Intelligent factory line body layout automatic generation method, device and equipment

Abstract

The invention discloses an intelligent factory line body layout automatic generation method, device and equipment, and relates to the technical field of artificial intelligence. The intelligent factory line body layout automatic generation method comprises the following steps: according to received production demand information, obtaining image data and text description information of a line body of the same type corresponding to the production demand information so as to construct a large-scale data set of the line body of the type; the text description information comprises production demand information; pre-training and reasoning the initial generation model according to the large-scale data set, so that the initial generation model learns a mapping relation from text description information to the line body planning image of the type, and a first generation model is obtained; performing reinforcement learning adjustment of manual feedback on the first generation model to obtain a second generation model; and inputting the production demand information into the second generation model, and outputting a line body planning image. According to the method, the workshop line body layout meeting the production requirements can be automatically generated, the planning efficiency of an intelligent factory is greatly improved, and the manual design cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device and equipment for automatically generating line layout of an intelligent factory. Background Art

[0002] With the continuous development of Industry 4.0 and smart manufacturing, smart factories have become a core trend in the future manufacturing industry. By integrating technologies such as the Internet of Things (IoT), big data, and artificial intelligence (AI), smart factories achieve automated and intelligent production processes, as well as high flexibility and efficiency. Production line automation is a crucial component of smart factories. A well-designed production line layout can significantly improve production efficiency, reduce operating costs, and enhance product quality. However, current production line designs still rely on engineers' experience and manual operations, resulting in low efficiency and difficulty in quickly responding to changes in production demand.

[0003] In related technologies, production line design methods mainly rely on the following steps: first, manually collect basic information such as the factory's production data and equipment layout; then, rely on engineers' experience and manual design tools to plan and optimize the production line; finally, apply the design results to the factory's actual production operations.

[0004] However, this approach faces many problems: on the one hand, the design process of the production line layout is complex and time-consuming, making it difficult to cope with frequent changes in demand and personalized customized production; on the other hand, the limitations of human experience make it difficult to further improve the efficiency and optimization effects of production line design. Summary of the Invention

[0005] The present application provides a method, device and equipment for automatically generating the line layout of an intelligent factory, which can solve the technical problems in the prior art of time-consuming and low-efficiency production line layout design.

[0006] In a first aspect, the present application provides a method for automatically generating a layout of a smart factory line, which comprises:

[0007] According to the received production demand information, image data and text description information of the same type of line corresponding to the above production demand information are obtained to construct a large-scale data set of the line type; the text description information includes the production demand information;

[0008] Based on the aforementioned large-scale dataset, the initial generative model is pre-trained and inferred to learn the mapping relationship from text description information to the line planning image of this type, thereby obtaining the first generative model.

[0009] Performing reinforcement learning adjustments using human feedback on the first generative model to obtain a second generative model;

[0010] The above production demand information is input into the above second generation model, and a line planning image is output.

[0011] In conjunction with the first aspect, in one embodiment, obtaining image data and text description information of the same type of line bodies corresponding to the above-mentioned production demand information specifically includes:

[0012] Obtain the smart factory corresponding to the above production demand information, and obtain the image data and text description information of the corresponding same type of line bodies in the smart factory to construct a multimodal dataset of the smart factory line bodies;

[0013] The image data and text description information of the same type of line bodies in the same industry of the above-mentioned smart factory are obtained to perform data enhancement and data storage on the above-mentioned multimodal dataset to obtain a large-scale dataset of this type of line bodies.

[0014] In conjunction with the first aspect, in one embodiment, pre-training the initial generative model based on the large-scale dataset specifically includes:

[0015] Using the text description information of the above large-scale dataset as input and the corresponding image data as output, an initial generative model is constructed;

[0016] Define a total loss function, and train the above initial generation model by minimizing the above total loss function to obtain a pre-trained model.

[0017] In combination with the first aspect, in one embodiment, the total loss function is a weighted sum of reconstruction loss, adversarial loss, and text alignment loss;

[0018] The above reconstruction loss is:

[0019] L r =Ε (T,I) [||IG(T;θ)|| 2 ];

[0020] The above adversarial loss is:

[0021] L a =Ε I [logD(I)]+E T [log(1-D(G(T)))]

[0022] The above text alignment loss is:

[0023] L t =-Ε (T,I) [cos(F(T),F(I))]

[0024] Among them, D is the discriminator, G(T;θ) is the initial generation model, θ is the model parameter, T is the text description information, I is the image data; F represents the embedding vector, and cos represents the cosine similarity of the embedding vector.

[0025] In conjunction with the first aspect, in one embodiment, after obtaining the pre-trained model, the method further includes reasoning the pre-trained model, specifically including:

[0026] Define the conditional probability distribution of forward diffusion, and forward diffuse the output image of the pre-trained model to gradually add noise to generate a noisy image;

[0027] Define the conditional probability distribution of reverse denoising and gradually denoise the above noisy image until a clear image is obtained;

[0028] The first generation model is obtained based on minimizing the difference between the noisy image and the clear image.

[0029] In combination with the first aspect, in one embodiment, the conditional probability distribution of the forward diffusion is:

[0030]

[0031] The conditional probability distribution of the above reverse denoising is:

[0032]

[0033] The loss function between the above noisy image and the clear image is:

[0034]

[0035] Among them, N represents Gaussian distribution, β t is the noise coefficient at time step t, I′ t is the noisy image at time step t, I′ t-1 is the noisy image at the t-1 time step, μ θ represents the mean function parameterized by the neural network, represents the variance of the noise.

[0036] In conjunction with the first aspect, in one embodiment, performing reinforcement learning adjustment of manual feedback on the first generation model specifically includes:

[0037] The reward model is trained using the image data and corresponding text description information output by the first generative model as input and the collected human feedback as output;

[0038] Based on the trained reward model, with the goal of maximizing the reward signal, the reinforcement learning method is used to optimize the above first generation model.

[0039] In conjunction with the first aspect, in one embodiment, the loss function when training the reward model is:

[0040] L R =Ε (T,I″) [||R φ (T,I″)-R h (T,I″)|| 2 ]

[0041] The above reward signal is:

[0042] L RL =Ε (T,I″) [R φ (T,I″)-λKL(G current ,G pretrained )]

[0043] Among them, R h (T,I″) is manual feedback, R φ (T,I″) is the reward model, I″ is the planning image generated by the first generative model, and λ is a hyperparameter.

[0044] In a second aspect, the present application provides a device for automatically generating a layout of a smart factory line, comprising:

[0045] A construction module is used to obtain image data and text description information of the same type of line bodies corresponding to the received production demand information, so as to construct a large-scale data set of the line bodies of the same type;

[0046] A model training module is used to pre-train and infer the initial generative model based on the above large-scale dataset, so that it learns the mapping relationship from text description information to the type of line planning image, and obtains the first generative model;

[0047] A human feedback reinforcement learning module, which is used to perform reinforcement learning adjustments based on human feedback on the first generative model to obtain a second generative model;

[0048] The automatic planning module is used to input the above production demand information into the above second generation model and output a line planning image.

[0049] In the third aspect, the present application provides an automatic generation device for the line layout of an intelligent factory, which includes a processor, a memory, and an automatic generation program for the line layout of an intelligent factory stored in the memory and executable by the processor. When the automatic generation program for the line layout of an intelligent factory is executed by the processor, the steps of the automatic generation method for the line layout of an intelligent factory are implemented.

[0050] The beneficial effects of the technical solution provided by this application include:

[0051] By receiving the production demand information from the user, the image data and text description information of the same type of line corresponding to the production demand information are obtained to construct a large-scale data set of this type of line, and then the initial generation model is pre-trained and inferred based on the large-scale data set to enable it to learn the mapping relationship from the text description information to the planning image of this type of line, to obtain a first generation model, and the first generation model is adjusted by reinforcement learning with manual feedback to obtain a second generation model, and then the production demand information is input into the second generation model to output the line planning image; effectively realize the automatic generation of workshop line layouts that meet production needs, greatly improve the planning efficiency of smart factories, reduce manual design costs, meet the requirements of flexible manufacturing and rapid response to market demand, and solve the technical problems of time-consuming and low efficiency in production line layout design in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of an embodiment of a method for automatically generating a layout of a smart factory line in the present application;

[0053] Figure 2 This is a flow chart of another embodiment of the method for automatically generating the layout of a smart factory line in the present application;

[0054] Figure 3 This is a schematic diagram of the functional modules of an embodiment of the device of the present application;

[0055] Figure 4 This is a timing diagram of each module in the embodiment of the present application;

[0056] Figure 5 This is a hardware structure diagram of the automatic generation equipment for the smart factory line layout involved in the embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0058] In a first aspect, an embodiment of the present application provides a method for automatically generating a line layout of a smart factory.

[0059] In one embodiment, referring to Figure 1 , Figure 1This is a flow chart of the first embodiment of the method for automatically generating a layout of a smart factory line in this application. The method includes:

[0060] S101. Based on the received user production demand information, obtain the image data and text description information of the same type of line corresponding to the above production demand information to construct a large-scale data set of the type of line; wherein the text description information includes the production demand information;

[0061] S102. Pre-training and reasoning the initial generative model based on the large-scale dataset to learn the mapping relationship from text description information to the type of line planning image, thereby obtaining a first generative model;

[0062] S103. Performing reinforcement learning adjustments using human feedback on the first generative model to obtain a second generative model;

[0063] S104. Input the above production demand information into the above second generation model and output the line planning image.

[0064] In this embodiment, the image data and text description information of the same type of line corresponding to the production demand information are obtained through the received user production demand information to construct a large-scale data set of this type of line, and then the initial generation model is pre-trained and inferred based on the large-scale data set to enable it to learn the mapping relationship from the text description information to the planning image of this type of line, to obtain a first generation model, and the first generation model is adjusted by reinforcement learning with manual feedback to obtain a second generation model, and then the production demand information is input into the second generation model to output the line planning image; effectively realize the automated generation of workshop line layouts that meet production needs, greatly improve the planning efficiency of smart factories, reduce manual design costs, meet the requirements of flexible manufacturing and rapid response to market demand, and solve the technical problems of time-consuming and low efficiency of production line layout design in related technologies.

[0065] Optionally, user production demand information includes line requirements and output requirements. After deploying the trained second generation model, when the user provides line requirements and output requirements, a layout plan for the smart factory line can be automatically generated, enabling rapid deployment of the smart factory.

[0066] Furthermore, in one embodiment, in the above step S1, obtaining the image data and text description information of the same type of line corresponding to the above production demand information specifically includes:

[0067] First, the smart factory corresponding to the above production demand information is obtained, and the image data and text description information of the corresponding same type of lines in the smart factory are obtained to construct a multimodal dataset of the lines in the smart factory.

[0068] Then, the image data and text description information of the same type of line bodies in the same industry of the above-mentioned smart factory are obtained to perform data enhancement and data storage on the above-mentioned multimodal dataset to obtain a large-scale dataset of this type of line bodies.

[0069] Preferably, the text description information includes text data and structured data.

[0070] In some embodiments, the introduction of artificial intelligence technologies, especially deep learning and reinforcement learning, provides new possibilities for production line automation design. Through training on large-scale data sets, AI models can learn patterns from historical production data and generate production line layout designs that adapt to different needs in a short period of time. However, most of the above studies focus on the application of single data sources, such as plain text descriptions and floor plans, which makes it difficult to fully utilize the rich multimodal data in smart factories (such as images, text, sensor data, etc.), thereby limiting the potential of AI models in production line design.

[0071] In this embodiment, the production line layout within a certain area specified by the user is obtained to construct a multimodal dataset of the smart factory line, specifically including:

[0072] A1. Obtain image data of smart factory production lines to ensure coverage of the production line layout for the same type of products, as well as large-scale text data and structured data covering the line attributes. Specifically, this includes "atomic-level" granular data of smart factory production lines, including the line's physical spatial dimensions, equipment attributes, equipment throughput per unit time, flexible manufacturing solutions, transport-type AGVs, manpower input, raw material input, energy input, etc.

[0073] A2. Preprocess the collected image data, text data, and structured data to ensure data quality and consistency. Specifically, the collected line image data is annotated in detail. Based on the annotations, the information data of the landmarks can be expanded, such as equipment type, production process, and operating steps.

[0074] A3. Align and match image data with corresponding text data and structured data to form an accurate multimodal dataset.

[0075] Furthermore, we will build a large-scale dataset at the pixel level for industrial lines, specifically including:

[0076] B1. Collect and integrate line images and corresponding text descriptions from the same industry;

[0077] B2. Preprocess the collected images and text descriptions to ensure their quality and consistency;

[0078] B3. Perform data enhancement on all images using methods such as rotation, scaling, cropping, and noise addition to enrich the diversity of image data and obtain image data;

[0079] B4. Enhance all text descriptions using natural language processing techniques such as synonym replacement, sentence transformation, and content expansion to obtain text description information;

[0080] B5. Perform multimodal alignment on the enhanced image data and text description information to ensure that each image accurately matches its corresponding text description, and store them to establish an efficient large-scale dataset that supports pixel-level data indexing and retrieval.

[0081] The above process may also include: Step B6. Verifying and quality controlling the constructed large-scale dataset to ensure that it meets the pixel-level accuracy and data integrity required for industrial applications.

[0082] In this embodiment, by collecting image and text description data from factories in the same industry, and unifying the format and standardizing the labels, we ensure that information such as images, equipment, and production capacity are consistent and easy to integrate; through image geometric transformation (such as rotation and scaling) and diversified text generation (such as rewriting descriptions or supplementing them with generative models), we expand the data set and simulate the production scenarios of the same type of products in different factories, so as to facilitate the pairing of enhanced image data and text description information to form a complete large-scale data set, ensuring that information such as images, production capacity, and layout descriptions are consistently associated.

[0083] Based on the above embodiment, in this embodiment, in the above step S2, the initial generation model is pre-trained according to the above large-scale data set, specifically including:

[0084] First, an initial generation model is constructed using the text description information of the above large-scale dataset as input and the corresponding image data as output.

[0085] Then, a total loss function is defined, and the initial generation model is trained by minimizing the total loss function to obtain a pre-trained model.

[0086] Furthermore, in this embodiment, the above-mentioned total loss function is a weighted sum of reconstruction loss, adversarial loss, and text alignment loss.

[0087] The above reconstruction loss is:

[0088] L r =Ε (T,I) [||IG(T;θ)|| 2 ];

[0089] The above adversarial loss is:

[0090] L a=Ε I [logD(I)]+E T [log(1-D(G(T)))]

[0091] The above text alignment loss is:

[0092] L t =-Ε (T,I) [cos(F(T),F(I))]

[0093] Among them, D is the discriminator, G(T;θ) is the initial generation model, θ is the model parameter, T is the text description information, I is the image data; F represents the embedding vector, and cos represents the cosine similarity of the embedding vector.

[0094] Furthermore, in this embodiment, after obtaining the pre-trained model, the pre-trained model is further reasoned. The reasoning of the pre-trained model specifically includes:

[0095] First, the conditional probability distribution of forward diffusion is defined, and the output image of the pre-trained model is forward diffused to gradually add noise to generate a noisy image.

[0096] Then, the conditional probability distribution of reverse denoising is defined, and the above noisy image is gradually denoised until a clear image is obtained;

[0097] Finally, based on minimizing the difference between the above-mentioned noisy image and the clear image, the above-mentioned first generation model is obtained.

[0098] In this embodiment, the pre-trained model is inferred through a diffusion model, which is a latent variable model consisting of two processes, including a forward diffusion process and a backward denoising process.

[0099] Preferably, in this embodiment, the conditional probability distribution of the forward diffusion is:

[0100]

[0101] The conditional probability distribution of the above reverse denoising is:

[0102]

[0103] The loss function between the above noisy image and the clear image is:

[0104]

[0105] Among them, N represents Gaussian distribution, β t is the noise coefficient at time step t, I t ′ is the noised image at time step t, I′ t-1is the noisy image at the t-1 time step, μ θ represents the mean function parameterized by the neural network, represents the variance of the noise.

[0106] Preferably, pre-training and reasoning are performed on the initial generative model, specifically including:

[0107] First, we build and configure the Stable Diffusion model, taking the text description information as the conditional input.

[0108] Secondly, we use text-image pairs from a multimodal dataset to pre-train the model, enabling it to learn the mapping relationship between text and image in the conditional diffusion process.

[0109] Then, by adjusting parameters such as the number of diffusion steps and noise level, the model performance is optimized to improve the quality and consistency of the generated images;

[0110] Finally, after completing pre-training, model inference is performed, text description information is input, and the corresponding line planning image is generated.

[0111] In this embodiment, a generative model (such as a GAN or Transformer model) is pre-trained using a large-scale dataset, so that the model learns the correlation features between image data and text description information during the pre-training process, and identifies the mapping relationship between equipment layout, production capacity requirements and space planning; wherein, the text description information is encoded into a feature vector, including production capacity requirements, equipment type and layout preferences, etc.; at the same time, the image data is encoded into pixel-level features; by aligning these multimodal features, the model can understand the generation logic from text to layout; then, in the inference stage, the image generated by the model is optimized again to generate a factory line layout diagram that meets the requirements, ensuring the rationality of the generation results in terms of equipment configuration, spatial distribution and production capacity output.

[0112] Furthermore, in one embodiment, in step S3, the first generation model is adjusted through reinforcement learning with human feedback, specifically including:

[0113] First, the reward model is trained using the image data and corresponding text description information output by the first generation model as input and the collected manual feedback as output.

[0114] Then, based on the trained reward model, the reinforcement learning method is used to optimize the first generation model with the goal of maximizing the reward signal.

[0115] In this embodiment, the RLHF scheme is used to align the output results of the artificial experience and the generative artificial intelligence model, so that the output line image is safe and controllable.

[0116] Furthermore, in one embodiment, the first generative model is adjusted through reinforcement learning with human feedback, specifically including:

[0117] C1. Collect manual feedback and manually evaluate the generated line images, annotating them for safety, accuracy, and compliance.

[0118] C2. Build a reward model, specifically a neural network or machine learning model that outputs a reward value or score based on the input line image data and text description information. This model converts human feedback into a reward signal, guiding model optimization and ultimately resulting in a trained reward model.

[0119] C3. Fine-tune the trained reward model and use deep reinforcement learning algorithms (such as DQN, DDPG, PPO, A3C, SAC, etc.) to train the first generative model to make it more consistent with human expectations;

[0120] C4. Integrate security policies and embed security control mechanisms into the first generation model to prevent the generation of non-compliant or harmful online content, thereby obtaining a second generation model.

[0121] The above process also includes: C5. Continuous monitoring and iteration, regularly evaluating the output of the second generation model, collecting new feedback, and continuously optimizing model performance to ensure that the output line body images are always consistent with human experience and are safe and controllable.

[0122] In this embodiment, when performing reinforcement learning adjustments with human feedback, evaluation criteria can be designed, inviting domain experts to score or evaluate the generated line layout diagrams, focusing on key indicators such as layout rationality, safety, and capacity matching, and providing improvement suggestions. Based on human feedback, a reward model is then constructed to quantify the quality of the generated results. The generated layout images receive different reward values ​​based on the scoring results to guide the model to optimize generation quality. The reward model is integrated into the training of the generation model, and the model parameters are adjusted through the RLHF method so that the generated line layouts gradually meet human experience and safety control requirements, ultimately achieving controllable and high-quality layout image output.

[0123] Furthermore, in this embodiment, the loss function when training the reward model is:

[0124] L R =Ε (T,I″) [||R φ (T,I″)-R h (T,I″)|| 2 ]

[0125] The above reward signal is:

[0126] L RL =Ε(T,I″) [R φ (T,I″)-λKL(G current ,G pretrained )]

[0127] Among them, R h (T,I″) is manual feedback, R φ (T,I″) is the reward model, I″ is the planning image generated by the first generative model, and λ is a hyperparameter.

[0128] Furthermore, the production demand information is input into the second generation model to output a line planning image, specifically including:

[0129] D1. Integrate the trained second generation model into the smart factory system;

[0130] D2. The second generation model receives and analyzes the production demand information input by the user, including line requirements and output requirements;

[0131] D3. The second generation model automatically generates a line layout plan that meets the requirements based on the analyzed demand parameters;

[0132] D4. Verify and optimize the generated layout to ensure feasibility and efficiency; provide the layout planning results to users for review and feedback;

[0133] D5. Based on user feedback, the model is iteratively adjusted to improve the layout design;

[0134] D6. Finalize and output the optimized line layout plan to support the rapid deployment of smart factories.

[0135] In this example, the user enters line requirements and output requirements. The smart factory system parses these requirements to derive core parameters, including spatial constraints, equipment layout requirements, and production targets. These parsed requirements are then fed into a second generation model. Based on the learned text-to-image mapping, the model generates a factory layout that meets the requirements—a line planning image—including details such as equipment locations, logistics channels, and operational space. The system then verifies the generated line planning image to ensure capacity matching and safety compliance. If any deviations are found, the system automatically adjusts the layout until it meets the requirements and presents the final result to the user.

[0136] Based on the above embodiment, in this embodiment, the above method specifically includes:

[0137] E1: Acquire image data of smart factory production lines and corresponding large-scale text data describing the line attributes. Collect structured data of factory production lines and construct a multimodal dataset of smart factory production lines. This includes:

[0138] E11: Select a type of factory based on user input;

[0139] E12: Obtain image data, text data, and structured data of similar lines in the factory. For example, collect multimodal datasets such as digital twins, images (jpg, png), text descriptions (word, txt), and structured data (excel) of similar lines in the factory's history.

[0140] E2: Based on the line image data and large-scale related data of descriptive texts in the same industry, namely text description information, data enhancement and data storage are performed on the multimodal dataset of smart factory production lines to construct a large-scale dataset at the pixel level of industrial lines.

[0141] This involves collecting relevant image data from factories in the same industry and performing geometric transformations, color conversions, image cropping, blending, and noise injection. We also collect text data from factories in the same industry and manually add, delete, and modify it. We also collect structured data from factories in the same industry and manually edit and supplement it. All of this data is aggregated to form a large-scale dataset for industrial lines.

[0142] E3: Pre-training and inference of AI generative models on the aforementioned large-scale datasets to learn the mapping relationship between text description information and planning images of factory workshop lines. Specifically, this includes:

[0143] E31: Input and output modeling, assuming that the input text description information is T and the output planning image is I. The goal is to learn the generative mapping G(T) from T to I, the formula is:

[0144] I=G(T;θ)

[0145] Where: G is the initial generative model (such as Stable Diffusion, GAN, VAE or Transformer, etc.), θ is the parameter of the model;

[0146] E32: In the pre-training stage, in order to make the generated image I′ close to the real image I, the total loss function L is defined as the weighted sum of reconstruction loss, adversarial loss, and text alignment loss. L is:

[0147] L=L r +λ1L a +λ2L t

[0148] Among them, λ1 and λ2 represent the weight coefficients of different losses respectively;

[0149] The reconstruction loss is used to constrain the similarity between the generated image and the real image in pixel space or feature space, L r for:

[0150] L r =Ε (T,I) [||IG(T;θ)|| 2 ]

[0151] Adversarial loss L a for:

[0152] L a =Ε I [logD(I)]+E T [log(1-D(G(T)))]

[0153] Among them, D is the discriminator, which distinguishes real images from generated images.

[0154] Text alignment loss L t for:

[0155] L t =-Ε (T,I) [cos(F(T),F(I))]

[0156] Among them, E represents averaging, F represents embedding, and cos represents the cosine similarity of the embedding vector.

[0157] By optimizing the total loss function L, the model parameters θ are updated:

[0158] θ * =argmin(L)

[0159] At this time, the input text description T is used to generate the pre-trained generative model G(T; θ * ) generates the corresponding initial planning image I′ of the factory workshop: I′=G(T;θ * ).

[0160] In the inference stage, the diffusion model Stable Diffusion is used to generate images. The generation process can be done through the following steps:

[0161] Forward Diffusion: Add noise to the initial planning image I′ to generate noise images I at multiple time steps t ′:

[0162]

[0163] Among them, q is the conditional probability distribution of forward diffusion, N is the Gaussian distribution, β t is the noise coefficient at time step t;

[0164] Reverse Generation: restore the clear image through conditional probability:

[0165]

[0166] Among them, μ θ represents the mean function parameterized by the neural network, Represents the variance of the noise and is usually fixed.

[0167] Model learning is achieved through the following loss function L d optimization:

[0168]

[0169] Finally, the image I″ of the production line layout is generated: I″=p θ (I′|T).

[0170] E4: Aligning human experience with generative AI model outputs through a human feedback-based reinforcement learning (RLHF) approach, specifically:

[0171] Design and train the reward model, input the final generated image I″ and the corresponding text description T. Use the feedback R provided by the expert h (T,I″) train reward model R φ (T,I″), the loss function is as follows:

[0172] L R =Ε (T,I″) [||R φ (T,I″)-R h (T,I″)|| 2 ]

[0173] Reward Model R φ Used to evaluate the safety, rationality, and human preference of generated images. Reinforcement learning can be used to optimize the generation model. The reinforcement learning method (PPO, Proximal Policy Optimization) is used to optimize the generation model with the goal of maximizing the reward signal:

[0174] L RL =Ε (T,I″) [R φ (T,I″)-λKL(G current ,G pretrained )]

[0175] Among them, the KL term limits the output distribution of the model to not deviate from the pre-training model G pretrained , that is, the first generation model, to ensure stability; G current is the optimized second generative model; λ is a hyperparameter used to balance the importance of the reward term and the KL divergence term.

[0176] E5: Deploy the trained second generation model. When the user provides line requirements and output requirements, the layout plan of the smart factory line is automatically generated to achieve rapid deployment of the smart factory.

[0177] The system analyzes user requirements (such as process type, equipment quantity, space constraints, and production targets) and converts them into standardized inputs. Using a second generative model, it generates an optimal layout solution that meets the multi-objective requirements of space utilization, safety, and production efficiency. The generated results can be directly output as standardized CAD files or used in digital twin simulations, automating the entire process from planning to deployment and accelerating the construction of smart factories.

[0178] The method of this embodiment makes full use of multimodal datasets, and automatically generates intelligent factory line planning diagrams through generative artificial intelligence models combined with reinforcement learning. During the pre-training stage of the generative model, the model parameters are optimized through reconstruction loss, adversarial loss, and text alignment loss, so that the model can generate high-quality factory line planning images from text descriptions, and the diffusion model is used to further improve the image generation effect. Combined with the RLHF scheme, a reward model is designed to evaluate the security, rationality and user preference of the generated images. The output of the generative model is optimized through reinforcement learning to ensure that the generated planning images are safe and controllable. Rapid deployment is achieved to meet the multi-objective requirements of production efficiency, safety and layout rationality.

[0179] Furthermore, if Figure 2 As shown, the method of this embodiment specifically includes:

[0180] S201: Receive production information input by the user, including the number of factory lines, total product output, etc.;

[0181] S202: The central server of the smart factory collects line image data, text data, and structured data from the current smart factory historical data based on user input; manually annotates the image data, text data, and structured data to make each type of data correspond to the internal logic; and simultaneously collects and pre-processes the image data and text description information of the same type of lines in the same industry, performs data enhancement, and stores the data, thereby forming a large-scale dataset of the smart factory lines;

[0182] S203: Determine the quality of the data set according to the data standard;

[0183] First, define the data quality standards. Based on industrial specifications, industry needs, and specific project goals, clarify the quality requirements of the multimodal dataset, including image resolution, completeness of text descriptions, accuracy of structured data, etc. For example, image resolution must reach a certain pixel threshold, and text descriptions must cover key attribute fields. Secondly, perform quality judgment and scoring. For image data: use image quality assessment algorithms, such as calculating the clarity, contrast, and color saturation of the image, to determine whether it meets the standards. For text data: use natural language processing (NLP) technology to detect the semantic integrity, keyword coverage, and language consistency of the text. For structured data: verify the range of numerical fields, the standardization of the format, and the logical consistency between fields. If the dataset quality is judged to be unqualified, proceed to S204. If the quality is judged to be qualified, proceed to S205.

[0184] S204: Data cleaning is performed to obtain a high-quality multimodal dataset. This involves removing blurry or low-resolution images, repairing image defects, cleaning redundant or noisy text, completing key descriptions, and addressing missing, outlier, and duplicate values ​​in structured data to ensure that field content is standardized and logically consistent. Furthermore, cross-modal consistency checks are performed to ensure that information across modalities is correctly mapped and unified. The process then proceeds to S205.

[0185] S205: Conduct multimodal large-scale model training to learn the logical relationship between text descriptions and line images, achieving a precise mapping from text input to image generation. By extracting features from multimodal data and modeling associations, the model grasps the deep correspondence between descriptive language and actual layout features.

[0186] S206: Evaluate the rationality, security, and user preference of the output by designing a reward model, and use reinforcement learning to optimize the generation process to ensure the safety, controllability, and high quality of line images.

[0187] S207: A comprehensive assessment is conducted across multiple aspects, including demand matching, layout rationality, and expert evaluation, to check whether the model-generated layout meets user requirements, including production processes, equipment quantity, space constraints, and output targets. The assessment also ensures high space utilization, rational material flow paths, optimized production efficiency, and compliance with industrial safety standards. If satisfied, the process proceeds to S208; otherwise, the process proceeds to S205.

[0188] Optionally, the layout solution can be further optimized by combining the evaluation of domain experts and user feedback to ensure that the generated results have practical usability and high reliability.

[0189] S208: By deploying a multimodal large model, the second generative model, the factory line layout can be automatically generated based on user-entered production requirements (such as line requirements, output requirements, process flow, equipment configuration, and space constraints). This input information is efficiently analyzed to generate an optimized layout solution that meets user needs, ensuring space utilization, safety, and production efficiency, enabling rapid planning and deployment of smart factories.

[0190] The method for automatically generating gland layouts in smart factories based on AIGC (artificial intelligence generated content) proposed in this embodiment integrates large-scale multimodal data, deep learning, generative models, and reinforcement learning techniques to address the problems of low efficiency and slow response in current production line design. It has the following advantages:

[0191] (1) High degree of automation and efficiency: AIGC technology automatically generates the line layout of the smart factory. Users only need to provide production demand information to quickly obtain a planning solution, which greatly improves layout efficiency and shortens the deployment cycle. It is particularly suitable for flexible manufacturing and rapid response to market demand environments;

[0192] (2) Controllability and security: The RLHF scheme is introduced to ensure that the generated layout plan not only meets the production demand information, but also has a high degree of security and controllability, reducing the risk of post-deployment adjustments;

[0193] (3) Reduce manual design costs: Utilize large-scale multimodal data sets and automated design processes to reduce reliance on manual planning and reduce labor costs, while improving the accuracy of layout design and adapting to changing production demand information.

[0194] In a second aspect, an embodiment of the present application also provides an automatic generation device for the line layout of an intelligent factory.

[0195] In one embodiment, the above-mentioned intelligent factory line layout automatic generation device includes a construction module, a model training module, a human feedback reinforcement learning module and an automatic planning module.

[0196] The construction module is used to obtain image data and text description information of the same type of line corresponding to the production demand information according to the received production demand information, so as to construct a large-scale data set of the line of the same type;

[0197] The model training module is used to pre-train and infer the initial generative model based on the large-scale dataset, so that it learns the mapping relationship from text description information to the type of line planning image, thereby obtaining a first generative model;

[0198] The human feedback reinforcement learning module is used to perform reinforcement learning adjustments based on human feedback on the first generative model to obtain a second generative model;

[0199] The automatic planning module is used to input the production demand information into the second generation model and output a line planning image.

[0200] The device of this embodiment pre-trains and infers generative models on large-scale data sets to learn the relationship from text descriptions to factory production line image generation. By introducing human experience and feedback, it generates safe and controllable line image output results. After the user provides requirements, the line layout and planning of the smart factory can be automatically generated, thereby realizing the rapid construction of the smart factory.

[0201] Preferably, the above-mentioned building module includes a data acquisition submodule and a data enhancement and storage submodule.

[0202] The above-mentioned data acquisition submodule is used to obtain the smart factory corresponding to the above-mentioned production demand information, and obtain the image data and text description information of the corresponding same type of line bodies in the smart factory to construct a multimodal dataset of the smart factory line bodies.

[0203] The data enhancement and storage submodule is used to obtain image data and text description information of the same type of line bodies in the same industry of the above-mentioned smart factory, so as to perform data enhancement and data storage on the above-mentioned multimodal dataset to obtain a large-scale dataset of this type of line bodies.

[0204] Furthermore, in one embodiment, the model training module is used to:

[0205] Using the text description information of the above large-scale dataset as input and the corresponding image data as output, an initial generative model is constructed;

[0206] Define a total loss function, and train the above initial generation model by minimizing the above total loss function to obtain a pre-trained model.

[0207] Furthermore, in one embodiment, the model training module is further used to:

[0208] Define the conditional probability distribution of forward diffusion, and forward diffuse the output image of the pre-trained model to gradually add noise to generate a noisy image;

[0209] Define the conditional probability distribution of reverse denoising and gradually denoise the above noisy image until a clear image is obtained;

[0210] The first generation model is obtained based on minimizing the difference between the noisy image and the clear image.

[0211] Furthermore, in one embodiment, the human feedback reinforcement learning module is used to:

[0212] The reward model is trained using the image data and corresponding text description information output by the first generative model as input and the collected human feedback as output;

[0213] Based on the trained reward model, with the goal of maximizing the reward signal, the reinforcement learning method is used to optimize the above first generation model.

[0214] Furthermore, if Figure 3 As shown, the above-mentioned device includes a factory module 301, a construction module, a model training module 304, a human feedback reinforcement learning RLHF module 305 and an automatic planning module 306.

[0215] The above-mentioned factory module 301 is responsible for uploading various types of factory data and performing antenna downtilt angle adjustment operations; the above-mentioned construction module includes a data acquisition submodule 302 and a data enhancement and storage submodule 303. The data acquisition submodule collects pictures, texts, and structured data of the factory according to a set fixed period; the data enhancement and storage submodule performs data enhancement on factories of the same type, filters out contaminated data, obtains a pure data set, and stores the corresponding multimodal big data; the model training module trains multimodal data and obtains the logical relationship between discourse description and line layout through statistical schemes; the RLHF module fine-tunes the pre-trained large model with manual feedback; the automatic planning module deploys the large model and generates automatic planning schemes for the production line.

[0216] like Figure 4 As shown, the temporal relationship between each module includes: the factory module 301 sends the factory's line image data, text data, and structured data to the data acquisition sub-module 302; the data acquisition sub-module 302 passes the collected traffic data to the data enhancement and storage sub-module 303 for enhancement and storage; the pre-processed data is sent to the model training module 304 of the AI ​​large model, through feature extraction and model training of the multimodal large model; then, the data enters the RLHF module 305, and the model output results are aligned with manual experience; the generated line layout results are given by the automatic planning module 306 of model deployment and automatic planning, and sent to the factory module 301 to realize the automatic generation of production lines in the area.

[0217] This embodiment supports full-process automation from demand input to layout generation, and automatically generates optimized factory line layout plans based on the process flow, equipment configuration, space constraints, and production targets input by the user. Through comprehensive judgments such as demand matching, layout rationality, and expert evaluation, it ensures that the generated results meet the multiple goals of space utilization, production efficiency, and safety, while having high availability and reliability; in addition, the practicality of the system is further improved, and the various modules work together to realize the full-link functions of data collection, enhancement, model training, and deployment. The generated layout plan can be directly sent to the factory module to support flexible production demand information of multiple factories, significantly accelerate the planning and implementation process of smart factories, and provide innovative automation solutions for industrial production.

[0218] Among them, the functional implementation of each module in the above-mentioned intelligent factory line layout automatic generation device corresponds to the various steps in the above-mentioned intelligent factory line layout automatic generation method embodiment, and its functions and implementation processes will not be repeated here one by one.

[0219] On the third aspect, an embodiment of the present application provides an automatic generation device for the line layout of an intelligent factory. The automatic generation device for the line layout of an intelligent factory can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0220] Reference Figure 5 , Figure 5 Schematic diagram of the hardware structure of the device for automatically generating the layout of a smart factory line involved in the embodiment of the present application. In the embodiment of the present application, the device for automatically generating the layout of a smart factory line may include a processor, a memory, a communication interface, and a communication bus.

[0221] The communication bus may be of any type and is used to interconnect the processor, memory, and communication interface.

[0222] Communication interfaces include input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces interconnect components within the intelligent factory layout automation device, as well as interfaces that connect the intelligent factory layout automation device to other devices (such as other computing devices or user devices). Physical interfaces can include Ethernet, fiber, or ATM interfaces; user devices can include displays and keyboards.

[0223] The memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0224] The processor may be a general-purpose processor, which may call the automatic generation program for the intelligent factory line layout stored in the memory and execute the automatic generation method for the intelligent factory line layout provided in the embodiments of the present application. For example, the general-purpose processor may be a central processing unit (CPU). The method executed when the automatic generation program for the intelligent factory line layout is called may refer to the various embodiments of the automatic generation method for the intelligent factory line layout of the present application, and will not be repeated here.

[0225] Those skilled in the art will understand that Figure 5 The hardware structure shown in the figure does not constitute a limitation to the present application and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0226] It should be noted that the serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0227] The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices. The terms "first", "second" and "third" are used to distinguish different objects, etc., and do not represent a sequence, nor do they limit the "first", "second" and "third" to different types.

[0228] In the description of the embodiments of this application, the words "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0229] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or, for example, A / B can mean A or B; “and / or” in the text is merely a description of the association relationship of associated objects, indicating that three relationships may exist, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in the description of the embodiments of the present application, “multiple” refers to two or more than two.

[0230] In some processes described in the embodiments of the present application, multiple operations or steps are included that appear in a specific order. However, it should be understood that these operations or steps may not be performed in the order in which they appear in the embodiments of the present application or may be performed in parallel. The sequence numbers of the operations are only used to distinguish between different operations, and the sequence numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations or steps may be performed in sequence or in parallel, and these operations or steps may be combined.

[0231] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for enabling a terminal device to execute the methods described in each embodiment of the present application.

[0232] The above are only preferred embodiments of the present application and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for automatically generating a line layout of an intelligent factory, characterized in that: The method comprises: According to the received production demand information, image data and text description information of the same type of line corresponding to the production demand information are obtained to construct a large-scale data set of the line type; the text description information includes the production demand information; Pre-training and reasoning the initial generative model based on the large-scale dataset to enable it to learn the mapping relationship from the text description information to the type of line planning image, thereby obtaining a first generative model; Performing reinforcement learning adjustment with manual feedback on the first generative model to obtain a second generative model; The production demand information is input into the second generation model, and a line planning image is output.

2. The method for automatically generating the layout of a smart factory line according to claim 1, characterized in that: Obtaining image data and text description information of the same type of line corresponding to the production demand information, specifically including: Obtaining a smart factory corresponding to the production demand information, and obtaining image data and text description information of the corresponding same type of line bodies in the smart factory, so as to construct a multimodal dataset of the line bodies in the smart factory; The image data and text description information of the same type of line bodies in the same industry of the smart factory are obtained to perform data enhancement and data storage on the multimodal dataset to obtain a large-scale dataset of the type of line bodies.

3. The method for automatically generating the layout of a smart factory line according to claim 1, wherein: Pre-training the initial generative model based on the large-scale dataset includes: An initial generative model is constructed using the text description information of the large-scale dataset as input and the corresponding image data as output; A total loss function is defined, and the initial generation model is trained by minimizing the total loss function to obtain a pre-trained model.

4. The method for automatically generating the layout of a smart factory line according to claim 3, wherein: The total loss function is the weighted sum of reconstruction loss, adversarial loss, and text alignment loss; The reconstruction loss is: L r =E (T,I) [||IG(T;θ)|| 2 ]; The adversarial loss is: L a =Ε I [logD(I)]+Ε T [log(1-D(G(T)))] The text alignment loss is: L t =-Ε (T,I) [cos(F(T),F(I))] Among them, D is the discriminator, G(T;θ) is the initial generation model, θ is the model parameter, T is the text description information, I is the image data; F represents the embedding vector, and cos represents the cosine similarity of the embedding vector.

5. The method for automatically generating the layout of a smart factory line according to claim 3, wherein: After obtaining the pre-trained model, the process also includes reasoning about the pre-trained model, specifically including: Defining a conditional probability distribution for forward diffusion, and forward diffusing the output image of the pre-trained model to gradually add noise to generate a noisy image; Defining a conditional probability distribution for reverse denoising, and gradually denoising the noisy image until a clear image is obtained; The first generative model is obtained based on minimizing the difference between the noisy image and the clear image.

6. The method for automatically generating the layout of a smart factory line according to claim 5, characterized in that: The conditional probability distribution of the forward diffusion is: The conditional probability distribution of the reverse denoising is: The loss function between the noisy image and the clear image is: Among them, N represents Gaussian distribution, β t is the noise coefficient at time step t, I t ′ is the noised image at time step t, I′ t-1 is the noisy image at the t-1 time step, μ θ represents the mean function parameterized by the neural network, represents the variance of the noise.

7. The method for automatically generating the layout of a smart factory line according to claim 1, wherein: Performing reinforcement learning adjustments on the first generative model using human feedback specifically includes: Training a reward model using the image data and corresponding text description information output by the first generative model as input and the collected human feedback as output; Based on the trained reward model, the first generation model is optimized using a reinforcement learning method with the goal of maximizing the reward signal.

8. The method for automatically generating the layout of a smart factory line according to claim 7, wherein: The loss function when training the reward model is: L R =Ε (T,I″) [||R φ (T,I″)-R h (T,I″)|| 2 ] The reward signal is: L RL =Ε (T,I″) [R φ (T,I″)-λKL(G current ,G pretrained )] Among them, R h (T,I″) is manual feedback, R φ (T,I″) is the reward model, I″ is the planning image generated by the first generative model, and λ is a hyperparameter.

9. An automatic generation device for intelligent factory line layout, characterized in that: The device comprises: A construction module is used to obtain image data and text description information of the same type of line bodies corresponding to the received production demand information, so as to construct a large-scale data set of the line bodies of the same type; A model training module is used to pre-train and infer the initial generation model based on the large-scale dataset, so that it learns the mapping relationship from text description information to the type of line planning image to obtain a first generation model; A human feedback reinforcement learning module, configured to perform reinforcement learning adjustments using human feedback on the first generative model to obtain a second generative model; An automatic planning module is used to input the production demand information into the second generation model and output a line planning image.

10. An automatic generation device for intelligent factory line layout, characterized in that: The intelligent factory line layout automatic generation device includes a processor, a memory, and an intelligent factory line layout automatic generation program stored in the memory and executable by the processor, wherein when the intelligent factory line layout automatic generation program is executed by the processor, the steps of the intelligent factory line layout automatic generation method as described in any one of claims 1 to 8 are implemented.

Images