A workshop equipment layout generation method, device and storage medium based on a mixed diffusion model

Abstract

The present invention discloses a method for generating workshop equipment layout based on a hybrid diffusion model, which belongs to the field of artificial intelligence and computer vision. The method comprises the following steps: defining a set of equipment containing discrete semantic labels and continuous geometric attributes, and using the workshop floor plan as a boundary constraint; constructing a hybrid diffusion model, and gradually injecting noise into the equipment attributes through independent discrete domain and continuous domain noise functions; using a denoising network with a time-varying Transformer as the core, and taking the floor plan features as conditions, iteratively recovering from the noise and generating a reasonable equipment layout. The method provided by the present invention can collaboratively process and generate discrete equipment semantic labels and continuous equipment geometric attributes, significantly improving design efficiency and flexibility.

Description

Technical Field

[0001] The present invention relates to the technical field of the combination of computer vision and artificial intelligence, and in particular to a method, device and storage medium for generating workshop equipment layout based on a hybrid diffusion model. Background Art

[0002] In the fields of industrial design and manufacturing, workshop equipment layout is a fundamental and critical process. Traditional layout design methods often rely on preset rules, standardized templates, or the designer's experience. This approach is somewhat efficient when handling standardized, repetitive tasks. However, as the manufacturing industry evolves towards flexibility and customization, the demand for flexibility and diversity in layout solutions is increasing. This has led to the limitations of traditional methods becoming increasingly prominent. They often struggle to quickly respond to and generate optimized solutions that meet specific personalized needs.

[0003] To overcome the shortcomings of traditional methods, the industry has begun to introduce data-driven deep learning technologies for layout generation. While these methods have improved the automation level of scene generation to a certain extent, challenges remain in practical application. For example, when processing mixed data such as shop floor equipment, which contains both discrete categorical information and continuous geometric information, existing models struggle to balance both.

[0004] Specifically, while traditional diffusion models applied to continuous data (such as images) can generate content, they suffer from a high degree of randomness when controlling continuous geometric properties such as the position and size of generated objects, making it difficult to ensure accurate and reasonable layouts. On the other hand, diffusion models designed for discrete data, when used to represent three-dimensional space, can disrupt the continuity and smoothness of the scene in physical space, compromising the realism of the layout.

[0005] Therefore, there is an urgent need for a new technical solution that can effectively process mixed domain data consisting of discrete semantics and continuous geometry, and at the same time incorporate the physical boundaries of the workshop as hard constraints into the generation process, so as to achieve fast, accurate and constraint-compliant equipment layout generation. Summary of the Invention

[0006] The present invention provides a method for generating a workshop equipment layout based on a hybrid diffusion model to solve the problems mentioned in the background technology, comprising the following steps:

[0007] S1. Define the set of equipment to be laid out, where each equipment contains a discrete semantic label and continuous geometric attributes, and use the workshop floor plan as the boundary constraint of the layout;

[0008] S2. injecting discrete domain noise and continuous domain noise into the semantic label and geometric attributes of the device, respectively, to gradually convert the original data into noise data;

[0009] S3. Using a time-varying Transformer-based denoising network, conditional on the boundary features of the workshop plan, gradually recover and generate a device layout solution containing semantic labels and geometric attributes from the noisy data;

[0010] S4. Feed the user-specified constrained device information as additional conditions into the denoising network to generate a layout solution that meets the specific constraints.

[0011] Furthermore, the geometric properties of the device in step S1 include the center position, size and rotation angle of the device.

[0012] Furthermore, step S2 includes: using independent discrete domain noising functions and continuous domain noising functions to add noise to the discrete semantic labels and the continuous geometric attributes respectively, and performing independent sampling training on the two.

[0013] Furthermore, in step S3, the loss function for training the denoising network is composed of a discrete domain loss and a continuous domain loss, so that the model can accurately predict device labels and optimize their geometric properties.

[0014] Furthermore, the time-varying Transformer denoising network includes:

[0015] A multi-layer perceptron (MLP) for encoding equipment geometric attributes and a PointNet network for extracting boundary features of the workshop floor plan;

[0016] A time-varying Transformer decoder that takes time-step information as input, uses adaptive layer normalization and multi-head attention mechanism to jointly process semantic, geometric features and boundary constraints;

[0017] Two independent output MLPs are used to generate the categorical distribution of semantic labels and the Gaussian mean of geometric features, respectively.

[0018] Furthermore, the time-varying Transformer decoder injects time step information through adaptive layer normalization AdaLN and processes the conditional input consisting of planar graph boundary features through a multi-head cross attention mechanism.

[0019] Furthermore, in step S3, the boundary constraints of the workshop plan are processed. Specifically, the steps include: sampling feature points from the boundary of the workshop plan, and extracting plane features using the PointNet network.

[0020] Furthermore, in step S4, the processing step of the additional conditions includes: using a feature encoder to process the constraint device information, and providing the processed information to the denoising network to retrain the model.

[0021] A computer device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the program implements the above-mentioned workshop equipment layout generation method when executed by the processor.

[0022] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned workshop equipment layout generation method.

[0023] Beneficial effects:

[0024] This paper proposes an innovative hybrid diffusion model architecture that can collaboratively process and generate discrete device semantic labels and continuous device geometric attributes. By designing independent noise addition and denoising strategies for different data domains, this approach overcomes the technical bottleneck of a single model's difficulty in simultaneously optimizing two different data types, ensuring the accuracy of the generated results in both categorization and physical parameters.

[0025] This method uses the physical floor plan of the workshop as a boundary constraint, extracting its key geometric features through a PointNet network. Using a time-varying Transformer decoder, the generation process strictly adheres to the constraints of the physical space. This not only ensures that all generated devices are located within the valid area, significantly improving the practical feasibility and rationality of the layout solution, but also makes the layout more adaptable to workshop environments of varying shapes and sizes.

[0026] The layout method provided by this invention offers the flexibility to support user-defined constraints. Users can pre-specify the locations or attributes of key equipment as additional conditions for model generation. The model then optimizes the layout based on this prior information, generating customized solutions tailored to specific process flows or user preferences. This significantly enhances the method's practicality and human-computer interaction capabilities. This comprehensive solution, combining hybrid data processing, boundary constraints, and user customization, ultimately generates a rationally arranged, precisely detailed, and practically tailored equipment layout, effectively improving design efficiency and quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is the transformer-based network model architecture of the present invention;

[0028] Figure 2 It is a workflow diagram for generating workshop equipment layout. DETAILED DESCRIPTION

[0029] 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 making creative efforts are within the scope of protection of the present invention.

[0030] Example 1:

[0031] The diffusion model used in the present invention is based on a Markov process consisting of forward noise addition and backward noise removal.

[0032] The forward noise addition process iterates over the geometric properties of the device Introducing controlled noise to obtain the hidden variable sequence after chaos The reverse process is done through the denoising network Gradually eliminate the noise of latent variable x, Obtained by training the denoising network. The loss function of network training is usually used to minimize the negative log probability ,in , represents the initial noise distribution With prior distribution KL divergence; usually Gaussian distribution is used to represent , , the forward process and the inverse process of model prediction are combined to optimize the denoising process at each time step t; Describing Model Predictions Step real data so that the generated data is as close to the true distribution as possible.

[0033] This embodiment provides a method for generating workshop equipment layout based on the mixed diffusion model, such as Figure 2 As shown, the following steps are included:

[0034] S1. Define the set of equipment to be laid out, where each equipment contains a discrete semantic label and continuous geometric attributes, and use the workshop floor plan as the boundary constraint of the layout.

[0035] The equipment set to be arranged in the workshop is ,equipment The label is , The center position of the body The product bounding box is , the rotation angle around the z axis relative to the initial position , 、 and Representation device , based on the invention method equipment layout combination optimization C devices, give the layout , x is the hidden variable of the mixed diffusion process.

[0036] In this embodiment, the above-mentioned equipment set is formed by selecting part of the workshop and equipment data from the data set 3D-FRONT and downloading part of the related workshop and equipment data from the Internet to form the data set that conforms to this embodiment.

[0037] In terms of data form, discrete semantic labels are used to define "what equipment is," specifically identifiers used to distinguish different equipment categories. For example, a workshop environment might include equipment with different functions, such as "machine tools," "conveyor belts," "robotic arms," ​​"inspection tables," or "shelves." Semantic labels assign each piece of equipment a specific category name. During data processing, these text labels are typically converted into a computer-readable format, such as one-hot encoded vectors, to facilitate subsequent processing in discrete diffusion models. Continuous geometric attributes refer to a series of continuous numerical values ​​that accurately describe the state of a device in three-dimensional physical space. Specifically, they include the coordinates of the device's 3D center position, its 3D dimensions, and the rotation angle of the device around its vertical axis. Together, these attributes define the device's location, size, and orientation within the workshop.

[0038] For the above datasets, the datasets are parsed by workshop type and the parsing results are saved in the corresponding directory; the datasets are preprocessed by workshop type and saved in the corresponding subdirectories.

[0039] S2. Inject discrete domain noise and continuous domain noise into the semantic label and geometric attributes of the device, respectively, to gradually transform the original data into noise data. The specific operations are as follows:

[0040] In the discrete denoising diffusion model and the continuous domain diffusion model, and Adding domain-specific noise ,in is the discrete domain noise function, is the noise function for the continuous domain;

[0041] right and Perform independent sampling training to calculate the posterior probability of the mixed domain diffusion model .

[0042] S3. A denoising network based on a time-varying Transformer is used to gradually recover and generate a device layout solution containing semantic labels and geometric attributes from the noisy data, using the boundary features of the workshop plan as a condition. The specific steps are as follows:

[0043] The loss function for training the denoising network in this step consists of a discrete domain loss and a continuous domain loss. This function aims to enable the model to accurately predict device labels and optimize their geometric properties. The specific design is as follows:

[0044] Designing the network Calculate the probability distribution of latent variables in discrete and continuous domains to make it consistent with the noise addition process ;

[0045] Compute discrete domain and continuous domain specific loss functions and and :

[0046]

[0047]

[0048]

[0049] .

[0050] in:

[0051] is the diffusion loss in the continuous domain, is the diffusion loss in the discrete domain;

[0052] z is a discrete variable, representing the semantic label of the object;

[0053] x is a continuous variable, representing the geometric properties of the object;

[0054] y is the conditional input, representing the boundary characteristics of the workshop plan;

[0055] t is the time step of the diffusion process;

[0056] q is the forward noise addition process in the diffusion model;

[0057] p is the reverse de-anxiety process in the diffusion model;

[0058] D KL represents the KL divergence, which is used to measure the difference between two distributions;

[0059] E is the expected value;

[0060] Calculate the loss function and the variational bounds of discrete and continuous domains ,in In order for the model to correctly predict the label type of the device object, is the regularization coefficient.

[0061] The denoising network design method based on time-varying Transformer used in this step is as follows:

[0062] Geometric attributes are encoded through MLP and combined with trainable semantic embeddings; 256 points are sampled from the floor plan image along the plane boundary of the workshop and the supernormal vector is calculated; the plane features are extracted using the PointNet network to better capture the equipment layout boundary of the workshop; the plane features are mixed with the learned position embedding to form a conditional vector; the time-varying transformer decoder accepts object features, injects time steps through adaptive layer normalization, uses a multi-head attention mechanism to capture boundary constraints, adjusts the reverse input, processes the conditional input through a multi-head cross-attention layer, and jointly processes semantic and geometric features to predict diffusion updates; the output of the transformer decoder is passed into two MLPs to generate a categorical distribution of semantic labels and a Gaussian mean of geometric features.

[0063] The de-anxiety network obtained based on the above method is as follows Figure 1 As shown in the figure, the network consists of an input, 8 transformer blocks, and a decoder. Semantic attributes are embedded into a linear vector of length 512, and geometric attributes are mapped to a 512-dimensional space through a 3-layer MLP with a dimension of [512, 1024] and then fed into the transformer decoder. A 64-dimensional shop floor plan feature is extracted using a 4-layer PointNet with a dimension of [64, 64, 512]. The Transformer block includes an adaptive normalization layer AdaLN, an 8-head multi-head attention mechanism with a dimension of 512, and a feedforward layer with a dimension of 2048. The output of the Transformer decoder is fed into two independent MLPs to decode the semantic and geometric predictions. The semantic feature decoder produces a categorical distribution on z; the geometric feature decoder consists of a 3-layer MLP and outputs the 8-dimensional Gaussian mean of x.

[0064] The input data input to the network consists of preprocessed semantic attributes and geometric attributes. The semantic attributes are generated into semantic vectors through the semantic embedding module, and the geometric attributes are transmitted to the MLP encoder to generate embedding vectors. The semantic vectors and geometric vectors are added together to form the preprocessed input data.

[0065] The PointNet network extracts spatial features from the planar graph constraints, and the spatial features are combined with position embedding to form a spatial position representation.

[0066] The input of the Transformer decoder includes the current time step t , the state at time step t-1 , and gradually generate the denoised output through iteration. The output of the Transformer decoder is generated by the MLP module .

[0067] S4. The constrained device information specified by the user is fed into the denoising network as an additional condition to generate a layout solution that satisfies specific constraints. Specifically, a feature encoder is used to process the constrained device information, and the processed information is provided to the denoising network to retrain the model. This method pre-specifies the positions or attributes of some key devices as additional conditions for generating the model. The model can perform layout optimization based on this prior information, thereby generating a customized solution that meets specific process flows or user preferences, greatly enhancing the practicality and human-computer interaction capabilities of the method.

[0068] The workshop equipment layout generation method based on the mixed diffusion model provided in this embodiment is as follows Figure 2 As shown, the present invention uses the physical plan of the workshop as a boundary constraint, extracts its key geometric features through the PointNet network, and uses the time-varying Transformer decoder to ensure that the generation process strictly follows the limitations of the physical space. This not only ensures that all generation equipment is located within the effective area, significantly improving the practical usability and rationality of the layout plan, but also enables the layout to better adapt to workshop environments of different shapes and sizes. By designing independent noise addition and denoising strategies for different data domains, this method overcomes the technical bottleneck that a single model is difficult to simultaneously optimize two data of different properties, ensuring the dual accuracy of the generation results in terms of category and physical parameters.

[0069] Example 2:

[0070] This embodiment provides a computer device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the program is executed by the processor, the above-mentioned workshop equipment layout generation method is implemented.

[0071] Example 3:

[0072] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the above-mentioned workshop equipment layout generation method is implemented.

[0073] In addition to the above embodiments, the present application also includes other implementation methods. Any technical solutions formed by equivalent transformation or equivalent replacement should fall within the scope of protection of the claims of the present invention.

Claims

1. A method for generating workshop equipment layout based on a mixed diffusion model, characterized in that: The following steps are involved: S1. Define the set of equipment to be laid out, where each equipment contains a discrete semantic label and continuous geometric attributes, and use the workshop floor plan as the boundary constraint of the layout; S2. injecting discrete domain noise and continuous domain noise into the semantic label and geometric attributes of the device, respectively, to gradually convert the original data into noise data; S3. Using a time-varying Transformer-based denoising network, conditional on the boundary features of the workshop plan, gradually recover and generate a device layout solution containing semantic labels and geometric attributes from the noisy data; S4. feeding the user-specified constraint device information as additional conditions into the denoising network to generate a layout solution that satisfies the constraints; The method of injecting discrete domain noise and continuous domain noise into the semantic label and geometric attributes of the device in step S2 is: In the discrete denoising diffusion model and the continuous domain diffusion model, and Add domain-specific noise: ； in Device labels, including discrete semantic labels and continuous geometric attributes, is the hidden variable of the mixed diffusion process, is the discrete domain noise function, is the noise function for the continuous domain; The method of gradually transforming the original data into noise data is: right and Perform independent sampling training and calculate the posterior probability of the mixed domain diffusion model: ； The loss function of the denoising network of the time-varying Transformer in step S3 is composed of discrete domain loss and continuous domain loss. The specific design steps include: Designing the network Calculate the probability distribution of latent variables in discrete and continuous domains to make it consistent with the noise addition process ; Compute discrete domain and continuous domain specific loss functions and and : ； ； ； ； in: is the diffusion loss in the continuous domain, is the diffusion loss in the discrete domain; z is a discrete variable, representing the semantic label of the object; x is a continuous variable, representing the geometric properties of the object; y is the conditional input, representing the boundary characteristics of the workshop plan; t is the time step of the diffusion process; q is the forward noise addition process in the diffusion model; p is the reverse de-anxiety process in the diffusion model; D KL represents the KL divergence, which is used to measure the difference between two distributions; E is the expected value; Then the loss function of the variational boundary of the discrete domain and the continuous domain is: ； in To enable the model to correctly predict the label type of the device object, the auxiliary loss is the regularization weight coefficient.

2. The method according to claim 1, wherein: The geometric properties of the device in step S1 include the center position, size and rotation angle of the device.

3. The method according to claim 1, characterized in that Step S2 includes: using independent discrete domain noising functions and continuous domain noising functions to add noise to the discrete semantic labels and the continuous geometric attributes respectively, and performing independent sampling training on the two.

4. The method according to claim 1, wherein: In step S3, the loss function for training the denoising network is composed of a discrete domain loss and a continuous domain loss, so that the model can accurately predict device labels and optimize their geometric properties.

5. The method according to claim 1, wherein In step S3, the time-varying Transformer denoising network includes: A multi-layer perceptron (MLP) for encoding equipment geometric attributes and a PointNet network for extracting boundary features of the workshop floor plan; A time-varying Transformer decoder that takes time-step information as input, uses adaptive layer normalization and multi-head attention mechanism to jointly process semantic, geometric features and boundary constraints; Two independent output MLPs are used to generate the categorical distribution of semantic labels and the Gaussian mean of geometric features, respectively.

6. The method according to claim 5, characterized in that The time-varying Transformer decoder injects time-step information through the adaptive layer normalized AdaLN and processes the conditional input consisting of planar graph boundary features through a multi-head cross-attention mechanism.

7. The method according to claim 1, characterized in that In step S3, the boundary constraints of the workshop plan are processed. The specific steps include: sampling feature points from the boundary of the workshop plan and extracting plane features using the PointNet network.

8. The method according to claim 1, characterized in that In step S4, the processing step of the additional conditions includes: using a feature encoder to process the constraint device information, and providing the processed information to the denoising network to retrain the model.

9. A computer device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein when the program is executed by the processor, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images