A manufacturing enterprise cost prediction method based on multi-dimensional data fusion

Abstract

The application discloses a manufacturing enterprise cost prediction method based on multi-dimensional data fusion, which is used in the technical field of mold manufacturing and comprises the following steps: integrating multi-source heterogeneous data, preprocessing the data to form unified and standardized multi-dimensional input features; constructing and training a multi-modal deep learning model based on the multi-dimensional input features; using the trained multi-modal deep learning model to infer project data in a manufacturing enterprise, generating a cost prediction report; verifying the cost prediction report according to preset business rules, and automatically correcting the cost prediction report in combination with a dynamic compensation mechanism to obtain a corrected cost prediction result; comparing the actual cost of a project with the corrected cost prediction result, analyzing errors according to a comparison result, and continuously optimizing the multi-modal deep learning model through incremental training. The cost prediction report of the application not only contains total cost, sub-item cost and confidence interval, but also provides key influence factor ranking and improvement suggestions, and the accuracy of prediction is realized.

Description

Technical Field

[0001] This application relates to the field of mold manufacturing technology, and more specifically, to a method for predicting manufacturing enterprise costs based on multi-dimensional data fusion. Background Technology

[0002] Automotive molds are specialized process equipment used for the mass production of automotive parts. They mainly include stamping, injection molding, and die casting, and are known as the mother of the automotive industry. They are characterized by large size, high precision (critical dimension tolerances are often within ±0.05mm), and long service life. Each new model requires the development of a complete set of customized molds, which is a typical high-complexity single-piece or small-batch production mode.

[0003] Cost forecasting is a core support for manufacturing companies' business decisions, directly impacting accurate pricing, project profitability, and market competitiveness. Dynamic forecasting can identify risks such as raw material fluctuations and process bottlenecks in advance, transforming cost control from post-event accounting to pre-event early warning and in-event control. At the same time, intelligent forecasting based on multi-source data fusion can accumulate expert experience and improve estimation consistency, making it a key means for companies to achieve lean management and sustainable development.

[0004] The shortcomings of existing technologies include a lack of sufficient interpretability in cost prediction results, failure to provide confidence intervals and analysis of key influencing factors; insufficient processing of multidimensional data, making it impossible to effectively integrate multimodal data; in addition, instability may occur during model training, resulting in slow convergence speed and insufficient optimization of model performance.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To overcome the above problems, this application aims to propose a cost prediction method for manufacturing enterprises based on multi-dimensional data fusion. The purpose is to address the shortcomings of existing technologies, including the lack of sufficient interpretability of cost prediction results, failure to provide confidence intervals and analysis of key influencing factors; insufficient processing of multi-dimensional data, and inability to effectively integrate multi-modal data; in addition, instability may occur during model training, and the convergence speed is slow, resulting in the model performance not being fully optimized.

[0007] Therefore, the specific technical solution adopted in this application is as follows: A cost prediction method for manufacturing enterprises based on multi-dimensional data fusion, the method comprising: S1. Integrate multi-source heterogeneous data and preprocess it to form unified and standardized multi-dimensional input features; S2. Construct and train a multimodal deep learning model based on multidimensional input features, and use the trained multimodal deep learning model to infer project data in manufacturing enterprises and generate cost prediction reports. S3. Verify the cost forecast report according to the preset business rules, and automatically correct it in combination with the dynamic compensation mechanism to obtain the corrected cost forecast result. S4. Compare the actual project cost with the revised cost prediction results, analyze the error based on the comparison results, and continuously optimize the multimodal deep learning model through incremental training.

[0008] Optionally, the preprocessed data forms a standardized multidimensional input feature, including: Extract raw data from the supplier management system, including production materials, trial molding materials management, working hours management, procurement management, inventory management, project management, quality anomalies, and financial accounts, to build a unified data lake; The raw data in the data lake is cleaned and processed to obtain the basic dataset; Based on the basic dataset, static attributes, dynamic time series, process maps and derived indicators are constructed, and the coding of categorical variables and the normalization of numerical variables are completed to form a unified and standardized multidimensional input feature.

[0009] Alternatively, the method for constructing and training a multimodal deep learning model based on multidimensional input features is as follows: Based on the static attributes in the multidimensional input features, a transformer branch is designed, a long short-term memory network branch is designed based on the dynamic time sequence, and a graph neural network branch is designed based on the process map. The feature vectors output by the three branches are then weighted and fused through an attention mechanism to obtain the fused comprehensive features. Based on the integrated features after fusion, a multi-task output layer is designed to predict the total cost, material cost, and labor cost respectively, thus constructing a multimodal deep learning model.

[0010] Alternatively, the method for obtaining the fused integrated features is as follows: Design converter branch architecture based on static attributes in multidimensional input features, design long short-term memory network branch architecture based on dynamic timing, and design graph neural network branch architecture based on process map; The multidimensional input features are input into the corresponding branches, and the static feature vector, temporal feature vector and spectral feature vector are extracted through the forward computation of each branch. The three extracted feature vectors are input into the attention mechanism layer, and weighted fusion is performed through weight calculation to output the fused comprehensive features.

[0011] Optionally, the trained multimodal deep learning model is used to infer project data in manufacturing enterprises, including: Set a weighted multi-task loss function and optimizer, and configure a cosine annealing learning rate decay strategy for the multimodal deep learning model; The configured multimodal deep learning model is trained using the fused integrated features, and the performance is monitored and the optimal multimodal deep learning model parameters are saved through the validation set. The new project's input features are fed into an optimized multimodal deep learning model for inference, and a cost prediction report is generated based on the output prediction values.

[0012] Optionally, a weighted multi-task loss function and optimizer are set, and a cosine annealing learning rate decay strategy for the multimodal deep learning model is configured, including: The main task is determined to use Hubel loss, the auxiliary task to use mean squared error loss and mean absolute error loss, and the weight coefficients of the three are set to define a weighted multi-task loss function. Based on the defined weighted multi-task loss function, the adaptive moment estimator W optimizer is selected and the initial learning rate is set. Add a cosine annealing learning rate decay strategy to the configured optimizer, and integrate the weighted multi-task loss function, the adaptive moment estimation W optimizer, and the cosine annealing learning rate decay strategy.

[0013] Optionally, a cost forecast report is generated based on the output forecast values, including: Obtain the total cost prediction, material cost prediction, labor cost prediction, and fusion weight coefficient of the attention mechanism layer from the model inference output as the basic inference data; Based on the basic inference data, the total cost confidence interval is calculated, the ranking of key influencing factors is extracted, and improvement suggestions are generated to obtain data to be reported containing complete information. Based on the data to be reported, generate a complete cost forecast report that includes the total cost confidence interval, a pie chart of cost components, a ranking of key influencing factors, and improvement suggestions.

[0014] Optionally, the data to be reported containing complete information is obtained, including: Calculate the total cost confidence interval based on the total cost forecast and confidence interval estimation method in the basic inference data; Add confidence intervals to the base inference data to obtain extended data containing confidence intervals; Interpretive analysis is performed based on the fusion weight coefficients and the importance of branch features in the extended data to extract the ranking of key influencing factors, and improvement suggestions are generated in combination with cost decomposition items. By adding the ranking of key influencing factors and improvement suggestions to the extended data, a reportable data set containing complete information is obtained.

[0015] Optionally, the formula for calculating the total cost confidence interval is: ; In the formula, Indicates the confidence interval for total cost; This represents the total cost forecast; Indicates confidence level 1- Degrees of freedom correspond t Two-sided quantiles of the distribution; This represents the total cost fusion weight coefficient of the attention mechanism layer output; This represents the total cost fusion weight coefficient of the attention mechanism layer output; Indicates the coefficient of variation of material costs; This represents the coefficient of variation for labor costs; Indicates the coefficient of variation of other manufacturing overhead costs; Indicates the number of historical cost samples; This represents the process complexity correction system for manufacturing enterprises.

[0016] Compared with the prior art, this application has the following beneficial effects: 1. This application integrates static attributes, time-series features, and process maps through a multimodal model, and combines an attention mechanism to achieve weighted feature fusion. At the same time, it uses a weighted multi-task loss function, optimizer, and learning rate strategy to ensure the model training effect. The final cost prediction report not only includes total cost, itemized cost, and confidence interval, but also provides rankings of key influencing factors and improvement suggestions, thus achieving prediction accuracy.

[0017] 2. The cost forecast report of this application not only provides predicted values ​​for total cost, material cost and labor cost, but also enhances the interpretability and credibility of the results by calculating confidence intervals and extracting key influencing factors. Combined with the fusion weight coefficient of the attention mechanism, the report can analyze the accuracy of cost forecast in detail and provide targeted improvement suggestions.

[0018] 3. This application designs different branch architectures to extract specialized feature vectors through deep processing of multi-dimensional input features, and then uses an attention mechanism for weighted fusion to ensure that the model can comprehensively consider a variety of complex correlation information and improve prediction accuracy.

[0019] 4. In the model training process of this application, a weighted multi-task loss function and a cosine annealing learning rate decay strategy are adopted to optimize the stability and convergence speed of the training process. At the same time, the validation set is used to monitor the model performance and the optimal parameters are saved for inference, ensuring the efficiency and accuracy of the final model. Attached Figure Description

[0020] The above-mentioned features, characteristics, and advantages of this application, as well as their implementation methods, will become clearer and more understandable in conjunction with the following description of the embodiments, which are illustrated in detail with reference to the accompanying drawings. Schematic diagrams are shown here: Figure 1 This is a flowchart of the manufacturing enterprise cost prediction method based on multi-dimensional data fusion in this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0022] According to embodiments of this application, a method for cost prediction in manufacturing enterprises based on multi-dimensional data fusion is provided. This method fuses static attributes, time-series features, and process maps through a multi-modal model, and combines an attention mechanism to achieve weighted feature fusion. Simultaneously, a weighted multi-task loss function, optimizer, and learning rate strategy are used to ensure model training effectiveness. The final cost prediction report not only includes total cost, itemized costs, and confidence intervals, but also provides rankings of key influencing factors and improvement suggestions, thus achieving accurate predictions. Figure 1 As shown, the method includes: S1. Integrate heterogeneous data from multiple sources and preprocess them to form unified and standardized multidimensional input features.

[0023] Preferably, the preprocessed multidimensional input features are standardized and include: Extract raw data from the supplier management system, including production materials, trial molding materials management, working hours management, procurement management, inventory management, project management, quality anomalies, and financial accounts, to build a unified data lake; The raw data in the data lake is cleaned and processed to obtain the basic dataset; Based on the basic dataset, static attributes, dynamic time series, process maps and derived indicators are constructed, and the coding of categorical variables and the normalization of numerical variables are completed to form a unified and standardized multidimensional input feature.

[0024] It should be noted that the contents of the data lake are shown in Table 1, as detailed below: Table 1 Basic Dataset name Data types Main uses Productive substances Material Bill of Materials (BOM), Standard Price, Inventory Status Provide raw material cost basis Procurement Management Purchase order, goods receipt, supplier quotation Obtain actual purchase price and cycle Inventory Management Material requisition form, material issuance form, and return form Analyze material loss rate and turnover efficiency Trial material management Trial molding consumables record Calculate the cost of non-standard materials Work Hour Management Process time, personnel efficiency, overtime situation Estimate labor costs MES system Production schedule, equipment utilization rate, downtime Reflecting manufacturing efficiency and indirect costs project management Project phases, budget allocation, and change logs Support dynamic cost adjustment Quality abnormality Number of defective products and number of rework Quantifying the cost of quality loss Financial Management Financial accounts and expense allocation Integrate with the financial system to ensure consistency. All data is accessed via API or direct database connection to form a unified data lake, providing support for subsequent modeling. Cleaning processes include: Remove duplicate orders, empty records, and outliers (such as negative unit prices); Fill in missing fields: use the mean or interpolation based on similar items; Unified unit system: The unit of weight is standardized to kg, and the unit of time is standardized to hour.

[0025] Based on the basic dataset, static attributes, dynamic time series, process maps, and derived indicators are constructed. The encoding of categorical variables and the normalization of numerical variables are completed to form a unified and standardized multidimensional input feature. The specific implementation is as follows: More than 80 high-order features were constructed, covering the following categories, as shown in Table 2, as follows: Table 2 Feature Construction Table category Specific feature examples Static properties Product model, material grade, process complexity rating, and whether it contains imported components. Dynamic Trends Average material price increase and per capita output change rate over the past 6 months Time series Daily output fluctuations, monthly labor cost growth rate Relationship Correlation between historical failure rates of a certain process and a certain piece of equipment Derivative Indicators Unit labor cost = Total labor cost / Total labor hours; Material utilization rate = Actual consumption / Theoretical consumption; First pass rate = Number of qualified products / Total production quantity The coding and standardization are as follows: 1. Categorical variables: Use One-Hot coding or Embedding; 2. Numerical variables: Z-score normalization or Min-Max standardization; 3. Time series: Slide a window to slice (e.g., T=30 days) to form a fixed-length input sequence.

[0026] S2. Construct and train a multimodal deep learning model based on multidimensional input features, and use the trained multimodal deep learning model to infer project data in manufacturing enterprises and generate cost prediction reports.

[0027] Preferably, the method for constructing and training a multimodal deep learning model based on multidimensional input features is as follows: Based on the static attributes in the multidimensional input features, a transformer branch is designed, a long short-term memory network branch is designed based on the dynamic time sequence, and a graph neural network branch is designed based on the process map. The feature vectors output by the three branches are then weighted and fused through an attention mechanism to obtain the fused comprehensive features. Based on the integrated features after fusion, a multi-task output layer is designed to predict the total cost, material cost, and labor cost respectively, thus constructing a multimodal deep learning model.

[0028] It should be noted that this application adopts a deep learning framework of dual-stream fusion + multi-task learning, with the following specific structure: [Input Layer] [Branch 1: Transformer] → [Global Feature Understanding] [Branch 2: LSTM] → [Time Evolution Modeling] [Branch 3: GNN] → [Process Path Map Analysis] [Feature Fusion Layer] → [Weighted Concatenation] [Main Task: Total Cost Forecasting] [Side Task 1: Material Cost Forecasting] [Auxiliary Task 2: Labor Cost Forecasting] [Output layer]; The specific contents of the transformer branch, the long short-term memory network branch, and the graph neural network branch are as follows: 1. The Transformer branch contains the following details: Input: Static feature vectors of the current project (such as BOM, process route); Function: Capture the dependencies between different components and identify key components that affect the total cost; Example: It was found that "sloping top mechanism" and "hot runner system" often appear together, and both of them drive up costs.

[0029] 2. The specific content of the Long Short-Term Memory (LSTM) branch is as follows: Input: Historical cost time series (e.g., actual cost of similar products over the past 12 months). Function: Models cost trends over time, identifying seasonal fluctuations, raw material price cycles, etc. Use bidirectional LSTM to improve context awareness.

[0030] 3. The specific content of the Graph Neural Network (GNN) branch is as follows: Input: Process flow diagram (nodes represent processes, edges represent their sequence); Function: Analyze the coupling effect between processes, such as "grinding must be performed after electro-erosion", skipping it may lead to increased rework costs; Cost-influencing factors are disseminated through messaging mechanisms.

[0031] The purpose is to obtain the predicted total cost, predicted material cost, and predicted labor cost.

[0032] Preferably, the method for obtaining the fused integrated features is as follows: Design converter branch architecture based on static attributes in multidimensional input features, design long short-term memory network branch architecture based on dynamic timing, and design graph neural network branch architecture based on process map; The multidimensional input features are input into the corresponding branches, and the static feature vector, temporal feature vector and spectral feature vector are extracted through the forward computation of each branch. The three extracted feature vectors are input into the attention mechanism layer, and weighted fusion is performed through weight calculation to output the fused comprehensive features.

[0033] Preferably, the trained multimodal deep learning model is used to infer project data in manufacturing enterprises, including: Set a weighted multi-task loss function and optimizer, and configure a cosine annealing learning rate decay strategy for the multimodal deep learning model; The configured multimodal deep learning model is trained using the fused integrated features, and the performance is monitored and the optimal multimodal deep learning model parameters are saved through the validation set. The new project's input features are fed into an optimized multimodal deep learning model for inference, and a cost prediction report is generated based on the output prediction values.

[0034] It should be explained that the training process is as follows: A weighted multi-task loss function is used: ; in: ; ; ; Weighting coefficient : : =0.6:0.2:0.2; Optimizers and hyperparameters; Optimizer: AdamW, initial learning rate 1e−4; BatchSize: 64; Epochs: 200; Dropout: 0.3; Learning rate decay strategy: Cosine Annealing with Warm Restarts; The training environment is as follows: (1) GPU acceleration: NVIDIA A100×2; (2) Framework: PyTorch + Lightning; (3) Data partitioning: training set (70%), validation set (15%), test set (15%).

[0035] The above parameters are explained as follows: This represents the total loss function, which is a weighted sum for multi-task learning. The weighting coefficient represents the loss of the main task (total cost forecasting); The weighting coefficient represents the loss from auxiliary task 1 (material cost prediction); The weighting coefficients represent the losses from auxiliary task 2 (labor cost prediction); The loss of the main task is usually represented by Huber Loss. The loss from the material cost forecasting task is represented by mean squared error (MSE). This indicates the loss from the labor cost forecasting task (the specific type was not specified in the original text). HuberLoss represents the Huber loss function, which combines mean squared error and absolute error, making it more robust to outliers. MSE stands for Mean Squared Error, a commonly used loss function in regression tasks. AdamW represents the adaptive moment estimator W optimizer, an improvement on Adam that introduces decoupled weight decay; BatchSize represents the batch size, which is the number of samples used in each training iteration. Epochs represents the number of training rounds, the number of times the complete dataset is traversed; Dropout refers to the dropout rate, which randomly ignores some neurons during training to prevent overfitting. Cosine AnnealingwithWarmRestarts represents a cosine annealing with warm restart learning rate decay strategy, causing the learning rate to change periodically; GPU: Graphics Processing Unit, used to accelerate deep learning computations; NVIDIA A100 indicates that the GPU is an A100 model. PyTorch stands for PyTorch, an open-source deep learning framework based on Python. Lightning stands for PyTorchLightning, a library that simplifies the model training and validation process.

[0036] Preferably, setting a weighted multi-task loss function and optimizer, and configuring a cosine annealing learning rate decay strategy for the multimodal deep learning model, includes: The main task is determined to use Hubel loss, the auxiliary task to use mean squared error loss and mean absolute error loss, and the weight coefficients of the three are set to define a weighted multi-task loss function. Based on the defined weighted multi-task loss function, the adaptive moment estimator W optimizer is selected and the initial learning rate is set. Add a cosine annealing learning rate decay strategy to the configured optimizer, and integrate the weighted multi-task loss function, the adaptive moment estimation W optimizer, and the cosine annealing learning rate decay strategy.

[0037] Preferably, a cost forecast report is generated based on the output forecast values, including: Obtain the total cost prediction, material cost prediction, labor cost prediction, and fusion weight coefficient of the attention mechanism layer from the model inference output as the basic inference data; Based on the basic inference data, the total cost confidence interval is calculated, the ranking of key influencing factors is extracted, and improvement suggestions are generated to obtain data to be reported containing complete information. Based on the data to be reported, generate a complete cost forecast report that includes the total cost confidence interval, a pie chart of cost components, a ranking of key influencing factors, and improvement suggestions.

[0038] Preferably, the data to be reported, containing complete information, includes: Calculate the total cost confidence interval based on the total cost forecast and confidence interval estimation method in the basic inference data; Add confidence intervals to the base inference data to obtain extended data containing confidence intervals; Interpretive analysis is performed based on the fusion weight coefficients and the importance of branch features in the extended data to extract the ranking of key influencing factors, and improvement suggestions are generated in combination with cost decomposition items. By adding the ranking of key influencing factors and improvement suggestions to the extended data, a reportable data set containing complete information is obtained.

[0039] Preferably, the formula for calculating the total cost confidence interval is: ; In the formula, Indicates the confidence interval for total cost; This represents the total cost forecast; Indicates confidence level 1- Degrees of freedom correspond t Two-sided quantiles of the distribution; This represents the total cost fusion weight coefficient of the attention mechanism layer output; This represents the total cost fusion weight coefficient of the attention mechanism layer output; Indicates the coefficient of variation of material costs; This represents the coefficient of variation for labor costs; Indicates the coefficient of variation of other manufacturing overhead costs; Indicates the number of historical cost samples; This represents the process complexity correction system for manufacturing enterprises.

[0040] It is necessary to explain the specific implementation steps of end-to-end cost forecasting. 1. The specific prediction steps are as follows: Step 1: Input Initialization Users create a new cost forecasting task in the "Cost Management" module of the SAM system and fill in the following information: Project number; Product model; Business unit; Expected production start date; Preliminary BOM list; Process route sketch (optional); 2. Automatically retrieve the following from the database: (1) Current standard price of materials; (2) Cost data of similar historical projects; (3) Standard working hours for relevant processes; (4) Supplier delivery cycle; (5) Equipment availability status; Based on the above, the generated features are as follows: Automatically generate an input vector X with 100+ dimensions, and organize it in the following format: Static features ∈ R 50 ; Time series ∈ R T×30 ; The number of nodes in the graph structure is ≤20; The Transformer branch processes static features and outputs a global context representation. LSTM branches process time series data and output long-term trend features; GNN branches analyze the process diagram and output the influence weights between processes; The three features are weighted and fused using an attention mechanism; Generate a "Cost Forecast Report", which includes: (1) Total cost forecast (including confidence interval); (2) Cost breakdown pie chart (materials, labor, management, quality losses, etc.); (3) Ranking of key influencing factors (Top 5); (4) Recommended improvement points (e.g., "reducing the time for a certain process can save 12%").

[0041] In addition, this application incorporates multiple judgment logics to ensure that the prediction results are reasonable and reliable; like If the cost is less than 0.7 × min (historical cost), a warning will be issued: the cost is abnormally low. Please check the completeness of the BOM (Bill of Materials). like If the value is 1.5 × max (historical cost), it indicates a significant risk and recommends a reassessment of the process plan. If material costs account for more than 60% of total costs, it indicates that material costs are too high and alternative materials should be considered. If the growth rate of labor costs exceeds 20% per month, it indicates a labor shortage and the need to coordinate production schedules.

[0042] S3. Verify the cost forecast report according to the preset business rules, and automatically correct it in combination with the dynamic compensation mechanism to obtain the corrected cost forecast result.

[0043] It should be explained that, in order to cope with uncertainties and external disturbances, this application introduces a variety of compensation mechanisms: 1. Supply chain risk compensation: If the supplier's fulfillment rate is less than 80%, increase the procurement premium by 5%-10%; If international logistics are involved, an exchange rate fluctuation factor (default +8%) will be added. 2. Compensation for process changes: When the process route changes (such as adding a heat treatment step), the system automatically identifies the change and adds the corresponding cost. It supports manually adding "temporary expense items" and updating the model input synchronously; 3. Compensation for quality loss: Calculate the average defect rate for this type of product based on historical data from the "Quality Anomalies" module; If the expected defect rate is >10%, then add a corresponding proportion of scrap costs to the total cost; 4. Compensation for production capacity bottlenecks: If MES data shows that the target equipment load has reached 90%, then overtime pay will be increased by 20% in labor costs. If the production schedule is extended by more than 10 days, the management fee will be automatically increased. 5. Manual intervention interface: Experienced cost engineers are allowed to revise the forecast results and indicate the reasons for the revisions; The modified data will be used as negative samples for model retraining, forming a closed-loop feedback.

[0044] S4. Compare the actual project cost with the revised cost prediction results, analyze the error based on the comparison results, and continuously optimize the multimodal deep learning model through incremental training.

[0045] It should be explained that this application establishes a complete model self-evolution mechanism: Upon completion of each project, the system automatically collects the final settlement cost; compares the actual cost with the predicted value and calculates the error; runs an incremental training program once a month to fine-tune the model using new data; sets up an "abnormal case pool" to store projects with deviations greater than 15% for expert review; and supports an online learning mode to gradually absorb new knowledge without affecting service.

[0046] It should be noted that the calculation formulas and all parameters involved in the calculations in this application have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0047] Although the present application has disclosed the preferred embodiments above, the embodiments are merely examples for the purpose of illustration and are not intended to limit the present application. Those skilled in the art can make some modifications and refinements without departing from the spirit and scope of the present application. The scope of protection claimed by the present application should be determined by the claims.

Claims

1. A method for cost prediction in manufacturing enterprises based on multi-dimensional data fusion, characterized in that, The method includes: S1. Integrate multi-source heterogeneous data and preprocess it to form unified and standardized multi-dimensional input features; S2. Construct and train a multimodal deep learning model based on multidimensional input features, and use the trained multimodal deep learning model to infer project data in manufacturing enterprises and generate cost prediction reports. S3. Verify the cost forecast report according to the preset business rules, and automatically correct it in combination with the dynamic compensation mechanism to obtain the corrected cost forecast result. S4. Compare the actual project cost with the revised cost prediction results, analyze the error based on the comparison results, and continuously optimize the multimodal deep learning model through incremental training.

2. The manufacturing enterprise cost forecasting method according to claim 1, characterized in that, The preprocessed, standardized multidimensional input features include: Extract raw data from the supplier management system, including production materials, trial molding materials management, working hours management, procurement management, inventory management, project management, quality anomalies, and financial accounts, to build a unified data lake; The raw data in the data lake is cleaned and processed to obtain the basic dataset; Based on the basic dataset, static attributes, dynamic time series, process maps and derived indicators are constructed, and the coding of categorical variables and the normalization of numerical variables are completed to form a unified and standardized multidimensional input feature.

3. The manufacturing enterprise cost forecasting method according to claim 1, characterized in that, The method for constructing and training a multimodal deep learning model based on multidimensional input features is as follows: Based on the static attributes in the multidimensional input features, a transformer branch is designed, a long short-term memory network branch is designed based on the dynamic time sequence, and a graph neural network branch is designed based on the process map. The feature vectors output by the three branches are then weighted and fused through an attention mechanism to obtain the fused comprehensive features. Based on the integrated features after fusion, a multi-task output layer is designed to predict the total cost, material cost, and labor cost respectively, thus constructing a multimodal deep learning model.

4. The manufacturing enterprise cost forecasting method according to claim 3, characterized in that, The method for obtaining the fused comprehensive features is as follows: Design converter branch architecture based on static attributes in multidimensional input features, design long short-term memory network branch architecture based on dynamic timing, and design graph neural network branch architecture based on process map; The multidimensional input features are input into the corresponding branches, and the static feature vector, temporal feature vector and spectral feature vector are extracted through the forward computation of each branch. The three extracted feature vectors are input into the attention mechanism layer, and weighted fusion is performed through weight calculation to output the fused comprehensive features.

5. The manufacturing enterprise cost forecasting method according to claim 4, characterized in that, The method of using a trained multimodal deep learning model to infer project data in manufacturing enterprises includes: Set a weighted multi-task loss function and optimizer, and configure a cosine annealing learning rate decay strategy for the multimodal deep learning model; The configured multimodal deep learning model is trained using the fused integrated features, and the performance is monitored and the optimal multimodal deep learning model parameters are saved through the validation set. The new project's input features are fed into an optimized multimodal deep learning model for inference, and a cost prediction report is generated based on the output prediction values.

6. The manufacturing enterprise cost forecasting method according to claim 5, characterized in that, The setting of the weighted multi-task loss function and optimizer, and the configuration of the cosine annealing learning rate decay strategy for the multimodal deep learning model, include: The main task is determined to use Hubel loss, the auxiliary task to use mean squared error loss and mean absolute error loss, and the weight coefficients of the three are set to define a weighted multi-task loss function. Based on the defined weighted multi-task loss function, the adaptive moment estimator W optimizer is selected and the initial learning rate is set. Add a cosine annealing learning rate decay strategy to the configured optimizer, and integrate the weighted multi-task loss function, the adaptive moment estimation W optimizer, and the cosine annealing learning rate decay strategy.

7. The manufacturing enterprise cost forecasting method according to claim 5, characterized in that, The process of generating a cost forecast report based on the output forecast values ​​includes: Obtain the total cost prediction, material cost prediction, labor cost prediction, and fusion weight coefficient of the attention mechanism layer from the model inference output as the basic inference data; Based on the basic inference data, the total cost confidence interval is calculated, the ranking of key influencing factors is extracted, and improvement suggestions are generated to obtain data to be reported containing complete information. Based on the data to be reported, generate a complete cost forecast report that includes the total cost confidence interval, a pie chart of cost components, a ranking of key influencing factors, and improvement suggestions.

8. The manufacturing enterprise cost forecasting method according to claim 7, characterized in that, The obtained reportable data containing complete information includes: Calculate the total cost confidence interval based on the total cost forecast and confidence interval estimation method in the basic inference data; Add confidence intervals to the base inference data to obtain extended data containing confidence intervals; Interpretive analysis is performed based on the fusion weight coefficients and the importance of branch features in the extended data to extract the ranking of key influencing factors, and improvement suggestions are generated in combination with cost decomposition items. By adding the ranking of key influencing factors and improvement suggestions to the extended data, a reportable data set containing complete information is obtained.

9. The manufacturing enterprise cost forecasting method according to claim 8, characterized in that, The formula for calculating the total cost confidence interval is as follows: ； In the formula, Indicates the confidence interval for total cost; This represents the total cost forecast; Indicates confidence level 1- Degrees of freedom correspond t Two-sided quantiles of the distribution; This represents the total cost fusion weight coefficient of the attention mechanism layer output; This represents the total cost fusion weight coefficient of the attention mechanism layer output; Indicates the coefficient of variation of material costs; This represents the coefficient of variation for labor costs; Indicates the coefficient of variation of other manufacturing overhead costs; Indicates the number of historical cost samples; This represents the process complexity correction system for manufacturing enterprises.

Images