Multi-feature constraint-oriented data analysis model
Through a data analysis model for multi-feature constraints, the generation adversarial network and dual discriminator mechanism is used, combined with multiple network structures and constraints, the accuracy and efficiency of resource allocation in the enterprise's production process is solved, efficient and reasonable resource allocation is achieved, and production efficiency is improved.
Patent Information
- Application Number
- CN202510564948.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-19
AI Technical Summary
The existing enterprise production process resource allocation algorithm is insufficient in the face of complex and variable factors, relies on manual experience, has low computing efficiency, and cannot respond quickly to changes in production demand, resulting in unreasonable resource allocation.
A data analysis model for multi-feature constraints is adopted, including data acquisition, improved GAN network and configuration optimization module. By generating an adversarial network and a dual discriminator mechanism, combining transform encoder, UNet decoder and attention selector, constraint conditions and punishment mechanisms are set to achieve multi-dimensional optimization of resource configuration.
It improves the accuracy and rationality of resource allocation, can quickly respond to diversified order needs, generate optimized resource allocation plans, and improve production efficiency.
Smart Images
Figure CN120509641A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data analysis technology, and in particular to a data analysis model under multi-feature constraints. Background Art
[0002] In a company's production operations, rational resource allocation is crucial. This resource allocation determines key factors such as production efficiency, cost control, and product quality. By optimizing resource allocation, companies can effectively reduce production costs, improve production efficiency, and quickly respond to changes in market demand, thereby enhancing their market competitiveness. For example, in the manufacturing industry, precise raw material procurement arrangements, rational equipment scheduling, and optimized human resource allocation can ensure efficient production line operation and avoid idle resources or shortages.
[0003] Currently, the development of resource allocation in enterprises' production processes has achieved some success. Some large enterprises have implemented preliminary information-based management of production resources using advanced management systems. However, the overall situation still faces many challenges. Firstly, the numerous complex and variable factors in the production process, such as fluctuations in raw material supply, uncertainty in order demand, and equipment failures, complicate resource allocation. Secondly, traditional resource allocation methods often rely on manual experience and lack precise quantitative analysis, making them difficult to adapt to the rapidly changing production environment of modern enterprises.
[0004] The application of intelligent algorithms in resource allocation in production processes is gradually gaining popularity. They can rapidly analyze and process large amounts of data, thereby providing optimized resource allocation solutions. For example, machine learning algorithms can analyze historical production data to predict resource requirements over different time periods. However, the current application of intelligent algorithms also has shortcomings and challenges. First, the accuracy of intelligent algorithms is highly dependent on the quality and completeness of the data. Data bias or omissions can lead to significant errors in the algorithm's results. Second, in terms of analytical comprehensiveness, existing resource allocation algorithm models often focus on primary production factors, insufficiently considering minor factors that may have a significant impact in specific scenarios. Regarding computational efficiency, the exponential growth in production data volumes significantly increases the computation time of complex intelligent algorithms. When faced with urgent production adjustments, the algorithms are unable to quickly recalculate and optimize resource allocation, resulting in a delay in responding to production demands. Summary of the Invention
[0005] The present invention aims to provide a data analysis model under multi-feature constraints, which can efficiently complete the resource allocation of the enterprise production process, and the generated configuration plan is better, which helps to maximize production efficiency.
[0006] The basic solution provided by the present invention is: a data analysis model under multi-feature constraints, including a data acquisition module, an improved GAN network module and a configuration optimization module;
[0007] The data acquisition module is used to collect the company's production basic data and production demand data; the production basic data includes the company's production equipment, production manpower and production materials; the production demand data includes production orders;
[0008] The improved GAN network module is used to generate resource allocation plans; the improved GAN network module is provided with an input layer, a generator network, a discriminant network, and an output layer; the input layer takes production basic data and production demand data as input; the generator network is used to generate basic allocation plans, and the discriminant network is used to discriminate the generated basic allocation plans and analyze the effectiveness of each basic allocation plan to obtain a score; the output layer is used to output the basic allocation plan with an effectiveness score higher than a threshold as the resource allocation plan;
[0009] The configuration optimization module is used to impose constraints on the generation process of the generation network to optimize the generation process.
[0010] Furthermore, it also includes a preprocessing module; the preprocessing module is used to preprocess production basic data and production demand data; the preprocessing includes data cleaning, normalization processing and category coding.
[0011] Furthermore, the discriminant network is a binary discriminant network, including a discriminant network P and a discriminant network N; the discriminant network P and the discriminant network N are provided with the same initial loss function; the discriminant network N is provided after the generative network, and the discriminant network P is provided independently.
[0012] Furthermore, the improved GAN network module is pre-trained using a specified data set; the specified training set includes real data samples based on historical production basic data, historical production demand data and historical resource configuration data, as well as virtual data samples marked as false.
[0013] Furthermore, the improved GAN network module includes the following training steps during training:
[0014] For the discriminant network, the specified data set is used as a training set to input the discriminant network P, and the discriminant network P is trained as a binary classifier;
[0015] Determine a penalty factor for the binary classifier to force the probability of identifying real data samples to be as close to 1 as possible and the probability of identifying generated samples to be as close to 0 as possible; the generated samples are generated by the generative network and are provided with false labels;
[0016] When the discriminant network P is trained, its weight parameters are shared with the discriminant network N, so that the discriminant network N and the discriminant network P have the same ability to distinguish true and false samples;
[0017] The false labels of the generated samples of the generator network are changed to true, and used as the training set of the discriminant network N, forcing the discriminant network N to expect the generated samples to have a probability close to 1, and not to judge the real data samples; if the discriminant network N expects the generated samples to have a probability close to 0, it will form a contrasting feedback with the former, forcing the generator network to change its parameters to generate pseudo samples that are more similar to the real data samples;
[0018] The training is terminated when the output probabilities of the discriminant network P and the discriminant network N for true and false samples are close and reach 1 / 2.
[0019] Furthermore, the configuration optimization module is also used to set a penalty mechanism for the generation model; when the basic configuration solution generated by the generation model violates the constraint conditions, the penalty mechanism is triggered.
[0020] Furthermore, the constraints include: hard constraints set based on the Lagrange multiplier method - production capacity upper limit, cost threshold and order deadline.
[0021] Furthermore, the constraint conditions include: soft constraint conditions set based on fuzzy logic - market volatility index and supply chain stability index.
[0022] Furthermore, the generative network is provided with a transform encoder, a UNet decoder and an attention selector.
[0023] Furthermore, a sliding window incremental learning method is used to update and improve the GAN network module in real time.
[0024] The working principle and advantages of the present invention are:
[0025] The present invention provides a data analysis model for multi-feature constraints that can efficiently complete resource allocation in an enterprise's production process, and the generated configuration plan is more optimal, helping to maximize production benefits. The key points are:
[0026] First, this solution integrates multimodal information, comprehensively considering the impact of multiple factors on resource allocation. By leveraging the generative adversarial network mechanism, the generative and discriminative networks interact to optimize resource allocation. Furthermore, a separate configuration optimization module is implemented to set constraints, enabling it to match diverse order requirements in actual production, improving the accuracy and rationality of resource allocation.
[0027] Second, a new generative adversarial network structure is designed in this solution to achieve better resource allocation. The generative network integrates a transform encoder, a UNet decoder, and an attention selector. The transformer can capture long-distance dependencies in resource allocation; the UNet excels at optimizing local details. The combination of the two enables resource allocation at multiple scales. Furthermore, combined with the attention selector, it can adapt to dynamically changing production environments. The discriminant network is specially designed as a binary discriminant network, which can form a dual-gradient feedback mechanism. The generative network must simultaneously meet the evaluation criteria of the two discriminant networks, which can avoid a single discriminator being too strong and causing training imbalance, and can form a complementary effect, thereby achieving the purpose of optimizing generation quality.
[0028] Third, this solution designs a new network training method that can further improve the network's solution generation efficiency and quality. Specifically, based on the traditional GAN framework, this solution introduces a dual discriminator (P and N) and a staged adversarial mechanism. Through label inversion, parameter sharing, and dynamic termination conditions, it constructs a more stable and efficient generative adversarial training paradigm. Through dual adversarial feedback and discriminator capability inheritance, it can force the generator to approximate the real data distribution in multiple dimensions, thereby improving the quality of the configuration solution and helping to maximize production benefits. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 This is a schematic structural diagram of an embodiment of a data analysis model under multi-feature constraints according to the present invention;
[0030] Figure 2 This is a schematic diagram of the improved GAN network module structure of an embodiment of the present invention for a data analysis model under multi-feature constraints. DETAILED DESCRIPTION
[0031] The following is a further detailed description through specific implementation methods:
[0032] Example 1
[0033] The embodiment is basically as shown in the attached Figure 1 Shown: A data analysis model under multi-feature constraints, including a data acquisition module, a preprocessing module, an improved GAN network module and a configuration optimization module.
[0034] The data acquisition module is used to collect the company's basic production data and production demand data; the basic production data includes the company's production equipment, production manpower and production materials; the production demand data includes production orders.
[0035] Among them, the production equipment includes equipment type, equipment quantity, equipment status (including equipment operating status, maintenance records and failure rate), and the corresponding processing scope of the equipment; the production manpower includes the number of personnel, personnel idleness (including workload) and the corresponding functional scope of each personnel; the production materials include material type, material quantity, material status (including inventory level, procurement cycle and supplier information), and the corresponding products of the materials; the production order includes order demand, order priority, and order time.
[0036] The preprocessing module is used to preprocess production base data and production demand data; this preprocessing includes data cleaning, normalization, and category coding. Data cleaning effectively addresses missing values and outliers in the data. Normalization converts production base data and production demand data into dimensionless values, facilitating unified analysis. Category coding converts non-numeric data in production base data and production demand data into numerical form, further facilitating improved GAN network module analysis. In practical applications, either normalization or category coding can be used.
[0037] Furthermore, the preprocessing module is also used to construct a resource state feature matrix (corresponding to the production base data) and an order feature tensor (corresponding to the production demand data) based on the preprocessed production base data and production demand data. In this embodiment, a resource state feature matrix with a three-dimensional tensor is constructed based on production equipment, production manpower and production materials. In this matrix, each dimension corresponds to a resource type (production equipment, production manpower or production material), wherein the characteristic attributes corresponding to the production equipment include equipment availability and equipment processing capacity, such as [equipment 1_availability, equipment 1_processing capacity]; the characteristic attributes corresponding to the production manpower include personnel availability and personnel capacity, such as [personnel 1_availability, personnel 1_capacity]; the characteristic attributes corresponding to the production materials include inventory index and material quality, such as [material 1_inventory index, material 1_quality]. The above attribute data are obtained based on the corresponding production base data.
[0038] Based on production demand data, a four-dimensional tensor can be constructed, with each dimension corresponding to a different feature attribute (order demand, order priority, order time); for example, [Order 1_Requirement_Specification], [Order 1_Requirement_Quantity, Order 1_Requirement_Special Requirements], [Order 1_Priority_Urgency, Order 1_Priority_Contract Terms], and [Order 1_Time Window_Start Time, Order 1_Time Window_End Time, Order 1_Time Window_Deadline]. The above attribute data is obtained based on the production demand data.
[0039] By constructing a resource status feature matrix and an order feature tensor, we can effectively integrate various resource and demand information, thereby helping to improve the accuracy of resource allocation. Furthermore, the resource status feature matrix can clearly reflect the distribution and availability of resources, providing basic data for subsequent resource allocation. For example, by comprehensively evaluating the operating status of equipment, manpower availability, and material inventory levels, we can accurately determine which resources can be used for order production. The setup based on the order feature tensor can better capture the differences and relationships between orders, helping to generate resource allocation plans that meet order requirements. For example, resources can be allocated preferentially to high-priority orders with tight deadlines.
[0040] The improved GAN network module is used to generate resource allocation plans; the improved GAN network module is equipped with an input layer, a generating network, a discriminant network, and an output layer; the input layer takes production basic data and production demand data as input; the generating network is used to generate basic configuration plans, the discriminant network is used to discriminate the generated basic configuration plans, and analyzes the effectiveness score of each basic configuration plan (its value range is 0 to 1, and the higher the score, the better the configuration effect of the plan); the output layer is used to output the basic configuration plan with a validity score higher than the threshold as the resource allocation plan. In actual application, the threshold can be adjusted according to actual needs. In this embodiment, the generated resource allocation plan includes equipment resource configuration corresponding to the production order (including equipment allocation, equipment usage time arrangement, etc.), human resource configuration (including personnel allocation, personnel work time arrangement, etc.), material resource configuration (including material allocation, etc.) and time resource configuration (including order production sequence, etc.).
[0041] Specifically, in this embodiment, the generation network is provided with a transform encoder, a UNet decoder and an attention selector. Among them, the transformer has the ability to capture global features and can capture long-distance dependencies in resource allocation; UNet is good at optimizing local details. The combination of the two can perform resource allocation at multiple scales. For example, when considering the resource allocation of the entire production system, the transformer can obtain the global demand and supply situation, while the UNet can make detailed resource adjustments to the local production links. Furthermore, combined with the attention selector, it can adapt to dynamically changing production environments. Specifically, because different features have different degrees of influence on resource allocation under different production situations; the attention selector (i.e., a feature selector embedded in the attention mechanism) can dynamically adjust the feature weights so that the network can focus on the most important features in the current production environment, thereby improving the accuracy of the resource allocation plan. For example, when the supply of materials is tight, the weights of features related to the materials will increase.
[0042] like Figure 2 As shown, the discriminant network is a binary discriminant network, comprising a discriminant network P and a discriminant network N. Both networks share the same initial loss function. The discriminant network N is placed after the generative network, while the discriminant network P is independently configured. This binary discriminant network configuration creates a dual gradient feedback mechanism, requiring the generative network to simultaneously meet the evaluation criteria of both discriminant networks. This prevents training imbalance caused by a single overly strong discriminator. If one discriminant network fails (e.g., due to overfitting), the other can still provide effective gradients, improving model robustness. Furthermore, the detailed division of labor between the discriminant networks creates a complementary effect, thereby optimizing generation quality. Furthermore, the discriminant network N is placed after the generative network, with the output of the generative network directly fed into the discriminant network N, forming a closed-loop feedback loop. This shortens the gradient backpropagation path and makes optimization more direct. The independently configured discriminant network P allows for the introduction of external supervisory signals (e.g., a small amount of real data), preventing the degradation of the discriminant network N's discriminative ability due to its limited exposure to generated samples.
[0043] The configuration optimization module is used to impose constraints on the generation process of the generative network to optimize the generation process. The module is also used to set a penalty mechanism for the generation model; when the basic configuration solution generated by the generation model violates the constraints, the penalty mechanism is triggered. This penalty mechanism can force the configuration solution generated by the network to comply with the constraints, helping to improve the feasibility of the solution.
[0044] The constraints include hard constraints set using the Lagrange multiplier method—capacity cap, cost threshold, and order deadline. These cap, cost threshold, and order deadline can all be set based on an existing objective function. Optionally, in actual applications, additional or fewer constraints can be added based on actual configuration needs.
[0045] The improved GAN network module is pre-trained using a specified dataset; the specified training dataset contains real data samples based on historical production basic data, historical production demand data, and historical resource configuration data, as well as virtual data samples marked as false. The real data samples are labeled as true.
[0046] The improved GAN network module includes the following training steps during training:
[0047] For the discriminant network, the specified data set is used as a training set to input the discriminant network P, and the discriminant network P is trained as a binary classifier;
[0048] Determine a penalty factor for the binary classifier to force the probability of identifying real data samples to be as close to 1 as possible and the probability of identifying generated samples to be as close to 0 as possible; the generated samples are generated by the generative network and are provided with false labels;
[0049] When the discriminant network P is trained, its weight parameters are shared with the discriminant network N, so that the discriminant network N and the discriminant network P have the same ability to distinguish true and false samples;
[0050] The false labels of the generated samples of the generator network are changed to true, and used as the training set of the discriminant network N, forcing the discriminant network N to expect the generated samples to have a probability close to 1, and not to judge the real data samples; if the discriminant network N expects the generated samples to have a probability close to 0, it will form a contrasting feedback with the former, forcing the generator network to change its parameters to generate pseudo samples that are more similar to the real data samples;
[0051] The training is terminated when the output probabilities of the discriminant network P and the discriminant network N for true and false samples are close and reach 1 / 2.
[0052] Through the above training steps, based on the traditional GAN framework, this scheme introduces a dual discriminator (P and N) and a staged adversarial mechanism. Through label inversion, parameter sharing and dynamic termination conditions, it constructs a more stable and efficient generative adversarial training paradigm. Through dual adversarial feedback and discriminator capability inheritance, it can force the generator to approximate the real data distribution in multiple dimensions.
[0053] First, the collaborative training of the two discriminators provides complementary supervision, helping to improve training stability and feedback diversity. Second, through label inversion, during training phase N, the labels of generated samples are forced to be 1 (real), while N's goal is to classify them as 1. If the quality of generated samples is insufficient (N's output approaches 0), the generator receives a reverse gradient signal, forcing it to adjust its parameters to improve fidelity. This design effectively introduces a self-adversarial mechanism, where the generator must simultaneously optimize against the conflicting objectives of the two discriminators. This pressure pushes the generated samples to approximate real data in multiple dimensions, achieving better sample generation. Third, through parameter sharing, N is initially equipped with the same discriminative power as P, avoiding the "discriminator lag" problem that can occur when the generator and discriminator are optimized simultaneously in traditional GANs, and preventing N from overfitting the generated samples. Fourth, a dynamic termination condition is set: training terminates when the output probabilities of both discriminators P and N for real and fake samples reach 1 / 2, indicating that the generated samples are indistinguishable from real samples from the discriminator's perspective (reaching a Nash equilibrium). Traditional GANs usually rely on experience to set training rounds or observe the quality of generated samples. However, this solution automatically determines convergence through quantitative indicators (output probability), which can reduce human intervention from the root.
[0054] This embodiment provides a data analysis model for multi-feature constraints, which can efficiently complete the resource allocation of the enterprise production process, and the generated configuration plan is better, which helps to maximize production efficiency.
[0055] Example 2
[0056] A data analysis model for multi-feature constraints is improved as follows based on the first embodiment.
[0057] The discriminant network P and the discriminant network N are both provided with a GCN layer and a TCN layer set up in parallel. Specifically, since in the production system, there is often a complex spatial topological relationship between equipment, manpower and materials, in this solution, the GCN layer (graph convolutional network layer) set up can effectively analyze this topological structure and analyze the resource allocation plan from the spatial dimension. Furthermore, in a production environment, the production process often has time dependence, such as the sequence of orders, the maintenance cycle of equipment, etc. In this solution, the TCN layer (time convolutional network layer) set up can identify this timing pattern and judge whether the resource allocation plan is reasonable in time and whether it can meet the delivery time requirements of the order. The combination of the two can judge the resource allocation plan from multiple dimensions and obtain the optimal output.
[0058] During training, the improved GAN network module uses a virtual production environment simulator for training and testing, and sets the reward function during training to R = λ1·Profit-λ2·Cost-λ3·DelayPenalty. Here, λ1, λ2, and λ3 are weight coefficients; Profit is profit, Cost is cost, and DelayPenalty is the delay penalty. The virtual production environment simulator can simulate a real production environment, providing a secure testing and optimization platform for the improved GAN network module. In this simulation environment, different resource allocation schemes can be tested without impacting actual production, thereby improving the performance of the improved GAN network module in a real production environment. Furthermore, by adjusting the values of λ1, λ2, and λ3, the degree of influence of different factors on the reward can be defined, allowing the model to optimize resource allocation schemes toward increasing profits, reducing costs, and minimizing delays.
[0059] Furthermore, a sliding window incremental learning method is used to update and improve the GAN network module in real time.
[0060] Specifically, the sliding window incremental learning method can use new data to update network parameters without retraining the entire network. This can improve network learning efficiency and reduce computing resource consumption while maintaining the network's adaptability to new data, making it suitable for frequently changing production environments.
[0061] This embodiment provides a data analysis model for multi-feature constraints, which can efficiently complete the resource allocation of the enterprise production process and can adapt to dynamically changing diversified production conditions, and has strong adaptability.
[0062] The above is only an embodiment of the present invention. Common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. Ordinary technicians in the relevant field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the guidance of this application. Some typical well-known structures or well-known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.
Claims
1. A data analysis model for multi-feature constraints, characterized by: Including data acquisition module, improved GAN network module and configuration optimization module; The data acquisition module is used to collect the company's production basic data and production demand data; the production basic data includes the company's production equipment, production manpower and production materials; the production demand data includes production orders; The improved GAN network module is used to generate a resource allocation plan; The improved GAN network module is equipped with an input layer, a generator network, a discriminant network, and an output layer. The input layer takes production basic data and production demand data as input. The generator network is used to generate basic configuration plans, and the discriminant network is used to discriminate the generated basic configuration plans and analyze the effectiveness of each basic configuration plan to obtain a score. The output layer is used to output the basic configuration plan with an effectiveness score higher than a threshold as the resource configuration plan. The configuration optimization module is used to impose constraints on the generation process of the generation network to optimize the generation process.
2. A data analysis model for multi-feature constraints according to claim 1, characterized in that: It also includes a pre-processing module; the pre-processing module is used to pre-process production basic data and production demand data; The preprocessing includes data cleaning, normalization and category coding.
3. The data analysis model under multi-feature constraints according to claim 1, characterized in that: The discriminant network is a binary discriminant network, including a discriminant network P and a discriminant network N; the discriminant network P and the discriminant network N are provided with the same initial loss function; the discriminant network N is provided after the generative network, and the discriminant network P is provided independently.
4. The data analysis model under multi-feature constraints according to claim 3, characterized in that: The improved GAN network module is pre-trained using a specified data set; the specified training set includes real data samples based on historical production basic data, historical production demand data and historical resource configuration data, as well as virtual data samples marked as false.
5. The data analysis model under multi-feature constraints according to claim 4, characterized in that: The improved GAN network module includes the following training steps during training: For the discriminant network, the specified data set is used as a training set to input the discriminant network P, and the discriminant network P is trained as a binary classifier; Determine a penalty factor for the binary classifier so that the probability of identifying a real data sample is as close to 1 as possible, and the probability of identifying a generated sample is as close to 0 as possible; the generated sample is generated by the generative network and is provided with a false label; When the discriminant network P is trained, its weight parameters are shared with the discriminant network N, so that the discriminant network N and the discriminant network P have the same ability to distinguish true and false samples; The false labels of the generated samples of the generator network are changed to true, and used as the training set of the discriminant network N, forcing the discriminant network N to expect the generated samples to have a probability close to 1, and not to judge the real data samples; if the discriminant network N expects the generated samples to have a probability close to 0, it will form a contrasting feedback with the former, forcing the generator network to change its parameters to generate pseudo samples that are more similar to the real data samples; The training is terminated when the output probabilities of the discriminant network P and the discriminant network N for true and false samples are close and reach 1 / 2.
6. The data analysis model under multi-feature constraints according to claim 1, characterized in that: The configuration optimization module is also used to set a penalty mechanism for the generation model; when the basic configuration solution generated by the generation model violates the constraint conditions, the penalty mechanism is triggered.
7. The data analysis model under multi-feature constraints according to claim 1, characterized in that: The constraints include: hard constraints set based on the Lagrange multiplier method - production capacity upper limit, cost threshold and order deadline.
8. The data analysis model under multi-feature constraints according to claim 1, characterized in that: The generative network is provided with a transform encoder, a UNet decoder and an attention selector.
9. The data analysis model under multi-feature constraints according to claim 1, characterized in that: The sliding window incremental learning method is used to update and improve the GAN network module in real time.