A life cycle assessment system, method, and computer program product

Abstract

This invention discloses a life cycle assessment system, method, and computer program product, belonging to the field of life cycle assessment technology. It includes a data acquisition module, an artificial intelligence prediction module, an environmental impact calculation module, and a feedback and visualization module. The data acquisition module collects raw data during equipment operation. The artificial intelligence prediction module receives the raw data and predicts resource flow data. The environmental impact calculation module receives the raw data and the resource flow data and converts them into standardized life cycle assessment indicators. The feedback and visualization module displays the life cycle assessment indicators and generates parameter adjustment instructions. This invention achieves real-time environmental impact assessment and sustainability optimization in hybrid manufacturing processes.

Description

Technical Field

[0001] This invention relates to the field of life cycle assessment technology, and in particular to a life cycle assessment system, method and computer program product. Background Technology

[0002] Hybrid manufacturing is a combined process flow that includes additive manufacturing, computer numerical control (CNC) machining, and laser polishing. Currently, artificial intelligence (AI) modules have been introduced into the field of manufacturing control to predict the mechanical properties of parts or production cycles; however, existing AI modules also take environmental factors into account when making predictions. From a technological standpoint, there is currently no integrated architecture that can simultaneously achieve real-time data acquisition from sensors, predict unmeasured flows (such as future power consumption and powder consumption) based on AI, and ultimately dynamically calculate the lifecycle assessment indicators of the target workpiece produced within the manufacturing equipment control cycle. Summary of the Invention

[0003] The present invention aims to overcome the inability of existing technologies to collaboratively collect real-time device data and predict life cycle assessment indicators, and provides a life cycle assessment system, method and computer program product.

[0004] To achieve the above objectives, the technical solution of this invention is as follows: a life cycle assessment system is provided, comprising a data acquisition module, an artificial intelligence prediction module, an environmental impact calculation module, and a feedback and visualization module; the data acquisition module collects raw data during equipment operation; the artificial intelligence prediction module receives the raw data and predicts resource flow data; the environmental impact calculation module receives the raw data and the resource flow data and converts them into standardized life cycle assessment indicators; the feedback and visualization module displays the life cycle assessment indicators and generates parameter adjustment instructions.

[0005] In one embodiment, the artificial intelligence prediction module employs a supervised learning regression model, combining a multilayer perceptron and a Gaussian process regressor to construct a hybrid model; the environmental impact calculation module employs a multi-stage data processing algorithm and uses an embedded Python computing engine, and the environmental impact calculation module calls feature factors from the Ecoinvent or ReCiPe database.

[0006] In one embodiment, a central control processor is also included, which manages data synchronization between modules, executes closed-loop control commands, and interfaces with the controllers of the devices; the central control processor and the controllers of each device use a standardized network time protocol to add timestamps to the original data packets.

[0007] In one embodiment, the artificial intelligence prediction module and the environmental impact calculation module are deployed on a cloud platform, while the data acquisition module and the feedback and visualization module run at the edge. The artificial intelligence prediction module adopts a multi-layer architecture, including: a data integrity pipeline for standardizing and denoising multi-source signals; feature engineering for manufacturing; and stacked model integration including calibration, uncertainty quantification, and online drift processing.

[0008] This invention also provides a life cycle assessment method for hybrid manufacturing, comprising the following steps when implementing the aforementioned life cycle assessment system:

[0009] Step S1: Start the hybrid manufacturing process and use the data acquisition module to acquire raw data;

[0010] Step S2: The artificial intelligence prediction module makes predictions based on the original data to obtain resource flow data;

[0011] Step S3: The environmental impact calculation module converts the resource flow data and the raw data into life cycle assessment indicators;

[0012] In step S4, the feedback and visualization module receives the lifecycle assessment indicators and transmits adjustment suggestions to the operator or control system.

[0013] In one embodiment, step S3 includes the following:

[0014] Step S31: Set the time window length according to the process cycle time, and aggregate the raw data and resource flow data acquired within the time window to obtain the resource flow vector of the current time window. And the data for the newly added time period is processed through a sliding time window mechanism;

[0015] Step S32: Model the resource flow vector using the mean-variance or quantile (10%, 50%, 90% quantile) form. The range of uncertainty;

[0016] Step S33, the resource flow vector with the uncertainty interval is... The data is matched with process datasets in the database, and the ReCiPe midpoint method is used to extract feature factors from the corresponding process datasets. ;

[0017] Step S34, based on the resource flow vector and characteristic factors The following formula is used to calculate the environmental impact indicators of each category within a single time window. ,

[0018]

[0019] in, The first in the current time window i Environmental impact indicators of class For the first j Class resource flow vector, For the first j The resource flow vector of class corresponds to the first i Characterization factors of environmental impact indicators of a class; The first time window i Class of environmental impact indicators,

[0020] and ,

[0021] The minimum environmental impact within the time window t; The maximum environmental impact within the time window t;

[0022] Step S35: Calculate the environmental impact indicators for each category obtained from each time window. By accumulating the uncertainties over time, life cycle assessment indicators are obtained. And output it to the feedback and visualization module;

[0023] In steps S31 and S32, the original data within the t-th time window is represented as follows:

[0024]

[0025] in, To measure the mean, The variance of the sensor measurements;

[0026] The resource flow data within the t-th time window is represented as follows:

[0027]

[0028] in, This represents the 10th percentile of the resource flow data; This represents the 50th percentile of the resource flow data; This represents the 90th percentile of the resource flow data;

[0029] A confidence-weighted distribution fusion method is used to assign weights to the confidence levels of the resource stream data and the original data.

[0030]

[0031]

[0032] in, The confidence weights of the original data are... The confidence weight of the resource flow data; the quantile interval width is the difference between the 90th quantile and the 10th quantile predicted by the artificial intelligence prediction module;

[0033] The mean-weighted average and uncertainty-weighted variance are used for calculation.

[0034]

[0035] The raw data and the resource flow data are integrated into a single resource flow vector. ;

[0036] In addition to the aggregate mean, it also transmits the uncertainty boundary, which is calculated using the following formula.

[0037]

[0038] Wherein, is the variance of the resource stream data calculated based on the quantile span, and the final variance This will be retained in subsequent stages.

[0039] In one embodiment, step S2 includes the following steps:

[0040] Step S21: Preprocess historical data and raw data collected by sensors to obtain standard data;

[0041] Step S22: Combine the feature set to perform feature selection on the standard data and identify key features;

[0042] Step S23: After hyperparameter tuning, the hybrid model is trained.

[0043] Step S24: Input the key features and the original data into the trained hybrid model to make predictions and obtain resource flow data;

[0044] Step S25: The resource flow data is transmitted to the environmental impact calculation module for evaluation;

[0045] Step S26: Optimize the model through model retraining.

[0046] In one embodiment, in step S21, the preprocessing includes standardization and interpolation; in step S22, during feature selection, key features related to energy consumption, powder mass flow, and gas consumption are screened based on correlation analysis and principal component analysis; in step S23, the hyperparameter tuning adopts grid search and cross-validation, including learning rate, number of iterations, and layer number parameters; the training of the hybrid model is achieved using gradient boosting and random forest ensemble learning algorithms; in step S24, the resource flow data includes energy consumption, powder mass flow, and gas consumption.

[0047] In one embodiment, in step S2, the data integrity pipeline performs time alignment and resampling on the original data; employs a Hamper filter or selected Kalman filter for noise reduction to suppress outliers; standardizes units; marks missing values ​​and generates quality labels;

[0048] The feature engineering receives time-series data and outputs a multi-dimensional feature matrix X aligned with the production cycle / layer / segment. The feature types include: geometrically derived features, toolpath statistical features, thermal load proxy features, and equipment status codes.

[0049] The stacked model ensemble includes a dual-branch learner comprising a tabular branch and a temporal branch. The tabular branch employs gradient boosting, including XGBoost or LightGBM, to handle static / segmented features, while the temporal branch uses 1D-CNN or LSTM to capture short-term sequence patterns in power / gas trajectories. The outputs of the dual branches are input to a meta-learner to reduce model bias and variance. Based on Bayesian hyperparameter optimization, using TPE or BOHB, the learning rate, tree depth, subsampling rate, CNN kernel size, and LSTM hidden unit parameters are optimized. Block-based time-series cross-validation is employed to ensure temporal order.

[0050] Calibration, uncertainty quantification, and physical constraints are performed. Calibration involves using isotonic regression or Platt scaling on the validation set to correct for systematic bias. Uncertainty quantification involves using quantile regression or Monte Carlo dropout to output the quantile intervals of the prediction results. Physical constraints: Boundary conditions are enforced through a set of rules.

[0051] During online monitoring and adaptive processes, error tracking is performed, and the average absolute percentage error or root mean square error is calculated on a rolling basis according to the process stage; drift detection is performed: ADWIN or KS test is used to detect residual / characteristic drift; if the drift exceeds the threshold, warm start refit, partial fitting or short-term fine-tuning is triggered, and the latest batch data is used;

[0052] In step S3, the environmental impact calculation module utilizes Quantile propagation boundaries are used to obtain interval impact values, and finally, a life cycle assessment index with confidence intervals is output.

[0053] The present invention also provides a computer program product, which, when executed by a processor, implements the steps of a life cycle assessment system as described above.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] Real-time environmental awareness: Environmental impact indicators are dynamically calculated during production (rather than after production ends), enabling sustainable control during the process; Predictive resource management: Hybrid models predict future energy and material consumption, allowing for early identification of high environmental impact conditions; Closed-loop optimization: The feedback and visualization modules transform sustainability data into actionable process adjustment plans, achieving measurable reductions in energy consumption and CO2 emissions; Hybrid process integration: The system integrates electron beam melting, computer numerical control, and laser polishing processes under a unified data framework, achieving complete lifecycle coverage; Industrial scalability: The modular architecture supports integration with existing factory control networks through standardized communication protocols, ensuring broad applicability in innovative manufacturing environments.

[0056] Compared with locally deployed systems, the beneficial effects of this invention are as follows:

[0057] Scalable multi-production line deployment: A single cloud server coordinates multiple hybrid manufacturing units, enabling unified environmental monitoring and centralized control; Centralized model maintenance: Hybrid models and lifecycle assessment databases are updated in the cloud and automatically synchronized to all edge nodes, ensuring consistent calculation accuracy across sites; Reduced equipment load: Edge controllers handle only data acquisition and preprocessing tasks, reducing on-site processor utilization and improving overall system stability; Seamless industrial integration: The architecture supports standardized data exchange with Manufacturing Execution Systems (MES) and Enterprise Resource Planning (ERP) systems, enabling automated sustainability reporting; Secure and traceable data flow: All transmissions are encrypted and logged to ensure data authenticity and compliance with industrial cybersecurity standards.

[0058] To make the above-mentioned features and advantages of the invention more apparent and understandable, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings. Attached Figure Description

[0059] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0060] Figure 2 This is a flowchart of the artificial intelligence prediction module in Embodiment 1 of the present invention.

[0061] Figure 3This is a flowchart of the environmental impact calculation module in Embodiment 1 of the present invention. Detailed Implementation

[0062] To make the objectives and technical solutions of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0063] Example 1

[0064] To achieve real-time lifecycle assessment and environmental impact optimization in hybrid manufacturing processes, this invention provides a lifecycle assessment system. The system includes a data acquisition module, an artificial intelligence prediction module, an environmental impact calculation module, and a feedback and visualization module. These modules are interconnected via a unified data bus and coordinated by a central control processor or industrial edge computer to perform data flow management and ensure synchronization between subsystems. The subsystems include an electron beam melting subsystem, a computer numerical control (CNC) subsystem, and a laser polishing subsystem.

[0065] Combination Figure 1 The overall workflow of the system includes the following steps:

[0066] Step S1: Start the hybrid manufacturing process and use the data acquisition module to acquire raw data;

[0067] Step S2: The artificial intelligence prediction module makes predictions based on the original data to obtain resource flow data;

[0068] Step S3: The environmental impact calculation module converts the resource flow data and the raw data into life cycle assessment indicators;

[0069] In step S4, the feedback and visualization module receives the lifecycle assessment indicators and transmits adjustment suggestions to the operator or control system.

[0070] The data acquisition module includes sensors connected to the electron beam melting equipment, computer numerical control equipment, and laser equipment. The raw data includes continuously acquired operating parameters such as power, beam current, feed rate, and scan rate, and is encapsulated into encrypted data packets for subsequent use.

[0071] Figure 2The workflow of the AI ​​prediction module is demonstrated. Based on sensor data, the module predicts real-time resource flow data and environmental impacts. The AI ​​prediction module employs a supervised learning regression model, combining a multilayer perceptron (MLP) for nonlinear correlation mapping and a Gaussian process regressor for uncertainty estimation to construct a hybrid model. This hybrid model is updated every 100 iterations through incremental learning, incorporating incoming, validated sensor data.

[0072] The operation steps of the artificial intelligence prediction module include:

[0073] Step S21: Preprocess historical data and raw data collected by sensors to obtain standard data;

[0074] Step S22: Combine the feature set to perform feature selection on the standard data and identify key features;

[0075] Step S23: After hyperparameter tuning, the hybrid model is trained.

[0076] Step S24: Input the key features and the original data into the trained hybrid model to make predictions and obtain resource flow data;

[0077] Step S25: The resource flow data is transmitted to the environmental impact calculation module for evaluation;

[0078] Step S26: Optimize the model through model retraining.

[0079] In step S21, preprocessing includes standardization and imputation. Standardization is performed to unify the data units, and missing values ​​are imputed to fill in the data gaps, ensuring the consistency and integrity of the input data.

[0080] In step S22, the feature set includes shape parameters (volume, surface area, mass ratio) and process settings. The key features are related to subsequent predictions. During feature selection, based on correlation analysis and principal component analysis, key features related to energy consumption, powder mass flow, and gas consumption are screened.

[0081] In step S23, before training, the hybrid model undergoes hyperparameter tuning using grid search and cross-validation, including parameters such as learning rate, number of iterations, and number of layers. Ensemble learning algorithms such as Gradient Boosting and Random Forest are employed, and training data is input to train the hybrid model, enabling it to predict energy and material consumption in real time.

[0082] In step S24, the prediction result is resource flow data, including energy and material consumption, specifically including short-term energy consumption (E) for the next control cycle. t+1 Powder mass flow (M) t+1 ) and gas consumption (G t+1 ).

[0083] In step S26, the hybrid model is continuously improved through model retraining, gradually enhancing its predictive capabilities using newly acquired data. This ensures that the system adapts to changing process conditions and continuously optimizes the prediction of resource flow data and environmental impacts.

[0084] In step S3, the environmental impact calculation module adopts an incremental environmental impact calculation method oriented towards uncertainty, transforming the resource flow data and raw data into life cycle assessment indicators, and combining them with... Figure 3 It can be seen that step S3 includes,

[0085] Step S31: Set the time window length according to the process cycle time, and aggregate the raw data and resource flow data acquired within the time window to obtain the resource flow vector of the current time window. And the data for the newly added time period is processed through a sliding time window mechanism;

[0086] Step S32: Model the resource flow vector using the mean-variance or quantile (10%, 50%, 90% quantile) form. The range of uncertainty;

[0087] Step S33, the resource flow vector with the uncertainty interval is... The data is matched with process datasets in the database, and the ReCiPe midpoint method is used to extract feature factors from the corresponding process datasets. ;

[0088] Step S34, based on the resource flow vector and characteristic factors The following formula is used to calculate the environmental impact indicators of each category within a single time window. ,

[0089]

[0090] in, The first in the current time window i Class of environmental impact indicators, For the first j Class resource flow vector, For the first j The resource flow vector of class corresponds to the first i Characterization factors of environmental impact indicators of a class; The first time window i Class of environmental impact indicators,

[0091] and ,

[0092] The minimum environmental impact within the time window t; The maximum environmental impact within the time window t;

[0093] Step S35: Calculate the environmental impact indicators for each category obtained from each time window. By accumulating the uncertainties over time, a life cycle assessment index for the current moment can be obtained. The data is then output to the feedback and visualization module.

[0094] In step S31, short-term energy consumption (E) is... t+1 Powder mass flow (M) t+1 ) and gas consumption (G t+1 Resource flow data, as well as raw data collected by sensors, are converted into unit process values, such as kWh / piece, kg / piece, and Nm³ / piece, to achieve the unification and standardization of data units.

[0095] In steps S31 and S32, the original data within the t-th time window is represented as follows:

[0096]

[0097] in, To measure the mean, The variance of the sensor measurements;

[0098] The resource flow data within the t-th time window is represented as follows:

[0099]

[0100] in, This represents the 10th percentile of the resource flow data; This represents the 50th percentile of the resource flow data; This represents the 90th percentile of the resource flow data;

[0101] A confidence-weighted distribution fusion method is used to assign weights to the confidence levels of the resource stream data and the original data.

[0102]

[0103]

[0104] in, The confidence weights of the original data are... The confidence weight of the resource flow data; the quantile interval width is the difference between the 90th quantile and the 10th quantile predicted by the artificial intelligence prediction module;

[0105] The mean-weighted average and uncertainty-weighted variance are used for calculation.

[0106]

[0107] The raw data and the resource flow data are integrated into a single resource flow vector. ;

[0108] In addition to the aggregate mean, it also transmits the uncertainty boundary, which is calculated using the following formula.

[0109]

[0110] Wherein, is the variance of the resource stream data calculated based on the quantile span. Final variance This will be retained in subsequent stages.

[0111] The database includes standardized environmental databases such as Ecoinvent and ReCiPe. Taking the Ecoinvent database as an example, it links power consumption with regional power grid datasets and powder usage with titanium powder manufacturing data.

[0112] In step S33, the ReCiPe midpoint method is used to call the characteristic factors in the corresponding process dataset. The environmental impact indicators for each single environment include the environmental impact indicators for electricity, powder, and gas.

[0113] In step S34, the life cycle assessment indicators include global warming potential (GWP), particulate matter formation potential (PMFP), human toxicity (HT), and water depletion potential (WDP). This step introduces incremental calculations, enabling low computational load for environmental assessment during the manufacturing process. Monte Carlo sampling or quantile propagation methods are used to... The uncertainty is transmitted to the environmental impact results.

[0114] In step S35, the life cycle assessment index is standardized to a part-level unit, such as kg·CO2e / piece.

[0115] The environmental impact calculation module is implemented using an embedded Python computing engine and runs on an internal SQL database containing the Ecoinvent characterization table. The transformation formula and characterization factor conform to the ISO 14040 / 14044 standard.

[0116] The feedback and visualization module includes a human-machine interface or sustainability dashboard to display real-time lifecycle assessment indicators and generate equipment parameter adjustment instructions or suggestions to reduce energy and material consumption during production. For example, if the global warming potential exceeds a preset threshold, the feedback and visualization module can suggest reducing beam power, optimizing toolpaths, and adjusting laser rates to reduce energy consumption and emissions.

[0117] The central control processor includes an edge controller and an industrial coordination gateway, which are used to manage data synchronization between modules, execute closed-loop control commands, and interface with the equipment controller to ensure deterministic operation and traceability.

[0118] During operation, the system utilizes an industrial Ethernet connection to successfully achieve stable data communication between the electron beam melting subsystem, the computer numerical control subsystem, and the laser polishing subsystem. The artificial intelligence prediction module continuously updates prediction results with an average refresh interval of 5 seconds; the environmental impact calculation module updates the total environmental impact index in real time, and the accuracy deviation compared with the OpenLCA calculation results after molding is ≤±10%. The dashboard interface displays the impact trend during the molding process in real time, allowing operators to observe instantaneous environmental load changes and identify high-consumption zones. When operators follow optimization suggestions (such as adjusting laser power or spindle speed), under the same production conditions, total energy consumption is reduced by approximately 12%–15% compared to baseline operation. In benchmark tests, closed-loop optimization reduced total energy consumption by 12%–15% and powder waste by 8%–10% compared to baseline production, verifying the quantitative advantages of the system in terms of sustainability.

[0119] Compared with the prior art, the beneficial effects of this embodiment are as follows:

[0120] Real-time environmental awareness: Environmental impact indicators are dynamically calculated during the production process (rather than after production ends), enabling sustainable control during the process;

[0121] Predictive resource management: Hybrid models predict future energy and material consumption and can identify high environmental impact conditions at an early stage;

[0122] Closed-loop optimization: The feedback and visualization module transforms sustainability data into actionable process adjustment plans, achieving measurable reductions in energy consumption and CO2 emissions;

[0123] Hybrid process integration: The system integrates electron beam melting, computer numerical control and laser polishing processes under a unified data framework to achieve full life cycle coverage;

[0124] Industrial scalability: The modular architecture supports integration with existing factory control networks through standardized communication protocols, ensuring broad applicability in innovative manufacturing environments.

[0125] Example 2

[0126] In this embodiment, the system described in Embodiment 1 is deployed using a distributed cloud-edge architecture to support enterprise-level applications across multiple hybrid manufacturing units. The system in this embodiment includes all the functional modules from Embodiment 1, including the data acquisition module, the artificial intelligence prediction module, the environmental impact calculation module, and the feedback and visualization module. These modules are coordinated by a central control processor. The main difference between the systems in Embodiment 1 and Embodiment 2 lies in the network topology and resource allocation. The artificial intelligence prediction module and the environmental impact calculation module are deployed on an industrial cloud platform to perform resource prediction and environmental impact calculations for electron beam melting equipment, computer numerical control equipment, and laser polishing equipment; the data acquisition module and the feedback and visualization module operate at the edge. The artificial intelligence prediction module and the environmental impact calculation module receive streaming data from multiple edge nodes, execute trained inference models, and return the predicted consumption curves to the edge gateway. The central control processor functions as both an edge gateway and a local coordinator, maintaining deterministic synchronization between the device controller and the cloud platform through industrial IoT protocols such as OPC-UA over TCP or MQTT with Quality of Service Level 2 (QoS 2). The central control processor uses locally cached AI predictions to perform real-time control loop adjustments, while the cloud platform handles long-term trend analysis and global model updates. This division of labor ensures sub-second feedback latency while retaining enterprise-level learning capabilities. The feedback and visualization module is deployed on the user interface or manufacturing execution terminal to display sustainability indicators in real time and issue control recommendations.

[0127] The specific process of the data synchronization mechanism is as follows: each device controller adds a timestamp to the sensor data packets using the standardized Network Time Protocol (NTP); the edge gateway integrates these data packets through the OPC-UA interface and achieves signal alignment within a ±20 millisecond tolerance window through interpolation; the generated synchronization dataset is stored as a time-series array for use by the artificial intelligence prediction module and the environmental impact calculation module. This deterministic synchronization process ensures the accuracy of lifecycle assessment indicators across different processes.

[0128] The workflow of this embodiment can be referred to Figure 1 .

[0129] In this embodiment, in step S1, the central control processor performs signal filtering on the raw data, compresses it into a standardized JSON payload, and transmits the processed raw data to the cloud platform through a secure channel for reception by the artificial intelligence prediction module and the environmental impact calculation module.

[0130] In this embodiment, in step S2, in order to improve the accuracy of prediction, a multi-layer architecture is adopted, including: (i) a data integrity pipeline for standardizing and denoising multi-source signals; (ii) manufacturing-oriented feature engineering; and (iii) a stacked model integration that includes calibration, uncertainty quantization and online drift processing.

[0131] The architecture and workflow of the artificial intelligence prediction module will be further explained below.

[0132] The data integrity pipeline is used to convert the synchronous time-series data (power, beam current, scanning speed, tool path length, gas / coolant flow rate, etc.), part metadata (alloy type, layer thickness), and environmental data of the electron beam melting equipment, the computer numerical control equipment, and the laser polishing equipment into clean, unit-uniform time-series data with quality tags. The workflow includes: time-aligning and resampling the raw data processed in step S1 (time step Δt = 1-5 seconds); using a Hampel filter (or Kalman smoothing) to denoise and suppress outliers; unit standardization (converting to kWh / piece, kg / piece, Nm³ / piece); marking missing values ​​and generating quality tags.

[0133] Feature engineering for manufacturing receives time-series data and outputs a multi-dimensional feature matrix aligned with the production cycle / layer / segment. X The specific feature types include: geometrically derived features (part volume, surface area / volume ratio, overhang ratio, support volume ratio); toolpath statistical features (total path length, average dwell time, duty cycle, fill density, number of layers); thermal load proxy features (∑(beam current × exposure time), linear energy density, interlayer cooling time); and equipment status coding (material type, scanning strategy, layer thickness, cavity temperature setting).

[0134] The aforementioned stacked model ensemble includes a dual-branch learner comprising a tabular branch (gradient boosting (XGBoost / LightGBM) for handling static / segmented features) and a temporal branch (1D-CNN or LSTM for capturing short-term sequence patterns in power / gas trajectories). The outputs of the dual branches are fed into a meta-learner (Ridge / LightGBM) to reduce model bias and variance. Based on Bayesian hyperparameter optimization (TPE / BOHB), parameters such as learning rate, tree depth, subsampling rate, CNN kernel size, and LSTM hidden units are optimized. Block-based time-series cross-validation (CV) is employed to ensure temporal order. The final output is the calibrated resource flow prediction values ​​for each prediction period. .

[0135] Further, calibration, uncertainty quantification, and physical constraints will be performed, namely: Calibration: Isotonic regression or Platt scaling will be used on the validation set to correct for systematic bias; Uncertainty quantification: Quantile regression (such as LightGBM quantiles) or Monte Carlo dropout will be used to output the quantile intervals of the prediction results. Physical constraints: Boundary conditions are enforced through rule sets, such as non-negative resource flow, unlimited equipment power, and quality balance of the recovery ratio. The final output includes constraint check values. This is for use by the environmental impact calculation module described later.

[0136] During online monitoring and adaptive processes, error tracking is performed, and the mean absolute percentage error (MAPE) / root mean square error (RMSE) is calculated on a rolling basis according to the process stage (electron beam melting, computer numerical control, laser polishing). Drift detection is performed: ADWIN or KS test is used to detect residual / feature drift; if the drift exceeds the threshold, warm-start refit, partial fitting (XGBoost), or short-term fine-tuning (LSTM) is triggered, using the latest batch of data. Finally, the updated model parameters are output; drift events and retraining audit logs are provided.

[0137] In step S3, the environmental impact calculation module will include resource flow data with uncertainty ( (Units: kWh, kg, Nm³), combined with characteristic factors (ReCiPe), through the formula Calculate and utilize Quantile propagation boundaries are used to obtain interval impact values, and finally, a real-time lifecycle assessment index with confidence intervals is output.

[0138] In step S4, the system also establishes a feedback loop between the cloud server and the edge devices to achieve continuous model updates, parameter optimization, and real-time visualization of sustainability indicators. In automatic mode, the edge controller modifies the settings of relevant devices.

[0139] All collected data is fed back to the cloud server for retraining and calibration of the hybrid model, ensuring continuous improvement in prediction accuracy. After each production cycle is completed, automatic retraining of the hybrid model is triggered (using validation sensor-influence data pairs stored in the cloud). Incremental learning adjusts the model weights without the need for full retraining, reducing computational load while improving prediction accuracy during continuous operation.

[0140] The system was validated in a production environment comprising two electron beam melting machines, one CNC workstation, and one laser polishing station. It maintained stable cloud-edge synchronization with an average transmission latency of 370 milliseconds. The energy consumption predictions generated by the AI ​​prediction module showed an average absolute error of <7% compared to post-processed measurements. Real-time environmental indicators were refreshed every 5 seconds. Closed-loop feedback achieved an 11% cumulative energy consumption reduction over continuous production cycles without sacrificing dimensional accuracy. During 72 hours of continuous operation, there were no data packet losses or timestamp mismatches. This architecture ensures scalability, cross-plant collaboration, and centralized management of prediction and environmental intelligence.

[0141] Compared with locally deployed systems, the beneficial effects of this embodiment are as follows:

[0142] Scalable multi-production line deployment: A single cloud server coordinates multiple hybrid manufacturing units to achieve unified environmental monitoring and centralized control;

[0143] Centralized model maintenance: The hybrid model and lifecycle assessment database are updated in the cloud and automatically synchronized to all edge nodes to ensure consistent calculation accuracy across all sites;

[0144] Reduce equipment load: The edge controller only handles data acquisition and preprocessing tasks, reducing the utilization rate of on-site processors and improving the overall stability of the system;

[0145] Seamless industrial integration: The architecture supports standardized data exchange with Manufacturing Execution Systems (MES) and Enterprise Resource Planning (ERP) systems, enabling automated sustainability reporting;

[0146] Secure and traceable data stream: All transmissions are encrypted and logged to ensure data authenticity and comply with industrial network security standards.

[0147] Although the present invention has been disclosed above by way of embodiments, it is not intended to limit the present invention. Anyone skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A life cycle assessment system, characterized by, The system comprises a data acquisition module, an artificial intelligence prediction module, an environmental impact calculation module, and a feedback and visualization module; the data acquisition module acquires original data in the operation of equipment; the artificial intelligence prediction module receives the original data and predicts resource flow data; the environmental impact calculation module receives the original data and the resource flow data and converts them into standardized life cycle assessment indicators; the feedback and visualization module displays the life cycle assessment indicators and generates parameter adjustment instructions; the environmental impact calculation module uses an uncertainty-oriented incremental environmental impact calculation method to convert the resource flow data and the original data into life cycle assessment indicators, including, According to the process beat setting time window length, the original data and the resource flow data obtained in the time window are aggregated to obtain a resource flow vector of a current time window And the data of the newly added time period is processed through the sliding time window mechanism. Modeling the resource flow vector in mean-variance or quantile form including 10%, 50%, 90% quantiles; The resource flow vector with uncertainty interval Matching with the process dataset in the database, using the ReCiPe midpoint method, extracting the characteristic factors in the corresponding process dataset ; According to the resource flow vector and the characteristic factor , the environmental impact index of each single class in a single time window is calculated by the following formula , wherein, is the environmental impact indicator of class i for the current time window, is the environmental impact indicator of class j is the resource flow vector of class is the resource flow vector of class j is the characteristicization factor of the environmental impact indicator of class i for the resource flow vector of class is the environmental impact indicator of class i for the previous time window, and , is the minimum environmental impact in the time window t; is the maximum environmental impact in the time window t; the environmental impact indicators of each single category calculated in each time window and its uncertainty are time accumulated to obtain the life cycle assessment indicator and output to the feedback and visualization module.

2. A life cycle assessment system as claimed in claim 1, wherein, The artificial intelligence prediction module uses a supervised learning regression model, combines a multi-layer perceptron and a Gaussian process regressor, and constructs a hybrid model; the environmental impact calculation module uses a multi-stage data processing algorithm and uses an embedded Python calculation engine; the environmental impact calculation module calls characteristic factors from an Ecoinvent or ReCiPe database.

3. A life cycle assessment system as claimed in claim 1, wherein, A central control processor is further included for managing data synchronization between modules, executing closed-loop control instructions, and interfacing with the controller of the equipment; the central control processor adds a timestamp to the original data packet using a standardized network time protocol.

4. A life cycle assessment system as claimed in claim 1, wherein, The artificial intelligence prediction module and the environmental impact calculation module are deployed on a cloud platform, and the data acquisition module and the feedback and visualization module run on an edge; the artificial intelligence prediction module uses a multi-layer architecture, including a data integrity pipeline for standardization and denoising of multi-source signals, manufacturing-oriented feature engineering, and integration of stacked models including calibration, uncertainty quantification, and online drift processing.

5. A life cycle assessment method characterized by, The life cycle assessment system of any one of claims 1-4 is executed, including the following steps, Step S1, starting a hybrid manufacturing process, using the data acquisition module to collect data and obtaining original data; Step S2, the artificial intelligence prediction module predicts based on the original data to obtain resource flow data; Step S3, the environmental impact calculation module converts the resource flow data and the original data into life cycle assessment indicators; Step S4, the feedback and visualization module receives the life cycle assessment indicators and transmits adjustment suggestions to the operator or the control system.

6. A life cycle assessment method as claimed in claim 5, characterised in that, In step S3, the original data in the tth time window is represented as: to measure the mean, to measure the variance of the sensor; The resource flow data in the tth time window is represented as: wherein, represents the 10th percentile of the resource flow data; represents the 50th percentile of the resource flow data; represents the 90th percentile of the resource flow data; A confidence weighted distribution fusion method is used to assign weights to the confidence levels of the resource flow data and the original data, wherein, is a confidence weight of the original data, is a confidence weight of the resource flow data; the quantile interval width is a difference between a 90% quantile and a 10% quantile predicted by the artificial intelligence prediction module. A mean weighted average and uncertainty weighted variance calculation is used, integrating the raw data and the resource flow data into a single resource flow vector ; In addition to the aggregated mean, the uncertainty boundary is also transmitted, and the following formula is used for calculation, where σ2is the variance of the resource stream data based on quantile span computation, and σ2finalis the final variance Will be retained throughout the subsequent stages.

7. A life cycle assessment method as claimed in claim 6, characterised in that, In step S2, the following steps are included, Step S21, pre-processing historical data and original data collected by sensors to obtain standard data; Step S22, combining a feature set, performing feature selection on the standard data to identify key features; Step S23, after hyperparameter tuning of the hybrid model, training is performed; Step S24, inputting the key features and the original data into the trained hybrid model for prediction to obtain resource flow data; Step S25, transmitting the resource flow data to the environmental impact calculation module for evaluation; Step S26, optimizing the model through model retraining.

8. A life cycle assessment method as claimed in claim 7, characterised in that, In step S21, the preprocessing includes standardization processing and interpolation; in step S22, in the feature selection process, based on correlation analysis and principal component analysis, key features related to energy consumption, powder mass flow and gas consumption are screened out; in step S23, the hyperparameter tuning adopts the grid search and cross-validation method, including learning rate, iteration number and layer number parameters; the gradient boosting and random forest ensemble learning algorithm are adopted to realize the training of the hybrid model; in step S24, the resource flow data includes energy consumption, powder mass flow and gas consumption.

9. A life cycle assessment method as claimed in claim 5, characterised in that, In step S2, the data integrity pipeline performs time alignment and resampling on the original data; Hanning filter or Kalman smoothing is used for denoising to suppress outliers; unit standardization; missing values are marked and quality labels are generated; The feature engineering receives time series data and outputs a multi-dimensional feature matrix X aligned with the production cycle / layer / segment, including feature types such as geometric derivative features, tool path statistical features, thermal load proxy features and device state encoding; The stacked model ensemble includes a double-branch learner, which includes a table branch and a time series branch, the table branch adopts gradient boosting, including XGBoost or LightGBM, to process static / segmented features, the time series branch adopts 1D-CNN or LSTM to capture short-term sequence patterns in power / gas trajectories; the outputs of the double-branch are input into a meta-learner to reduce model bias and variance; based on Bayesian hyperparameter optimization, TPE or BOHB is used to optimize learning rate, tree depth, subsampling rate, CNN kernel size, LSTM hidden unit parameters, and block time series cross-validation is used to ensure time sequence; Calibration, uncertainty quantification and physical constraints, Calibration: isotonic regression or Platt scaling on validation set to correct systematic bias; Uncertainty quantification: quantile regression or Monte Carlo dropout to output quantile intervals of predictions ; Physical constraints: boundary conditions are enforced through a rule set; In online monitoring and adaptive process, error tracking is performed, and mean absolute percentage error or root mean square error is calculated by process stage; drift detection is performed: ADWIN or KS test is used to detect residual / feature drift; If the drift exceeds the threshold, trigger warm start re-fitting, partial fitting or short-term fine-tuning using the latest batch data; In step S3, the environmental impact calculation module uses The quantile propagation boundary is obtained, and the interval impact value is obtained. Finally, the life cycle evaluation index with the confidence interval is output.

10. A computer program product, characterised in that, The computer program, when executed by a processor, implements the steps of a life cycle assessment system as claimed in any one of claims 1-4. The computer program, when executed by a processor, implements the steps of a life cycle assessment system as claimed in any one of claims 1-4.

Images