ), α′=α·(1-δ) (δ=0.15), correspondingly forcibly increasing the weight of the pressure feature or vibration feature to a specific level (e.g., 40%) to enhance the ability to identify abnormal conditions.
[0062] The local model training module, based on the aforementioned fusion features, uses a suitable time series analysis model (such as a 1DCNN-GRU hybrid network, LSTM, etc.) to train the local model and generate local model parameters.
[0063] In some embodiments, the edge computing node also supports a self-checking mechanism. This mechanism detects anomalies in the multimodal data of key components of the injection molding machine and automatically switches to offline prediction mode when anomalies occur, triggering a device alarm via the Modbus RTU protocol. Furthermore, when the missing rate of multimodal data exceeds a set value (e.g., 5%), the self-checking mechanism effectively fills in the missing data by sequentially executing a completion strategy of cubic spline interpolation, feature dimensionality reduction prediction, and loading a historical weighted average model. This reduces the impact of missing data on prediction accuracy and ensures the continuous and stable operation of the system.
[0064] After the local model training module generates local model parameters, they can be uploaded to the cloud server through the encrypted tunnel of the secure communication module. This allows the cloud server's adaptive federated aggregation module to dynamically weight and aggregate data based on the data quality score and the local model error, generating global model parameters. Specifically, the data quality scoring system of this adaptive federated aggregation module is based on a comprehensive calculation of signal-to-noise ratio, data integrity, time decay factor, and cosine similarity of operating condition labels. The calculation formula is as follows:
[0065] Q i =0.6Q base +0.2Q time +0.2Q match
[0066] Q base =(SNR) i / 20)·(1-MissingRate i )
[0067] Among them, SNR i MissingRate is the signal-to-noise ratio. i Q represents the data missing rate. time Q is the time decay factor. match The cosine similarity between the working condition label and the global distribution.
[0068] The adaptive federated aggregation module incorporates a Z-score standardization mechanism into its dynamic weighted aggregation process. This mechanism standardizes data quality scores and local model errors, and then calculates aggregation weights based on these standardized scores and errors, resulting in a more accurate and generalizable global model. The weighted aggregation formula is as follows:
[0069]
[0070] Among them, w i For aggregate weights, For standardized data quality scoring, For the standardized local model error, Qi E scores the data quality. i Let μQ be the local model error, μQ be the mean of all data quality scores, and σQ be the standard deviation of all data quality scores. E σ is the mean of all local model errors. E This represents the standard deviation of the errors for all local models.
[0071] In some embodiments, the convergence condition of the adaptive federated aggregation module is that the standard deviation of the MAE (mean absolute error) of the global model on the validation set is less than a set value (0.3%) for multiple consecutive rounds, and the mean is less than an industry threshold (7% for the screw prediction scenario). At this point, training is terminated and a new version of the model is released. The validation set is a global dataset spanning all nodes, with a set proportion of historical data randomly sampled in each round. This design makes the validation data more representative and diverse, enabling a more comprehensive evaluation of the model's performance under various data distributions and scenarios.
[0072] In some embodiments, the secure communication module of the cloud server also integrates a privacy enhancement mechanism, as shown in the following figures. Figure 3 As shown, this secure communication module combines differential privacy protection and block sparse pruning compression by allocating an independent privacy budget ε to each edge node. i Alternatively, a global budget can be used and dynamically allocated based on node contribution. Based on this privacy budget, Laplace noise is added to the local model parameters of each edge computing node to resist membership inference attacks and data reconstruction attacks, achieving differential privacy protection. The noise-added local model parameters are uploaded to the adaptive federated aggregation module via an encrypted tunnel of the secure communication module for weighted aggregation, generating global model parameters. Before distributing the global model parameters, the secure communication module performs block sparse pruning and compression on the weighted aggregated global model parameters, preserving the continuity of the key weight matrix to adapt to the parallel computing characteristics of the GPU. The pruned sparse model parameters are then distributed to each edge computing node, enabling each edge computing node to perform predictions of the remaining lifespan of corresponding key components of the injection molding machine.
[0073] In some embodiments, to ensure the effectiveness of aggregation, the adaptive federated aggregation module also employs an abnormal node removal mechanism to reduce the interference of abnormal factors on the system and prevent low-quality or erroneous parameters from negatively impacting the overall model. (See also...) Figure 4 As shown, by calculating the error of each node's parameters at the start of aggregation, when the error of a certain edge calculation node exceeds a threshold (e.g., node error E), the aggregation is completed. i >μE+2σ EIf the node fails to improve after two consecutive rounds, it can be marked as abnormal and skipped from the aggregation of its parameters. The self-check mechanism of the edge computing node will be triggered. Once the self-check of the edge computing node passes, it can resume participating in the weighted aggregation of the adaptive federated aggregation module. If the self-check fails, it will continue to be marked as abnormal until the self-check passes.
[0074] The following is an exemplary description of the prediction process of the distributed prediction system for the remaining life of key components of injection molding machines according to specific working conditions. Specifically, it involves screw remaining life prediction, hydraulic system abnormal working condition handling, and mold life prediction. Through dynamic weighting of features perceived by working conditions, federated learning emergency reconfiguration, and multimodal hybrid modeling, it achieves full-scenario coverage from conventional prediction to fault response. Moreover, all raw data is retained at the edge layer, meeting the ISO 20889 privacy compliance requirements.
[0075] To predict the remaining life of the screw, a vibration acceleration sensor (range ±50g, sampling rate 10kHz), a temperature thermocouple (accuracy ±0.5℃), and a pressure transmitter (0-50MPa) were used to collect parameters such as screw speed (50-200rpm), back pressure (5-30MPa), and time-domain waveform of the vibration signal. Then, an 8-layer wavelet packet decomposition was performed on the vibration signal using a data preprocessing module to extract the energy entropy in the 5-10kHz frequency band as wear features. When the kurtosis of the pressure signal was >3, the cavitation risk of the hydraulic system was marked, and the pressure feature weight was increased to 40%. Local training was performed using a 2-layer LSTM network (128 units per layer, tanh activation function, Adam optimizer, learning rate 0.001) in the local model training module, with a privacy budget ε. i =0.5, generating a round of local model parameters every 2 hours and uploading them to the cloud server. The adaptive federated aggregation module calculates weights based on data quality (60%) and model error (40%) during global aggregation, with an aggregation cycle of 24 hours. A model version update is triggered when the MAE on the validation set is less than 8% for three consecutive rounds.
[0076] When handling abnormal operating conditions in the hydraulic system, if the approximate entropy of the oil pressure signal exceeds the threshold of 1.5 (normal range 0.8-1.2), it is determined to be an abnormal fluctuation, triggering an emergency aggregation mechanism. The edge computing node automatically switches to fault feature extraction mode and adds the calculation of cross-correlation coefficients of flow signals. The adaptive federated aggregation module shortens the aggregation cycle from 72 hours to 4 hours, reduces the weight of data quality scoring to 20%, and increases the weight of local model error to 80%, prioritizing the adoption of local models with a recent MAE drop of 15% (with a weighting of up to 30%) to quickly aggregate abnormal operating condition data. When the remaining life prediction value is less than 200 hours, the edge computing node sends a warning signal to the PLC via the RS485 interface and triggers the automatic generation of equipment maintenance work orders.
[0077] When predicting mold life, the clamping mechanism of the injection molding machine is monitored using the clamping force curve (sampling frequency 100Hz, range 0-5000kN) and mold temperature distribution (infrared thermal imager, resolution 0.1℃, scan cycle 10 seconds). The dynamic time warping (DTW) distance is calculated from the clamping force signal; a distance > 0.8 indicates moderate cavity wear. Principal component analysis (PCA) is used to reduce the dimensionality of the temperature signal, and regions with temperature differences exceeding 5℃ are extracted as thermal fatigue features. A CNN-Transformer hybrid network is used for modeling; the CNN layer extracts the spatial features of the temperature field, and the Transformer layer captures the temporal dependence of the clamping force. A privacy budget ε is also used. i =0.6 (medium sensitive data protection level), output the predicted value of the remaining life of the mold parting surface.
[0078] In summary, this application adopts an edge-cloud collaborative architecture. By deploying edge computing nodes in the control cabinet of an injection molding machine in an industrial setting, and using Savitzky-Golay filtering and Z-score normalization to preprocess the collected multimodal data at the edge computing nodes, noise can be effectively removed and data distribution standardized. Then, a self-attention mechanism is used to dynamically adjust the weights of features in different domains to achieve the fusion of multiple feature vectors. The fused feature vectors are then used for local model training. The cloud server uses an adaptive federated aggregation framework to dynamically weight and aggregate the local model parameters of each edge computing node to generate global model parameters, which are then distributed to each edge computing node using a secure communication module to monitor and predict the remaining lifespan of key components of the injection molding machine. This application coordinates the data processing results of multiple edge computing nodes through a cloud server, realizing cross-factory collaborative data modeling, solving the legal compliance issues faced by traditional centralized modeling methods, and also improving the model's generalization ability under different operating conditions.
[0079] It should be noted that the embodiments described in this application are merely some embodiments, not all embodiments. The components of the embodiments of this application typically described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the above detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0080] The terms "first, second, third, etc." or similar terms such as module A, module B, module C, etc., used in the specification and claims are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that a specific order or sequence may be interchanged where permitted so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.
[0081] In the above description, the labels indicating the steps do not necessarily mean that the steps will be executed. They may include intermediate steps or be replaced by other steps. Where permissible, the order of the steps may be interchanged or executed simultaneously.
[0082] The term "comprising" as used in the specification and claims should not be construed as limiting itself to what follows; it does not exclude other elements or steps. Therefore, it should be interpreted as specifying the presence of the mentioned feature, integral, step, or component, but does not exclude the presence or addition of one or more other features, integrals, steps, or components, or groups thereof. Thus, the statement "device comprising means A and B" should not be limited to a device consisting solely of components A and B.
[0083] The terms "an embodiment" or "an embodiment" as used in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in at least one embodiment of this application. Therefore, the terms "in one embodiment" or "in an embodiment" appearing throughout this specification do not necessarily refer to the same embodiment, but may refer to the same embodiment. Furthermore, in the various embodiments of this application, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions between different embodiments are consistent and can be mutually referenced. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.
[0084] Note that the above are merely preferred embodiments and the technical principles employed in this application. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present application has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, all of which fall within the scope of protection of the present invention.
Claims
1. A distributed prediction system for the remaining life of key components of an injection molding machine, characterized in that, Includes multiple edge computing nodes and cloud servers; The edge computing node is deployed in the control cabinet of the injection molding machine in the industrial site and includes a data preprocessing module, a multimodal feature fusion module, and a local model training module. The data preprocessing module is used to perform Savitzky-Golay filtering and Z-score normalization on the multimodal data of the key components of the injection molding machine. The multimodal data includes at least one of vibration signal, pressure signal, and temperature signal. The data preprocessing module extracts wavelet packet energy entropy of a specific frequency band as a vibration feature vector based on the vibration signal, calculates kurtosis index as a pressure feature vector based on the pressure signal, and calculates standard deviation as a temperature feature vector based on the temperature signal. Based on the vibration feature vector, pressure feature vector, and temperature feature vector, the multimodal feature fusion module uses a self-attention mechanism to dynamically adjust the weights of the time domain, frequency domain, and statistical domain features to generate fused features. The local model training module is used to train a local model based on the fused features and generate local model parameters. The cloud server includes an adaptive federated aggregation module and a secure communication module. The adaptive federated aggregation module is used to receive local model parameters uploaded by each edge computing node, and dynamically weight and aggregate them based on data quality scores and local model errors to generate global model parameters. The secure communication module is used to receive local model parameters uploaded by each edge computing node through an encrypted tunnel, and to distribute the global model parameters to each edge computing node to perform predictions of the remaining lifespan of corresponding key components of the injection molding machine.
2. The system according to claim 1, characterized in that, The self-attention mechanism of the multimodal feature fusion module converts the preprocessed feature vectors into Q, K, and V matrices through a shared linear layer, and calculates the weights in the time domain, frequency domain, and statistical domain using scaled dot product attention; the calculation formula for this scaled dot product attention is as follows: ； in, For query, As key, For value, For a point in time, The dimension of the key; Based on the weights in the time domain, frequency domain, and statistical domain, fused features are generated: ； in, As a feature of fusion, For time domain characteristics, For frequency domain characteristics, For statistical domain features, , , These are the weights, and .
3. The system according to claim 1, characterized in that, The multimodal feature fusion module also employs a condition-sensitive weight compensation mechanism. When a specific abnormal condition is detected, the weight compensation algorithm is triggered to forcibly increase the weight of the abnormal feature.
4. The system according to claim 1, characterized in that, The edge computing node also employs a self-testing mechanism, which is used to detect abnormalities in the multimodal data of key components of the injection molding machine and automatically switches to offline prediction mode when an abnormality occurs, and triggers device alarms through the Modbus RTU protocol. The self-checking mechanism is also used to execute a completion strategy of cubic spline interpolation, feature dimensionality reduction prediction, and loading a historical weighted average model in sequence when the missing rate of the multimodal data exceeds a set value.
5. The system according to claim 1, characterized in that, The data quality scoring system of the adaptive federated aggregation module is based on a comprehensive calculation of signal-to-noise ratio, data integrity, time decay factor, and cosine similarity of operating condition labels. The calculation formula is as follows: ； ； in, For signal-to-noise ratio, For data missing rate, The time decay factor, The cosine similarity between the working condition label and the global distribution.
6. The system according to claim 1 or 5, characterized in that, The adaptive federated aggregation module introduces a Z-score normalization mechanism into its dynamic weighted aggregation process, and the weighted aggregation formula is as follows: ； ； ； in, For aggregate weights, For standardized data quality scoring, For standardized local model error, Score the data quality. For local model error, The mean of all data quality scores. The standard deviation of all data quality scores. The mean of all local model errors. This represents the standard deviation of the errors for all local models.
7. The system according to claim 1, characterized in that, The adaptive federated aggregation module also employs an abnormal node elimination mechanism. When the error of a node exceeds a threshold and fails to improve for two consecutive rounds, the aggregation of the node's parameters is skipped.
8. The system according to claim 1, characterized in that, The convergence condition of the adaptive federated aggregation module is that the standard deviation of the MAE of the global model on the validation set is less than a set value for multiple consecutive rounds and the mean is less than the industry threshold. At this time, training is terminated and a new version of the model is released. The validation set is a global dataset spanning all nodes, and a set proportion of historical data is randomly extracted in each round.
9. The system according to claim 1, characterized in that, The secure communication module also integrates a privacy enhancement mechanism, specifically including: Laplace noise is added to the local model parameters of the edge computing node to achieve differential privacy protection; The global model parameters after weighted aggregation are compressed by block sparse pruning, while preserving the continuity of key weight matrices to adapt to GPU acceleration.