A Deep Learning-Based Method for Improving Energy Efficiency in Trailer Production Environments

Abstract

This invention discloses a deep learning-based method for improving energy efficiency in trailer production environments, comprising the following steps: S1, real-time collection of trailer production environment data; S2, preprocessing the data to generate a standardized dataset; S3, constructing and training an energy efficiency optimization control model; S4, inputting the real-time generated standardized dataset into the energy efficiency optimization control model to generate energy consumption, production efficiency, and environmental stability indicators; S5, using an improved RVEA optimization algorithm to optimize the generated energy consumption, production efficiency, and environmental stability indicators, generating optimal production strategy parameters; S6, deploying the optimal production strategy parameters to the actual production execution system. This invention combines deep learning with an improved RVEA optimization method to intelligently optimize trailer production energy efficiency, possessing advantages such as strong adaptability, fast optimization speed, and excellent overall performance.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to a method for improving energy efficiency in trailer production environments based on deep learning. Background Technology

[0002] With the development of intelligent manufacturing and industrial internet technologies, trailer production, as an important part of transportation equipment manufacturing, is facing increasingly higher requirements for production energy efficiency and environmental stability. In the traditional trailer production process, problems such as high energy consumption in the production environment, large fluctuations in production efficiency, and unstable control of environmental parameters are common. These have become important bottlenecks restricting trailer manufacturers from improving quality and efficiency, saving energy and reducing emissions, and achieving green manufacturing. At present, common energy efficiency improvement methods in the industry mainly include equipment energy efficiency monitoring, process optimization, and simple energy consumption statistical analysis. Some advanced manufacturing enterprises have introduced rule-based energy consumption prediction systems to conduct offline analysis and optimization of energy utilization. However, existing technologies still have obvious limitations in practical applications.

[0003] Existing trailer production energy efficiency management typically relies on fixed energy consumption monitoring equipment and static energy efficiency benchmarks, lacking systematic collection and deep integration of multi-source heterogeneous production environment data. Multi-dimensional information such as equipment operating status, environmental temperature and humidity changes, process execution sequence, production cycle, and personnel operation behavior involved in the production process are often recorded in isolation and fail to form a unified data flow. Therefore, existing systems cannot achieve real-time and comprehensive perception of the trailer production environment, nor can they make timely and accurate energy efficiency control decisions based on dynamic environmental changes. In addition, due to weak data quality control capabilities, production data often contains problems such as missing data, anomalies, time asynchrony, and spatial location disorder, which further weakens the reliability of existing energy efficiency management systems in supporting data-driven decision-making.

[0004] In terms of energy efficiency optimization strategies, traditional trailer manufacturing often adopts experience-driven rule engines or energy consumption modeling methods based on linear regression, failing to fully incorporate deep learning technology. Faced with the complex and ever-changing trailer production environment, traditional methods struggle to effectively capture nonlinear and multi-scale implicit patterns, resulting in low accuracy in energy consumption prediction, limited room for improving production efficiency, and weak environmental stability assurance capabilities. In particular, in terms of multi-objective optimization, existing systems typically use a simple weighted summation method to handle the three sub-objectives of energy consumption, efficiency, and stability, ignoring the inherent conflict between multiple objectives. The optimization results tend to favor a single objective, making it difficult to achieve optimal overall performance in the trailer production process.

[0005] In the model training and application phases, existing technologies generally lack an intent recognition mechanism, making it impossible to effectively detect and respond to potential abnormal behaviors, spoofing operations, or process drift during the production process. Due to the lack of an intent recognition module, existing control systems can only passively adapt to external changes and cannot actively identify and correct potential energy efficiency deviation trends, resulting in insufficient robustness and intelligence of optimization strategies. At the same time, in terms of dynamic control model design, existing systems often adopt a single-objective output structure, lacking simultaneous modeling and joint optimization of multiple indicators such as energy consumption, efficiency, and stability, further limiting the potential for improving production energy efficiency.

[0006] In the application of multi-objective optimization algorithms, although traditional evolutionary optimization methods can handle multi-objective optimization problems, in the trailer production environment, due to intense conflicts of objectives and frequent dynamic changes, traditional algorithms have slow convergence speed and weak search directionality, and are prone to getting trapped in local optima. As a result, the generated production strategy parameters cannot adapt to the changing needs of the production environment in real time. In addition, existing optimization algorithms usually have fixed reference vector configurations and fail to dynamically adjust the optimization reference direction according to changes in the production environment, which causes the optimization effect to lag or fail in practical applications. Summary of the Invention

[0007] One objective of this invention is to propose a deep learning-based method for improving energy efficiency in trailer production environments. This invention combines deep learning modeling with an improved RVEA optimization algorithm to address the problems of high energy consumption, large efficiency fluctuations, and difficult environmental control in trailer production environments. It proposes a method based on joint optimization of real-time multi-source data acquisition, intent recognition, and dynamic control. Through dynamic reference vector updates, guidance direction filtering, and adaptive adjustment of multi-objective weights, it achieves simultaneous optimization of energy efficiency, production efficiency, and environmental stability, possessing advantages such as strong adaptability, fast optimization speed, excellent overall performance, and high system robustness.

[0008] A method for improving energy efficiency in trailer production environments based on deep learning, according to an embodiment of the present invention, includes the following steps:

[0009] S1. Real-time collection of production environment data of trailers to form multi-source data streams;

[0010] S2. Preprocess the multi-source data streams to generate a normalized dataset;

[0011] S3. Construct an energy efficiency optimization control model and train the energy efficiency optimization control model using a standardized dataset;

[0012] S4. During the actual operation of trailer production, the standardized dataset generated in real time is input into the energy efficiency optimization control model to generate energy consumption, production efficiency and environmental stability indicators.

[0013] S5. Using an improved RVEA optimization algorithm, the generated energy consumption, production efficiency, and environmental stability indicators are optimized to generate the optimal production strategy parameters.

[0014] S6. Deploy the optimal production strategy parameters to the actual production execution system.

[0015] Optionally, the production environment data includes equipment energy consumption parameters, equipment operating status, ambient temperature and humidity, process sequence, production speed, and personnel activity level data.

[0016] Optionally, the preprocessing includes missing value imputation, outlier removal, time synchronization alignment, spatial location mapping, and data normalization.

[0017] Optionally, the energy efficiency optimization control model includes an intent recognition module and a dynamic control module.

[0018] Optionally, the intent recognition module can be trained adversarially by introducing sample pairs of fake production behavior and sample pairs of abnormal production behavior.

[0019] Optionally, the dynamic control module adopts a multi-objective output structure and introduces Pareto optimization to improve output energy consumption, production efficiency, and environmental stability.

[0020] Optionally, S3 specifically includes:

[0021] S31. Input the normalized dataset into the energy efficiency optimization control model, which includes an intent recognition module and a dynamic control module;

[0022] S32. Input the normalized dataset into the intent recognition module, which processes it layer by layer through a multilayer perceptron:

[0023] ;

[0024] in, Indicates the first Layer intends to encode vectors, Indicates the first Layer weight matrix, Indicates the first Layer intends to encode vectors, Indicates the first Layer bias term, This represents the batch normalization function. This represents the modified linear unit activation function;

[0025] S33. Collect real production behavior sample pairs, generate fake production behavior sample pairs by inputting Gaussian noise into the automatic encoder, and generate abnormal production behavior sample pairs by simulating abnormal behavior.

[0026] S34. When training the intent recognition module, a joint adversarial optimization strategy is adopted. The adversarial loss for each pair of real production, fake production, and abnormal production behavior samples is calculated separately and combined to form the total intent recognition loss.

[0027] ;

[0028] in, This indicates an intention to identify the total loss. Represents the mathematical expectation. This represents the sample distribution of actual production behavior. This represents the sample distribution of disguised production behavior. This represents the sample distribution of abnormal production behavior. This represents the probability output by the discriminator. and Let represent the loss weighting coefficients for samples of fake production behavior and samples of abnormal production behavior, respectively. This represents a sample pair representing real production behavior. This represents a sample pair indicating spoofed production behavior. This represents a sample pair indicating abnormal production behavior. Indicates an automatic encoder. Indicates Gaussian noise. Represents a normalized dataset. Labels indicating correct operating procedures during production. This represents production environment data simulating abnormal behavior. Behavioral labels that represent simulated abnormal behavior;

[0029] S35. The received intent encoding vector is split by the dynamic control module, outputting energy consumption, production efficiency, and environmental stability indicators. Sub-target losses are calculated for each sub-target loss using the cross-entropy loss function, and the sub-target losses are combined to generate a multi-target loss vector.

[0030] ;

[0031] in, Represents the multi-objective loss vector. This indicates the loss of energy-consuming sub-targets. This represents the loss of the sub-objective of production efficiency. This represents the loss of the environmental stability sub-objective. These represent the parameters of the energy efficiency optimization control model;

[0032] S36. Using Pareto optimization, the weight relationships are autonomously set according to the optimization preferences of the three objectives of energy consumption, production efficiency, and environmental stability. A reference vector set is set according to uniform sampling, where each element in the reference vector set represents a unit norm vector. For the parameters of the current energy efficiency optimization control model, the multi-objective loss vector and the Euclidean projection distance of each reference vector are calculated:

[0033] ;

[0034] in, This represents the Euclidean projected distance between the multi-target loss vector and the reference vector. Represents Euclidean norm operations. Represents the reference vector;

[0035] Select the reference vector corresponding to the smallest Euclidean projected distance, calculate the dynamic weights, and combine the sub-target losses based on the dynamic weights to form the Pareto loss:

[0036] ;

[0037] in, This represents the Pareto loss after dynamic weighted combination. Indicates the first Individual target loss, Indicates the first Dynamic weights of individual sub-objectives Represents the reference vector In the Dimensional component values, Represents the reference vector In the Dimensional component values;

[0038] S37. Combine the losses of the intent recognition module and the dynamic control module to generate a comprehensive loss function:

[0039] ;

[0040] in, Represents the comprehensive loss function. This represents a balance factor between the loss from intent recognition and the loss from the control module.

[0041] S38. The energy efficiency optimization control model is trained by minimizing the comprehensive loss function, and the Euclidean distance between the current intention encoding vector and the previous stage intention encoding vector is calculated in real time. When the Euclidean distance between the two stages of intention encoding exceeds the set threshold, the local Pareto front reconstruction is triggered, the intention encoding vector is updated, and the training continues with the new comprehensive loss function until the model converges.

[0042] Optionally, the intent recognition module uses a generative adversarial network, and the dynamic control module uses a multilayer perceptron.

[0043] Optionally, S38 specifically includes:

[0044] S381. Obtain the current intent encoding vector. After completing each training stage, extract the current output intent encoding vector of the energy efficiency optimization control model, and at the same time read the recorded intent encoding vector of the previous stage.

[0045] S382. Calculate the change in intent encoding based on the Euclidean distance between the current intent encoding vector and the intent encoding vector of the previous stage:

[0046] ;

[0047] in, The intended encoding of the change amount This represents the current intent encoding vector. This represents the encoding vector of the intent from the previous stage. Operations on Euclidean norms;

[0048] S383. Determine whether the change in intent encoding exceeds the preset threshold. If it does, proceed to the dynamic optimization process; otherwise, continue with the regular training steps.

[0049] S384. After triggering the dynamic optimization process, collect the local sub-target trajectory and extract the nearest... The multi-target loss vectors from each batch constitute a set of local sub-target trajectories;

[0050] S385. Perform local Pareto front reconstruction, apply the fast non-dominated sorting algorithm to the local sub-target trajectory set, and identify a new set of local Pareto optimal reference vectors.

[0051] S386. Select the reference vector with the smallest Euclidean projection distance of the multi-target loss vector in the current stage from the reference vector set, and recalculate the dynamic weights.

[0052] S387. Perform intention encoding feature subspace compression and collect the most recent... The intent encoding vectors of each batch constitute the intent encoding matrix, and singular value decomposition is performed on the intent encoding matrix:

[0053] ;

[0054] in, Intent encoding matrix, Describes a left singular vector matrix. Represents a singular value diagonal matrix. Describes a right singular vector matrix. Indicates the transpose operation;

[0055] Principal component directions are selected by ensuring that the cumulative energy coverage is not less than the cumulative energy coverage threshold, thus forming a compressed intent encoding vector;

[0056] S388. Based on the compressed intent encoding vector and the updated dynamic weights, recalculate the comprehensive loss function;

[0057] S389. Perform dynamic optimization training, with the goal of minimizing the new comprehensive loss function. Apply an adaptive optimizer to update the parameters of the energy efficiency optimization control model and continue training until the convergence condition is met.

[0058] Optionally, S5 specifically includes:

[0059] S51. Combine energy consumption, production efficiency, and environmental stability indicators to form a performance indicator vector:

[0060] ;

[0061] in, This represents a vector of performance metrics under the current production status. Indicates energy consumption index, Indicates a production efficiency indicator. Indicates environmental stability indicators;

[0062] S52. Based on the performance index vector, the weight relationship is set autonomously according to the optimization preferences of the three objectives of energy consumption, production efficiency and environmental stability, and the initial reference vector set is set according to uniform sampling.

[0063] S53. Calculate the Euclidean distance between the reference vectors and performance index vectors in each initial reference vector set;

[0064] S54. Based on the magnitude of the Euclidean distance, use fast non-dominated sorting to select the Pareto optimal reference set;

[0065] S55. For the Pareto-optimal reference vector set, calculate the guiding vectors in each direction based on the current performance index vectors:

[0066] ;

[0067] in, Indicates the first One directional guiding vector. Indicates the first The Pareto optimal reference vector;

[0068] S56. In each round of optimization iteration, new candidate solution parameters are generated based on the current reference vector, candidate performance index vectors are inferred, and the movement vector is calculated. The selection is guided by the cosine of the angle between the guiding vector and the movement vector.

[0069] ;

[0070] in, Represents the cosine value of the included angle. Represents a vector of candidate performance metrics. Represents the movement vector;

[0071] If the cosine of the included angle is close to 1, it means that the improvement direction of the candidate solution is consistent with the direction of the guiding vector, and it should be retained.

[0072] If the cosine of the included angle is close to -1, it means that the candidate solution deviates from the direction of the guiding vector, and the candidate solution is eliminated.

[0073] S57. An improved RVEA optimization algorithm is adopted. In each optimization iteration, the reference vector is dynamically updated based on the current performance index vector, and the guiding vector is also updated.

[0074] ;

[0075] in, Indicates the updated number Reference vectors, This represents the reference vector before the update. This represents the reference vector update rate factor. Represents the current performance metric vector;

[0076] S58. In each iteration, the weights of the sub-objectives are dynamically adjusted based on the current prediction error of each sub-objective:

[0077] ;

[0078] in, Indicates the first The weights of each sub-objective are dynamically adjusted. Indicates the original sub-objective weights. Indicates parameters Next Item performance index vector, Indicates the first Project reference value, Indicates the weighting stretching factor. This indicates taking the maximum value;

[0079] S59. By comprehensively applying dynamic updating of the reference vector and real-time adjustment of sub-objective weights, a weighted multi-objective error function is defined. The optimization objective is to minimize the weighted multi-objective error function, ultimately obtaining the optimal production strategy parameters:

[0080] ;

[0081] in, This represents the weighted multi-objective error function.

[0082] The beneficial effects of this invention are:

[0083] First, the present invention proposes a deep learning-based method for improving the energy efficiency of trailer production environments. By introducing real-time acquisition and standardized processing of multi-source production environment data streams, the present invention constructs a unified, standardized, and continuous production data foundation, improves the system's ability to perceive the complex and dynamic environment of the production site, provides high-quality data support for subsequent deep learning model training, and overcomes the problems of data bias and poor applicability when performing energy efficiency modeling based on a single data source in the past.

[0084] Secondly, this invention, by designing an energy efficiency optimization control model that includes an intent recognition module and a dynamic control module, overcomes the limitations of traditional trailer production energy efficiency optimization methods that can only passively adapt based on historical data. The intent recognition module adopts a joint adversarial training method by introducing samples of masquerading production behavior and samples of abnormal production behavior, which effectively improves the sensitivity to changes in production intent and the ability to detect abnormal behavior. This enables the system to identify potential energy efficiency deviation trends in real time, correct production strategies in a timely manner, and enhance the initiative and robustness of the energy efficiency optimization process. The dynamic control module adopts a multi-objective output structure and uses the Pareto optimization mechanism to simultaneously model and jointly optimize energy consumption, production efficiency, and environmental stability indicators. This overcomes the problems of simple weighting and uncontrollable objective conflicts in traditional multi-objective optimization methods, and improves the balance and optimality of the overall performance of the production process.

[0085] Furthermore, in the multi-objective optimization process, this invention introduces an improved RVEA optimization algorithm. Through reference vector generation, guiding direction selection, fast non-dominated sorting, and dynamic weight adaptive adjustment mechanism, a multi-objective optimization framework is constructed for energy consumption, production efficiency, and environmental stability indicators. In each optimization iteration, the reference vector is dynamically updated in real time according to the current production status, and the guiding vector is adjusted in real time based on changes in performance indicators. During the candidate solution selection process, the cosine of the angle between the guiding directions is used as the optimization criterion, thereby realizing rapid and accurate evolution towards the optimal solution in complex trailer production environments. This effectively improves the directionality and convergence speed of multi-objective search, avoids the limitation of traditional optimization algorithms being prone to getting trapped in local optima, and through the dynamic sub-objective weight adjustment mechanism, the system can adaptively adjust the optimization focus according to the sub-objective deviation at different stages, achieving a comprehensive performance improvement with the lowest energy consumption, highest efficiency, and most stable environment in the production process.

[0086] Finally, this invention designs a dynamic local Pareto front reconstruction mechanism and an intent encoding subspace compression mechanism during the intent recognition training stage. These mechanisms enable timely reconstruction of the optimization direction and feature representation when production intent changes drastically, ensuring the continuous adaptive capability of the energy efficiency optimization control model under changes in the actual production environment. This significantly improves the model's stability and intelligence. By automatically deploying the optimal production strategy parameters to the production execution system, this invention achieves closed-loop control of the entire process, including data acquisition, model inference, strategy optimization, and execution application. This significantly improves the energy efficiency, capacity utilization, and production environment safety of trailer manufacturers in actual operation. Attached Figure Description

[0087] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0088] Figure 1 This is a flowchart of a deep learning-based method for improving energy efficiency in trailer production environments, as proposed in this invention.

[0089] Figure 2 This is a block diagram of the overall structure of the energy efficiency optimization control model for a deep learning-based method for improving energy efficiency in trailer production environments proposed in this invention.

[0090] Figure 3 This is a flowchart of the improved RVEA optimization algorithm for a deep learning-based method to improve energy efficiency in trailer production environments, as proposed in this invention. Detailed Implementation

[0091] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0092] refer to Figure 1-3 A deep learning-based method for improving energy efficiency in trailer production environments includes the following steps:

[0093] S1. Real-time collection of production environment data of trailers to form multi-source data streams;

[0094] S2. Preprocess the multi-source data streams to generate a normalized dataset;

[0095] S3. Construct an energy efficiency optimization control model and train the energy efficiency optimization control model using a standardized dataset;

[0096] S4. During the actual operation of trailer production, the standardized dataset generated in real time is input into the energy efficiency optimization control model to generate energy consumption, production efficiency and environmental stability indicators.

[0097] S5. Using an improved RVEA optimization algorithm, the generated energy consumption, production efficiency, and environmental stability indicators are optimized to generate the optimal production strategy parameters.

[0098] S6. Deploy the optimal production strategy parameters to the actual production execution system.

[0099] This invention constructs a deep learning-based method for improving energy efficiency in trailer production environments. It enables real-time data acquisition, data preprocessing, training and application of energy efficiency optimization control models, and automatic deployment of optimal production strategies, forming a complete closed-loop system for energy efficiency improvement. Compared to traditional rule-based energy efficiency control methods, this invention can dynamically sense changes in the production environment, predict energy consumption, efficiency, and environmental indicators in real time, and generate optimal control strategies through intelligent optimization algorithms. This significantly improves energy utilization efficiency, production stability, and system intelligence in trailer production, demonstrating remarkable real-time performance, accuracy, and adaptive optimization capabilities.

[0100] In this embodiment, the production environment data includes equipment energy consumption parameters, equipment operating status, ambient temperature and humidity, process sequence, production speed, and personnel activity data.

[0101] This invention defines the scope of trailer production environment data, covering equipment energy consumption parameters, operating status, ambient temperature and humidity, process sequence, production speed, and personnel activity data. It enables comprehensive collection and modeling of multi-dimensional features of the production process. Compared with traditional data collection methods that only focus on equipment energy consumption or a single environmental parameter, this invention achieves real-time capture of information on the entire trailer production chain and all its states. This provides a rich, comprehensive, and high-quality data foundation for subsequent deep learning model training and optimization, significantly improving the accuracy and applicability of energy efficiency prediction and production optimization decisions.

[0102] In this embodiment, the preprocessing includes missing value imputation, outlier removal, time synchronization alignment, spatial location mapping, and data normalization.

[0103] In the preprocessing of multi-source data streams from trailer production environments, this invention employs various processing methods such as missing value imputation, outlier removal, time synchronization alignment, spatial location mapping, and data normalization. These methods effectively improve data quality and consistency. Compared to traditional simple screening or direct filtering methods, this invention ensures the integrity, accuracy, and comparability of the input data, providing more stable and reliable input conditions for energy efficiency optimization control models. This significantly improves the convergence speed, prediction accuracy, and robustness of subsequent model training.

[0104] In this embodiment, the energy efficiency optimization control model includes an intent recognition module and a dynamic control module.

[0105] This invention, by designing an energy efficiency optimization control model, organically combines an intent recognition module with a dynamic control module. In the process of improving the energy efficiency of trailer production, it achieves synergistic optimization of dynamic recognition of production intent and prediction of multi-objective performance indicators. Compared with the traditional control system that simply adjusts based on fixed parameters or empirical rules, this invention can flexibly adjust the control strategy based on real-time perception of changes in production intent, taking into account the requirements of energy consumption, production efficiency and environmental stability, and significantly enhancing the intelligence, adaptability and overall optimization level of the trailer production system.

[0106] In this embodiment, the intent recognition module is trained adversarially by introducing sample pairs of fake production behavior and sample pairs of abnormal production behavior.

[0107] This invention significantly improves the ability to identify potential anomalies and changes in intent in the production process by introducing adversarial training of sample pairs of spoofed production behavior and sample pairs of abnormal production behavior in the intent recognition module. Compared with traditional intent recognition methods based on single supervised learning, this invention can more accurately detect abnormal operations or spoofed behaviors in the production process, adjust energy efficiency optimization control strategies in a timely manner, enhance the anomaly perception and intelligent correction capabilities of the trailer production system in complex and ever-changing environments, and improve the robustness and safety of the overall system.

[0108] In this embodiment, the dynamic control module adopts a multi-objective output structure and introduces Pareto optimization to improve output energy consumption, production efficiency, and environmental stability.

[0109] This invention employs a multi-objective output structure in the dynamic control module and introduces the Pareto optimization mechanism, enabling the simultaneous output of three indicators: energy consumption, production efficiency, and environmental stability. This overcomes the limitation of traditional control models that only support single-objective optimization. By balancing the conflict relationships between multiple objectives, this invention achieves comprehensive optimization of optimal energy efficiency, maximum efficiency, and environmental stability during trailer production, significantly improving the overall operational performance of the production line. It possesses the advantages of strong multi-objective collaborative optimization capabilities and adaptability to complex actual production environments.

[0110] In this embodiment, S3 specifically includes:

[0111] S31. Input the normalized dataset into the energy efficiency optimization control model, which includes an intent recognition module and a dynamic control module;

[0112] S32. Input the normalized dataset into the intent recognition module, which processes it layer by layer through a multilayer perceptron:

[0113] ;

[0114] in, Indicates the first Layer intends to encode vectors, Indicates the first Layer weight matrix, Indicates the first Layer intends to encode vectors, Indicates the first Layer bias term, This represents the batch normalization function. This represents the modified linear unit activation function;

[0115] S33. Collect real production behavior sample pairs, generate fake production behavior sample pairs by inputting Gaussian noise into the automatic encoder, and generate abnormal production behavior sample pairs by simulating abnormal behavior.

[0116] S34. When training the intent recognition module, a joint adversarial optimization strategy is adopted. The adversarial loss for each pair of real production, fake production, and abnormal production behavior samples is calculated separately and combined to form the total intent recognition loss.

[0117] ;

[0118] in, This indicates an intention to identify the total loss. Represents the mathematical expectation. This represents the sample distribution of actual production behavior. This represents the sample distribution of disguised production behavior. This represents the sample distribution of abnormal production behavior. This represents the probability output by the discriminator. and Let represent the loss weighting coefficients for samples of fake production behavior and samples of abnormal production behavior, respectively. This represents a sample pair representing real production behavior. This represents a sample pair indicating spoofed production behavior. This represents a sample pair indicating abnormal production behavior. Indicates an automatic encoder. Indicates Gaussian noise. Represents a normalized dataset. Labels indicating correct operating procedures during production. This represents production environment data simulating abnormal behavior. Behavioral labels that represent simulated abnormal behavior;

[0119] S35. The received intent encoding vector is split by the dynamic control module, outputting energy consumption, production efficiency, and environmental stability indicators. Sub-target losses are calculated for each sub-target loss using the cross-entropy loss function, and the sub-target losses are combined to generate a multi-target loss vector.

[0120] ;

[0121] in, Represents the multi-objective loss vector. This indicates the loss of energy-consuming sub-targets. This represents the loss of the sub-objective of production efficiency. This represents the loss of the environmental stability sub-objective. These represent the parameters of the energy efficiency optimization control model;

[0122] S36. Using Pareto optimization, the weight relationships are autonomously set according to the optimization preferences of the three objectives of energy consumption, production efficiency, and environmental stability. A reference vector set is set according to uniform sampling, where each element in the reference vector set represents a unit norm vector. For the parameters of the current energy efficiency optimization control model, the multi-objective loss vector and the Euclidean projection distance of each reference vector are calculated:

[0123] ;

[0124] in, This represents the Euclidean projected distance between the multi-target loss vector and the reference vector. Represents Euclidean norm operations. Represents the reference vector;

[0125] Select the reference vector corresponding to the smallest Euclidean projected distance, calculate the dynamic weights, and combine the sub-target losses based on the dynamic weights to form the Pareto loss:

[0126] ;

[0127] in, This represents the Pareto loss after dynamic weighted combination. Indicates the first Individual target loss, Indicates the first Dynamic weights of individual sub-objectives Represents the reference vector In the Dimensional component values, Represents the reference vector In the Dimensional component values;

[0128] S37. Combine the losses of the intent recognition module and the dynamic control module to generate a comprehensive loss function:

[0129] ;

[0130] in, Represents the comprehensive loss function. This represents a balance factor between the loss from intent recognition and the loss from the control module.

[0131] S38. The energy efficiency optimization control model is trained by minimizing the comprehensive loss function, and the Euclidean distance between the current intention encoding vector and the previous stage intention encoding vector is calculated in real time. When the Euclidean distance between the two stages of intention encoding exceeds the set threshold, the local Pareto front reconstruction is triggered, the intention encoding vector is updated, and the training continues with the new comprehensive loss function until the model converges.

[0132] This invention constructs a standardized dataset and inputs it into an energy efficiency optimization control model. It then uses a multilayer perceptron to process the data, combines it with generative adversarial examples to train the intent recognition module, and utilizes Pareto optimization to output multi-objective performance indicators. This enables the systematic training of an energy efficiency optimization control model in a trailer production environment. Unlike traditional single-data-input and simple regression modeling, this invention improves the model's adaptability and prediction accuracy in a variable production environment by simulating and recognizing complex production behaviors. This significantly enhances the systematic nature, accuracy, and robustness of model optimization.

[0133] In this embodiment, the intent recognition module uses a generative adversarial network, and the dynamic control module uses a multilayer perceptron.

[0134] This invention fully leverages the advantages of deep learning in complex feature extraction and nonlinear modeling by introducing a generative adversarial network structure in the intent recognition module and employing a multilayer perceptron structure in the dynamic control module. Compared with traditional shallow networks or manual feature methods, this invention significantly improves the sensitivity of intent recognition and the accuracy of control decisions in trailer production processes, enhances the model's adaptability to environmental changes, and ensures the ability to continuously achieve energy efficiency optimization and multi-objective performance balance in complex trailer production environments.

[0135] In this embodiment, S38 specifically includes:

[0136] S381. Obtain the current intent encoding vector. After completing each training stage, extract the current output intent encoding vector of the energy efficiency optimization control model, and at the same time read the recorded intent encoding vector of the previous stage.

[0137] S382. Calculate the change in intent encoding based on the Euclidean distance between the current intent encoding vector and the intent encoding vector of the previous stage:

[0138] ;

[0139] in, The intended encoding of the change amount This represents the current intent encoding vector. This represents the encoding vector of the intent from the previous stage. Operations on Euclidean norms;

[0140] S383. Determine whether the change in intent encoding exceeds the preset threshold. If it does, proceed to the dynamic optimization process; otherwise, continue with the regular training steps.

[0141] S384. After triggering the dynamic optimization process, collect the local sub-target trajectory and extract the nearest... The multi-target loss vectors from each batch constitute a set of local sub-target trajectories;

[0142] S385. Perform local Pareto front reconstruction, apply the fast non-dominated sorting algorithm to the local sub-target trajectory set, and identify a new set of local Pareto optimal reference vectors.

[0143] S386. Select the reference vector with the smallest Euclidean projection distance of the multi-target loss vector in the current stage from the reference vector set, and recalculate the dynamic weights.

[0144] S387. Perform intention encoding feature subspace compression and collect the most recent... The intent encoding vectors of each batch constitute the intent encoding matrix, and singular value decomposition is performed on the intent encoding matrix:

[0145] ;

[0146] in, Intent encoding matrix, Describes a left singular vector matrix. Represents a singular value diagonal matrix. Describes a right singular vector matrix. Indicates the transpose operation;

[0147] Principal component directions are selected by ensuring that the cumulative energy coverage is not less than the cumulative energy coverage threshold, thus forming a compressed intent encoding vector;

[0148] S388. Based on the compressed intent encoding vector and the updated dynamic weights, recalculate the comprehensive loss function;

[0149] S389. Perform dynamic optimization training, with the goal of minimizing the new comprehensive loss function. Apply an adaptive optimizer to update the parameters of the energy efficiency optimization control model and continue training until the convergence condition is met.

[0150] In the training process of the intent recognition and control model, this invention designs a dynamic local Pareto front reconstruction and intent encoding feature subspace compression mechanism to address the situation where the change in intent encoding vector exceeds a threshold. By updating the optimization direction through rapid non-dominated sorting and extracting principal components based on singular value decomposition, this invention effectively addresses the risk of control failure caused by sudden changes in intent in the trailer production environment. Compared with traditional fixed-structure training methods, this invention significantly improves the adaptability and continuous optimization capability of the energy efficiency optimization control model under dynamic environmental changes, ensuring the stability and intelligence of the production process.

[0151] In this embodiment, S5 specifically includes:

[0152] S51. Combine energy consumption, production efficiency, and environmental stability indicators to form a performance indicator vector:

[0153] ;

[0154] in, This represents a vector of performance metrics under the current production status. Indicates energy consumption index, Indicates a production efficiency indicator. Indicates environmental stability indicators;

[0155] S52. Based on the performance index vector, the weight relationship is set autonomously according to the optimization preferences of the three objectives of energy consumption, production efficiency and environmental stability, and the initial reference vector set is set according to uniform sampling.

[0156] S53. Calculate the Euclidean distance between the reference vectors and performance index vectors in each initial reference vector set;

[0157] S54. Based on the magnitude of the Euclidean distance, use fast non-dominated sorting to select the Pareto optimal reference set;

[0158] S55. For the Pareto-optimal reference vector set, calculate the guiding vectors in each direction based on the current performance index vectors:

[0159] ;

[0160] in, Indicates the first One directional guiding vector. Indicates the first The Pareto optimal reference vector;

[0161] S56. In each round of optimization iteration, new candidate solution parameters are generated based on the current reference vector, candidate performance index vectors are inferred, and the movement vector is calculated. The selection is guided by the cosine of the angle between the guiding vector and the movement vector.

[0162] ;

[0163] in, Represents the cosine value of the included angle. Represents a vector of candidate performance metrics. Represents the movement vector;

[0164] If the cosine of the included angle is close to 1, it means that the improvement direction of the candidate solution is consistent with the direction of the guiding vector, and it should be retained.

[0165] If the cosine of the included angle is close to -1, it means that the candidate solution deviates from the direction of the guiding vector, and the candidate solution is eliminated.

[0166] S57. An improved RVEA optimization algorithm is adopted. In each optimization iteration, the reference vector is dynamically updated based on the current performance index vector, and the guiding vector is also updated.

[0167] ;

[0168] in, Indicates the updated number Reference vectors, This represents the reference vector before the update. This represents the reference vector update rate factor. Represents the current performance metric vector;

[0169] S58. In each iteration, the weights of the sub-objectives are dynamically adjusted based on the current prediction error of each sub-objective:

[0170] ;

[0171] in, Indicates the first The weights of each sub-objective are dynamically adjusted. Indicates the original sub-objective weights. Indicates parameters Next Item performance index vector, Indicates the first Project reference value, Indicates the weighting stretching factor. This indicates taking the maximum value;

[0172] S59. By comprehensively applying dynamic updating of the reference vector and real-time adjustment of sub-objective weights, a weighted multi-objective error function is defined. The optimization objective is to minimize the weighted multi-objective error function, ultimately obtaining the optimal production strategy parameters:

[0173] ;

[0174] in, This represents the weighted multi-objective error function.

[0175] In the process of generating optimal production strategy parameters, this invention employs an improved RVEA optimization algorithm, combined with a direction-guided selection, dynamic reference vector update, and sub-objective dynamic weight adjustment mechanism, to achieve simultaneous and efficient optimization of energy consumption, production efficiency, and environmental stability. Compared with traditional static multi-objective optimization methods, this invention can quickly respond to environmental changes during trailer production, adaptively adjust the optimization direction, avoid getting trapped in local optima, and effectively improve the optimization convergence speed, the overall performance of the control strategy, and the practical application effect.

[0176] Example 1:

[0177] To verify the feasibility of this invention in practical applications, it was applied to the production base of a trailer manufacturing company. This base has multiple production lines for trailer frames, bodies, painting, and assembly. The workshop is equipped with a large number of energy-intensive devices, such as welding robots, automated painting lines, and assembly robotic arms. It also faces practical problems such as intensive personnel work, frequent changes in environmental conditions, and complex process switching. High energy consumption, large fluctuations in production efficiency, and poor environmental stability are common problems in production. Traditional energy efficiency management methods are insufficient to meet the company's dual requirements of energy conservation and efficiency improvement.

[0178] To implement the method of this invention at this base, various data acquisition devices, such as environmental temperature and humidity sensors, equipment energy consumption acquisition modules, and personnel motion capture devices, are first deployed in the production environment to collect data in real time, including equipment operating status, energy consumption parameters, temperature and humidity changes, process execution sequence, and personnel operation behavior. After the preprocessing module performs missing value filling, outlier removal, time synchronization alignment, and spatial location mapping, and after standardization processing, a unified and standardized dataset is formed.

[0179] An energy efficiency optimization control model was trained based on a standardized dataset. The intent recognition module employs a generative adversarial network (GAN) structure, which performs joint adversarial training by inputting real production behavior samples, spoofed production behavior samples, and abnormal behavior samples to identify different production intentions. The dynamic control module uses a multilayer perceptron structure and outputs energy consumption, production efficiency, and environmental stability indicators. During model training, a Pareto optimization mechanism is introduced to balance the three performance indicators, and an improved RVEA optimization algorithm is used to generate the optimal production strategy.

[0180] In practical applications, the system generates new performance index vectors in real time with a sampling cycle of 5 minutes per day, and dynamically optimizes production control parameters based on real-time indicators. During the optimization phase, the energy efficiency optimization control strategy continuously adapts to changes in the production environment through directional guidance screening and dynamic reference vector updates. After the optimization strategies were deployed on the trailer frame welding line, final assembly line and painting line respectively, the overall energy consumption level decreased significantly, production efficiency improved, and the fluctuation range of environmental stability was significantly reduced.

[0181] To verify the effectiveness of the present invention, key indicators were monitored and statistically analyzed three months before and after application.

[0182] Table 1. Comparison of Experiments on Improving Trailer Production Energy Efficiency Based on Deep Learning and RVEA Optimization

[0183] ;

[0184] In terms of production efficiency, this invention reduces the average production cycle time of a single trailer on the trailer frame welding line from 85 minutes to 74 minutes, an improvement of 12.94%. Through intelligent adjustment of production rhythm and energy efficiency control strategies, this invention can optimize equipment load distribution and personnel operation coordination, improve overall production smoothness, significantly shorten the manufacturing cycle, and thus improve the overall operation efficiency of the production line.

[0185] Data from the trailer assembly line also validated the effectiveness of the invention. After implementing the invention, the energy consumption per trailer on the assembly line decreased from 38.2 kWh to 33.1 kWh, a reduction of 13.41%. By dynamically monitoring equipment operating status and optimizing work processes, the invention effectively reduces energy loss in the final assembly process, improves the energy efficiency of the production system, and fully demonstrates its excellent energy optimization capabilities across different process stages.

[0186] In terms of production cycle time, the production time per trailer assembly line was reduced from 72 minutes to 64 minutes, with an efficiency improvement of 11.27%. This invention, through real-time intent recognition and adaptive adjustment of control strategies, makes the operation process more coherent, reduces equipment idling and waiting time, and significantly improves overall production efficiency, further demonstrating the practical application value of this invention in complex production environments.

[0187] In terms of production environment control stability, the present invention has also achieved significant optimization results. The temperature fluctuation range of the trailer frame welding area has been reduced from ±4.7℃ to ±2.3℃, and the humidity fluctuation range has been reduced from ±6.2%RH to ±3.1%RH, which improves temperature and humidity stability by 51.06% and 50.00% respectively. By predicting environmental change trends in real time and adjusting control strategies, the present invention significantly enhances the stability of the production environment and creates a more reliable working environment for high-quality production.

[0188] The ambient temperature fluctuation in the trailer assembly area also decreased from ±4.5℃ to ±2.5℃, a reduction of 44.44%. This invention's real-time environmental stability control mechanism effectively suppresses drastic temperature fluctuations, improves personnel comfort and equipment operational stability, and provides a better environmental guarantee for subsequent painting and precision assembly processes.

[0189] In coating production lines, this invention brings a particularly significant improvement in environmental stability. The temperature fluctuation range of the production line is reduced from ±5.0℃ to ±1.8℃, the humidity fluctuation range is reduced from ±7.0%RH to ±2.8%RH, and the air particulate matter concentration is reduced from 0.12mg / m³ to 0.06mg / m³, with all indicators showing substantial improvement. Through dynamic sensing and adaptive environmental control, this invention can effectively suppress environmental disturbances generated during the production process, significantly improve coating quality and process stability, and reduce environmental pollution levels, demonstrating the important application value of this invention in green manufacturing.

[0190] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for improving energy efficiency in trailer production environments based on deep learning, characterized in that, Includes the following steps: S1. Real-time collection of production environment data of trailers to form multi-source data streams; S2. Preprocess the multi-source data streams to generate a normalized dataset; S3. Construct an energy efficiency optimization control model and train the energy efficiency optimization control model using a standardized dataset; S4. During the actual operation of trailer production, the standardized dataset generated in real time is input into the energy efficiency optimization control model to generate energy consumption, production efficiency and environmental stability indicators. S5. Using an improved RVEA optimization algorithm, the generated energy consumption, production efficiency, and environmental stability indicators are optimized to generate the optimal production strategy parameters. S6. Deploy the optimal production strategy parameters to the actual production execution system; S3 specifically includes: S31. Input the normalized dataset into the energy efficiency optimization control model, which includes an intent recognition module and a dynamic control module; S32. Input the normalized dataset into the intent recognition module, which processes it layer by layer through the multilayer perceptron. S33. Collect real production behavior sample pairs, generate fake production behavior sample pairs by inputting Gaussian noise into the automatic encoder, and generate abnormal production behavior sample pairs by simulating abnormal behavior. S34. When training the intent recognition module, a joint adversarial optimization strategy is adopted. The adversarial loss corresponding to the sample pairs of real production, fake production and abnormal production behavior is calculated respectively, and combined to form the total intent recognition loss. S35. Receive the intent encoding vector, split it through the dynamic control module, output energy consumption, production efficiency and environmental stability indicators, and calculate the sub-target loss for each. The sub-target loss uses the cross-entropy loss function, and the sub-target losses are combined to generate a multi-target loss vector. S36. Using Pareto optimization, the weight relationship is set autonomously according to the optimization preferences of the three objectives of energy consumption, production efficiency and environmental stability. A reference vector set is set according to uniform sampling. Each element in the reference vector set represents a unit norm vector. For the parameters of the current energy efficiency optimization control model, the multi-objective loss vector and the Euclidean projection distance of each reference vector are calculated. Select the reference vector corresponding to the smallest Euclidean projected distance, calculate the dynamic weights, and combine the sub-target losses based on the dynamic weights to form the Pareto loss: ； in, This represents the Pareto loss after dynamic weighted combination. Indicates the first Individual target loss, Indicates the first Dynamic weights of individual sub-objectives Represents the reference vector In the Dimensional component values, Represents the reference vector In the Dimensional component values; S37. Combine the losses of the intent recognition module and the dynamic control module to generate a comprehensive loss function; S38. The energy efficiency optimization control model is trained by minimizing the comprehensive loss function, and the Euclidean distance between the current intention encoding vector and the previous stage intention encoding vector is calculated in real time. When the Euclidean distance between the two stages of intention encoding exceeds the set threshold, the local Pareto front reconstruction is triggered, the intention encoding vector is updated, and the training continues with the new comprehensive loss function until the model converges. S5 specifically includes: S51. Combine energy consumption, production efficiency and environmental stability indicators to form a performance index vector; S52. Based on the performance index vector, the weight relationship is set autonomously according to the optimization preferences of the three objectives of energy consumption, production efficiency and environmental stability, and the initial reference vector set is set according to uniform sampling. S53. Calculate the Euclidean distance between the reference vectors and performance index vectors in each initial reference vector set; S54. Based on the magnitude of the Euclidean distance, use fast non-dominated sorting to select the Pareto optimal reference set; S55. For the Pareto-optimal reference vector set, calculate the guiding vectors in each direction based on the current performance index vectors: ； in, Indicates the first One directional guiding vector. Indicates the first The Pareto optimal reference vector; S56. In each round of optimization iteration, new candidate solution parameters are generated based on the current reference vector, candidate performance index vectors are inferred, and the movement vector is calculated. The selection is guided by the cosine of the angle between the guiding vector and the movement vector. ； in, Represents the cosine value of the included angle. Represents a vector of candidate performance metrics. Represents the movement vector; If the cosine of the included angle is close to 1, it means that the improvement direction of the candidate solution is consistent with the direction of the guiding vector, and it should be retained. If the cosine of the included angle is close to -1, it means that the candidate solution deviates from the direction of the guiding vector, and the candidate solution is eliminated. S57. An improved RVEA optimization algorithm is adopted. In each optimization iteration, the reference vector is dynamically updated based on the current performance index vector, and the guiding vector is also updated. ； in, Indicates the updated number Reference vectors, This represents the reference vector before the update. This represents the reference vector update rate factor. Represents the current performance metric vector; S58. In each iteration, the weights of the sub-objectives are dynamically adjusted based on the current prediction error of each sub-objective: ； in, Indicates the first The weights of each sub-objective are dynamically adjusted. Indicates the original sub-objective weights. Indicates parameters Next Item performance index vector, Indicates the first Project reference value, Indicates the weighting stretching factor. This indicates taking the maximum value; S59. By comprehensively applying the dynamic updating of the reference vector and the real-time adjustment of the sub-objective weights, a weighted multi-objective error function is defined. The optimization objective is to minimize the weighted multi-objective error function, and finally obtain the optimal production strategy parameters.

2. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, The production environment data includes equipment energy consumption parameters, equipment operating status, ambient temperature and humidity, process sequence, production speed, and personnel activity level data.

3. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, The preprocessing includes missing value imputation, outlier removal, time synchronization alignment, spatial location mapping, and data normalization.

4. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, The intent recognition module is trained adversarially by introducing sample pairs of fake production behavior and sample pairs of abnormal production behavior.

5. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, The dynamic control module adopts a multi-objective output structure and introduces Pareto optimization to improve output energy consumption, production efficiency and environmental stability.

6. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, The intent recognition module uses a generative adversarial network, and the dynamic control module uses a multilayer perceptron.

7. The method for improving energy efficiency in trailer production environments based on deep learning according to claim 1, characterized in that, S38 specifically includes: S381. Obtain the current intent encoding vector. After completing each training stage, extract the current output intent encoding vector of the energy efficiency optimization control model, and at the same time read the recorded intent encoding vector of the previous stage. S382. Calculate the change in intent encoding based on the Euclidean distance between the current intent encoding vector and the intent encoding vector of the previous stage; S383. Determine whether the change in intent encoding exceeds the preset threshold. If it does, proceed to the dynamic optimization process; otherwise, continue with the regular training steps. S384. After triggering the dynamic optimization process, collect the local sub-target trajectory and extract the nearest... The multi-target loss vectors from each batch constitute a set of local sub-target trajectories; S385. Perform local Pareto front reconstruction, apply the fast non-dominated sorting algorithm to the local sub-target trajectory set, and identify a new set of local Pareto optimal reference vectors. S386. Select the reference vector with the smallest Euclidean projection distance of the multi-target loss vector in the current stage from the reference vector set, and recalculate the dynamic weights. S387. Perform intention encoding feature subspace compression and collect the most recent... The intent encoding vectors of each batch constitute the intent encoding matrix, and singular value decomposition is performed on the intent encoding matrix. Principal component directions are selected by ensuring that the cumulative energy coverage is not less than the cumulative energy coverage threshold, thus forming a compressed intent encoding vector; S388. Based on the compressed intent encoding vector and the updated dynamic weights, recalculate the comprehensive loss function; S389. Perform dynamic optimization training, with the goal of minimizing the new comprehensive loss function. Apply an adaptive optimizer to update the parameters of the energy efficiency optimization control model and continue training until the convergence condition is met.

Images