A method for predicting mold pricing based on multi-dimensional data fusion

Abstract

This invention relates to the field of mold price prediction technology, and in particular to a mold price prediction method based on multi-dimensional data fusion. A hybrid neural network model combining Transformer and LSTM is constructed. The global attention mechanism of the Transformer module is used to deeply mine the complex relationships between high-dimensional static features, while the LSTM module captures the temporal evolution patterns of historical prices. The two types of features are effectively fused, significantly improving the accuracy and stability of mold price prediction. Through a pre-processing layer and attention heads that process features of specific categories, the model is guided to perform structured modeling based on feature categories in the early stages of learning. This effectively distinguishes and focuses on the dependencies within different categories of features, such as static attributes, process attributes, and temporal costs, reducing noise interference between irrelevant features and improving the interpretability, efficiency, and generalization ability of the model's feature association learning.

Description

Technical Field

[0001] This invention relates to the field of mold price prediction technology, and in particular to a mold price prediction method based on multi-dimensional data fusion. Background Technology

[0002] As a fundamental process equipment in industrial production, the accuracy and timeliness of mold pricing directly affect a company's cost control and efficiency. Mold pricing is mainly determined by two categories of characteristics: one is the static attribute characteristics of the mold, such as mold type, material, geometric dimensions, structural complexity (e.g., whether it contains inclined ejectors and sliders), and process requirements (e.g., heat treatment grade). These characteristics contain a mixture of continuous and discrete values, and there are complex interactions and nonlinear patterns between them. For example, the processing cost of a specific material under a specific structure can increase nonlinearly. The other category is dynamic time-series characteristics, namely the historical pricing sequence of the mold, which reflects the combined effects of various macroeconomic factors such as raw material market fluctuations, industry capacity cycles, and supply chain cost changes.

[0003] Existing methods based on human experience are difficult to quantify the complex relationships between various features. Traditional time series models (such as ARIMA models) are difficult to incorporate the complex static attribute features of molds. If a simple machine learning model simply concatenates two types of features as input, it will be difficult to extract their inherent patterns and effectively fuse them due to the limitations of the model structure, resulting in unstable prediction accuracy of mold pricing. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a mold pricing prediction method based on multi-dimensional data fusion. This method solves the technical problem that existing technologies struggle to simultaneously capture and organically couple the static attribute features and dynamic temporal features of molds, leading to unstable accuracy in mold pricing prediction.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a mold pricing prediction method based on multi-dimensional data fusion, which specifically includes the following steps: S1. Collect historical mold pricing sequences and multi-dimensional feature data for predicting mold pricing at each time point to construct a training dataset; S2. Construct a mold price prediction model that includes a feature encoding module, a Transformer module, an LSTM module, and a fusion output module, and train the model using the training dataset; The feature encoding module is used to encode multi-dimensional feature data and historical mold quotation sequences; The Transformer module is used to process encoded multi-dimensional feature data through multiple self-attention blocks to learn global context features that characterize the importance weights among the features. The LSTM module is used to extract overall temporal features from historical mold quotation sequences; The fusion output module is used to fuse global context features and overall temporal features, and calculate the mold quotation based on the fusion result; S3. Obtain multi-dimensional data of the mold to be predicted and input it into the trained mold price prediction model to obtain the corresponding mold price.

[0006] Preferably, the feature encoding module is used to construct preset derived features based on multi-dimensional data, divide the multi-dimensional feature data and derived features into multiple data categories that include at least categorical features and continuous features, encode multiple features according to the preset encoding rules of different data categories, and divide the historical mold quotation sequence into sub-time series with fixed time lengths.

[0007] Preferably, the Transformer module includes multiple self-attention blocks stacked sequentially, and each self-attention block is sequentially provided with a multi-head attention mechanism, a first residual connection module, a feedforward neural network, and a second residual connection module; The multi-head attention mechanism is used to calculate the local context features in the input features that represent the importance weights between the features, and to concatenate and linearly transform the local context features output by each attention head to obtain the attention output features. The first residual connection module is used to add the attention output features to the input features of the multi-head attention mechanism and perform layer normalization to obtain the first normalized features; The feedforward neural network is used to perform a nonlinear transformation on the first normalized feature to obtain a nonlinear transformed feature. The second residual connection module is used to add the first normalized feature and the nonlinear transformation feature, and perform layer normalization to obtain the second normalized feature; The second normalized feature output from the last self-attention block is labeled as a global context feature.

[0008] Preferably, the Transformer module further includes a pre-processing layer set at the input end of the self-attention block. The pre-processing layer is used to divide the input features into feature categories that include at least static attribute features, process attribute features and time-series cost features according to their physical meaning, and to map the features of different feature categories to the corresponding feature subspaces through different linear transformation layers. Then, the features in each feature subspace are concatenated to form an initial feature matrix for inputting the first multi-head attention mechanism. The first multi-head attention mechanism includes multiple feature attention heads for processing corresponding feature categories. Each feature attention head calculates local context features in the initial feature matrix to represent the attention weights between features of the corresponding feature category. Then, the local context features output by each feature attention head are concatenated and linearly transformed to obtain the attention output features.

[0009] Preferably, the LSTM module is a bidirectional LSTM network, and the specific steps for extracting the overall time-series features from the historical mold quotation sequence are as follows: The forward and backward hidden state sequences are extracted from the input historical mold price quotation sequence through the forward and backward propagation layers of a bidirectional LSTM network, and the tanh function is used as the activation function inside both. By concatenating the forward hidden state and the backward hidden state at each time step, the temporal evolution features are obtained. Extract the feature vector of the last time step in the temporal evolution features to obtain the overall temporal features.

[0010] Preferably, the specific steps for calculating the mold quotation are as follows: The overall temporal features and global contextual features are concatenated to obtain fused features; The fused features are processed through at least two fully connected network layers, and Dropout is added to prevent overfitting to obtain the predicted quotes and corresponding prediction confidence intervals. The mold price is calculated based on the predicted price and the preset first price adjustment rule, and the mold price and the predicted confidence interval are output.

[0011] Preferably, the first price adjustment rule includes at least the following: Calculate the low price threshold below the historical lowest price based on the historical lowest price. If the predicted price is lower than the low price threshold, issue the first alarm signal and output the predicted price as the mold price, while also outputting the corresponding prediction confidence interval. Calculate the high price threshold that is higher than the historical high price based on the highest historical price. If the predicted price is higher than the high price threshold, issue a second alarm signal and output the predicted price as the mold price, and output the corresponding prediction confidence interval. Calculate the similarity between the current mold and historical molds, and mark the mold with a similarity greater than the similarity threshold and the closest time to the current mold quotation as a similar mold. When the error between the predicted quotation of the current mold and the mold quotation of the similar mold is greater than the preset error value, issue a third alarm signal and output the predicted quotation as the mold quotation, and output the corresponding prediction confidence interval. If the current mold order is urgent, an expedited fee will be added to the predicted quote to obtain the mold quote and the corresponding prediction confidence interval. If the current mold involves imported or exported materials, then the tariff will be added to the predicted price to obtain the mold price and the corresponding prediction confidence interval; If the price fluctuation of raw materials exceeds the preset material price fluctuation range, the raw material fluctuation fee will be added to the predicted price to obtain the mold price and the corresponding prediction confidence interval. If the average failure rate of this type of mold trial exceeds the preset failure rate threshold, then a processing fee reserve will be added to the predicted quotation to obtain the mold quotation and the corresponding prediction confidence interval.

[0012] Preferably, in step S3, after acquiring multi-dimensional data of the mold to be predicted and inputting it into the trained mold price prediction model to obtain the corresponding mold price, the method further includes: obtaining multiple suppliers with the highest comprehensive scores as candidate suppliers according to preset supplier scoring rules, and then calculating the final mold price after selecting each candidate supplier according to a preset compensation mechanism; the compensation mechanism is as follows: If the historical fulfillment rate of the candidate supplier is less than the preset value, a risk premium will be added to the mold quotation to form the final mold quotation. If the number of collaborations with candidate suppliers is less than the preset number of collaborations, a safety margin cost will be added to the mold quotation to form the final mold quotation.

[0013] By employing the above technical solution, the present invention provides a method for predicting mold pricing based on multi-dimensional data fusion, which has at least the following beneficial effects: 1. This invention constructs a hybrid neural network model that combines Transformer and LSTM. By using the global attention mechanism of the Transformer module to deeply mine the complex relationships between high-dimensional static features, and at the same time using the LSTM module to capture the temporal evolution of historical quotations, the two types of features are effectively fused together, which solves the problem that traditional single models are difficult to process static and dynamic data in a coordinated manner, thereby significantly improving the overall accuracy and stability of mold quotation prediction.

[0014] 2. This invention guides the model to perform structured modeling based on the category of features in the early stage of learning by designing a pre-processing layer and attention heads that process features of specific categories. This effectively distinguishes and focuses on the dependencies within different categories of features such as static attributes, process attributes, and time-series costs, reduces noise interference between irrelevant features, and improves the interpretability, efficiency, and generalization ability of the model's feature association learning.

[0015] 3. By introducing a quotation adjustment rule with multiple verification logic and a supplier risk compensation mechanism, this invention combines data-driven forecasting with business rules and intelligent risk management. It achieves automatic early warning of abnormal quotations, dynamic compensation for specific cost items (such as expedited or imported materials), quantitative assessment of the performance risks of different suppliers, and finally outputs decision support information with confidence intervals and risk adjustments, which greatly improves the practicality and business reliability of the forecast results. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the mold pricing prediction method based on multi-dimensional data fusion according to the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0018] To address the technical problem of unstable mold price prediction accuracy caused by the inability of existing technologies to simultaneously capture and organically couple the static attribute features and dynamic temporal features of molds, and considering the significant limitations of traditional single neural network models in processing mold price prediction tasks (e.g., LSTM alone is insufficient to effectively model the nonlinear coupling between different static attribute features, Transformer alone lacks sensitivity to time series evolution patterns, and simple stacking or concatenation of the two structures can easily lead to information redundancy or gradient vanishing), this invention provides a mold price prediction method based on multi-dimensional data fusion. It constructs a price prediction model that combines the global attention mechanism of Transformer with the temporal modeling capabilities of LSTM. The Transformer's global attention mechanism captures the complex relationships between static attribute features, while LSTM processes the time series variation patterns of historical price data. The outputs of both are weighted and fused before being fed into a regression head, ultimately generating accurate price prediction results. This significantly improves the model's generalization ability and prediction stability in nonlinear, high-dimensional, and heterogeneous data environments. Figure 1 As shown, the method specifically includes the following steps: S1. First, for each completed mold project, collect its historical quotation sequence. Usually, this sequence is arranged in chronological order, such as recording the average market transaction price or actual quotation of various molds in each quarter over the past 36 months. Then, multi-dimensional feature data for predicting mold quotations at each moment in the historical mold quotation sequence are acquired synchronously. The multi-dimensional feature data includes material standard prices, procurement records, inventory status, supplier ratings, mold BOM, trial mold material usage, etc., to construct a training dataset. This data mainly comes from the enterprise's internal SAM (Supply Chain and Production Management) system, ERP (Enterprise Resource Planning) system, and MES (Manufacturing Execution System). These systems have accumulated data on the entire lifecycle of molds from design, procurement, manufacturing to delivery. Taking a certain SAM system as an example, parameters such as material configuration table, material density, heat treatment requirements, and surface treatment grade can be obtained from the basic configuration management module of SAM. The benchmark unit price of mainstream materials in the current market can be extracted from the material standard price module. Information such as historical purchase orders, trial mold material requisition forms, actual purchase prices, and payment cycles can be collected from the procurement management module. Operation records such as receiving slips, material requisition forms, material issuance forms, and return slips can be read from the inventory management module to analyze material flow efficiency and loss rate. The quantity of special materials consumed during the trial molding process and their corresponding costs can be counted from the trial mold material management module. All of these data from the SAM system can be used as multi-dimensional data.

[0019] S2. Construct a mold pricing prediction model that includes a feature encoding module, a Transformer module, an LSTM module, and a fusion output module. Train the model using a training dataset. The detailed functions and processing procedures of each module are described below: The feature encoding module is used to encode multi-dimensional feature data and historical mold quotation sequences. Different encoding rules are used for different features, so they first need to be classified. Before classification, some features are difficult to obtain directly and need to be calculated and transformed based on the acquired raw multi-dimensional data. Therefore, the feature encoding module needs to construct preset derived features based on the multi-dimensional data, such as average processing cost per kilogram, trial molding times / total working hours ratio, and supplier response delay days. Then, the multi-dimensional feature data and derived features are divided into multiple data categories, each containing at least categorical and continuous features. Multiple features are then encoded according to the preset encoding rules for different data categories. For example, categorical variables (such as mold type, material number, and supplier ID) are vectorized using One-Hot encoding or Embedding, while continuous variables are processed using Min-Max normalization or Z-score standardization. For instance, H13 steel is mapped to a 512-dimensional dense vector, with the dimension length normalized to the [0,1] interval. The historical mold quotation sequence is divided into sub-time series with a fixed time length (e.g., 12 months).

[0020] Traditional fully connected networks struggle to explicitly model the complex, non-linear interactions between static attribute features. For example, the choice of a specific material can significantly impact the cost of subsequent heat treatment and polishing processes. The Transformer module, based on a self-attention mechanism, can automatically identify which feature combinations have the most critical impact on the final price. For instance, it identifies using H13 material and requiring mirror polishing as a high-cost feature combination. The output of the Transformer module is a feature representation rich in global contextual information, encoding the joint effect of all static features after complex interactions. Therefore, the Transformer module processes encoded multi-dimensional feature data through multiple self-attention blocks to learn global contextual features that characterize the importance weights between features. Specifically, the Transformer module includes multiple stacked self-attention blocks, each containing a multi-head attention mechanism, a first residual connection module, a feedforward neural network, and a second residual connection module. The components of each self-attention block are further explained below: 1. The multi-head attention mechanism is used to calculate the local context features in the input features that represent the importance weights between the features, and to identify the key factor combinations that affect the price. The local context features output by each attention head are concatenated and linearly transformed to obtain the attention output features.

[0021] 2. The first residual connection module is used to add the attention output features to the input features of the multi-head attention mechanism and perform layer normalization to obtain the first normalized features.

[0022] 3. The feedforward neural network is used to perform a nonlinear transformation on the first normalized feature to obtain the nonlinear transformed feature.

[0023] 4. The second residual connection module is used to add the first normalized feature and the nonlinear transformation feature, and perform layer normalization to obtain the second normalized feature, and mark the second normalized feature output by the last self-attention block as the global context feature.

[0024] The hyperparameters of the Transformer module can be set as follows: model dimensions Multi-head attention head count h=8, feedforward network hidden layer size The number of layers L=6, and the dropout rate=0.1.

[0025] Furthermore, considering that the original, encoded feature vectors, although uniform in format, contain features with drastically different physical meanings (such as material, size, and supplier), feeding these features indiscriminately into the attention mechanism could lead to the model expending significant computational resources to learn associations that shouldn't exist between these features. For example, meaninglessly associating supplier ID with mold length. Therefore, the Transformer module is further improved by including a pre-processing layer at the input of the self-attention block. This pre-processing layer categorizes the input features according to their physical meaning into feature categories that at least include static attribute features, process attribute features, and time-series cost features. Static attribute features refer to the inherent and unchanging properties of the mold (such as material codes like H13 and P20), basic mold types (such as injection molds and stamping molds), and basic geometric shape classifications. Process attribute features refer to attributes related to the manufacturing process, such as a list of machining operations (milling, drilling, and EDM), surface treatment requirements (such as mirror polishing), and precision levels. Time-series cost features refer to attributes related to time and market fluctuations, such as the historical purchase price sequence of materials and current market index factors. Different linear transformation layers are used to map features of different feature categories to corresponding feature subspaces. The features in each feature subspace are then concatenated to form the initial feature matrix used as input to the first multi-head attention mechanism.

[0026] The first multi-head attention mechanism includes multiple feature attention heads for processing corresponding feature categories. Each feature attention head uses different linear transformation layers to process the features of these categories, allowing the features of different categories to be distributed in different, more specific vector subspaces after transformation, facilitating further processing. Specifically, each feature attention head calculates the local context features in the initial feature matrix that represent the attention weights between features of the corresponding feature category. For example, the first head mainly focuses on features within the subspace of static attribute features, calculating the attention weights between these features to learn how static factors such as material and mold type affect cost. Its calculation process is similar to standard attention, but its query, key, and value vectors are mainly or entirely generated by the features of the corresponding subspace. This design forces the model to first deeply explore the inherent patterns within similar features in the initial learning stage, such as the cost synergy between different process steps, rather than prematurely getting bogged down in noisy associations between cross-category features. This improves the accuracy and interpretability of feature association learning. Then, the local context features output by each feature attention head are concatenated and linearly transformed to obtain the attention output features.

[0027] After the initial multi-head attention mechanism, the output features are already the result of preliminary fusion of information from various categories. For subsequent stacked self-attention blocks with more general structures, their inputs are already high-quality, semantically rich fused features. Therefore, they can focus on learning deeper and more complex cross-category feature interaction patterns. Through this structured Transformer module, the blindness of Transformer module learning is effectively reduced, training convergence is accelerated, and the importance weights learned in the end are more in line with business logic, enhancing the robustness and reliability of the entire prediction model.

[0028] Considering that mold pricing is affected by factors such as macroeconomic conditions and material price fluctuations, exhibiting obvious time trends and cyclical patterns, a sequence model capable of capturing long-term dependencies is needed. LSTM, through its internal gating mechanism (i.e., input gate, forget gate, and output gate), can selectively remember or forget information in the sequence, thereby effectively capturing long-term trends, cyclical fluctuations, and other temporal patterns in pricing data. The output of the LSTM module is a feature representation that condenses the historical price evolution pattern, reflecting the macroeconomic cost environment in which the current pricing occurs. Therefore, the LSTM module is used to extract overall temporal features from historical mold pricing sequences.

[0029] However, mold pricing is not only influenced by past history (such as the increase in material costs in the past few months), but may also implicitly contain expectations of near-term future trends (such as the upcoming industry peak season). A bidirectional structure can scan the sequence simultaneously from both forward and backward directions, capturing this contextual information more comprehensively. To effectively capture the long-term trend, periodic fluctuations, and contextual dependencies of mold pricing over time, a bidirectional LSTM network is set as the LSTM module. The specific steps for extracting the overall temporal features from the historical mold pricing sequence are as follows: First, the forward and backward propagation layers of the bidirectional LSTM network are used to extract the forward hidden state sequence and the backward hidden state sequence from the input historical mold pricing sequence, respectively. The hidden state sequence consists of a forward hidden state sequence that encodes evolutionary information from the past to the present, and a backward hidden state sequence that encodes contextual information from the future to the present. The tanh function is used as the activation function within both sequences. To fuse the complementary information extracted from the forward and backward layers and form a complete contextual representation for each time step, the forward and backward hidden states of each time step are concatenated to obtain the temporal evolution features. The hidden state of the last time step theoretically contains the information encoded and passed on from all previous time steps, serving as a summary of the entire sequence. Therefore, the feature vector of the last time step is extracted from the temporal evolution features to obtain the overall temporal features.

[0030] The final price of a mold is the result of the sum of static material and process costs and dynamic market environment costs. For example, the price of a mold with a complex structure (static feature) will be significantly higher during periods of soaring raw material prices (temporal feature) than during periods of stable material prices. Simple feature splicing cannot capture this cross-modal interaction effect, and a special fusion and decision layer needs to be designed. Therefore, the fusion output module is used to fuse global context features and global temporal features, and calculate the mold price based on the fusion result. The fusion process can be carried out in the following steps: First, the global temporal features and global context features are spliced ​​to obtain a richer fused feature. Then, the fused feature is processed through at least two fully connected network layers. The first fully connected network layer can compress the original 512-dimensional feature to 256 dimensions, and then the second fully connected network layer compresses it to 1 dimension to obtain the predicted price. Dropout is also added in this process to prevent overfitting.

[0031] Finally, the mold price is calculated based on the predicted price and the preset first price adjustment rule, and the mold price and the predicted confidence interval are output. Monte Carlo Dropout or multi-task learning can be used to output the predicted confidence interval (e.g., 95% CI). The first price adjustment rule can be either not set (i.e., no first price adjustment rule is set, and the predicted price is directly output as the mold price), or the predicted price can be further adjusted based on the preset first price adjustment rule to obtain the mold price. Based on a summary of past business experience, this embodiment designs a series of parallel judgment and adjustment logics as the first price adjustment rule. The system will trigger these rules according to conditions. The first price adjustment rule includes at least the following: 1. Calculate the low price threshold below the historical lowest price based on the historical lowest price (for example, set the low price threshold to 70% of the historical lowest price). If the predicted price is lower than the low price threshold, issue the first alarm signal and output the predicted price as the mold price. At the same time, output the corresponding prediction confidence interval. At this time, the first alarm signal can indicate: The price is abnormally low. Please check if there are any missing items.

[0032] 2. Calculate the high price threshold based on the highest historical price (e.g., set the high price threshold to 150% of the highest historical price). If the predicted price is higher than the high price threshold, issue a second alarm signal and output the predicted price as the mold price, along with the corresponding prediction confidence interval. At this time, the second alarm signal can output a prompt: the cost risk is high, and it is recommended to optimize the design scheme.

[0033] 3. Calculate the similarity between the current mold and historical molds. The similarity calculation can be based on pre-designed rules, and molds with similarity greater than the similarity threshold and closest to the current mold quotation time are marked as similar molds. For example, attributes such as material, size, processing technology and precision are set as multiple evaluation dimensions. If the number of two molds that are the same in multiple evaluation dimensions is greater than the preset number, they can be judged to be similar. When the error between the predicted quotation of the current mold and the mold quotation of the similar mold is greater than the preset error value (e.g., 15%), a third alarm signal is issued and the predicted quotation is output as the mold quotation. At the same time, the corresponding prediction confidence interval is output. The third alarm signal can output a prompt: It is recommended to re-evaluate the process route.

[0034] 4. If the current mold is an urgent order, add an expedited fee (e.g., increase the forecasted price by 15%) to the forecasted price to obtain the mold price and the corresponding forecast confidence interval.

[0035] 5. If the current mold involves imported or exported materials, add tariffs (usually 10% of the predicted price) to the predicted price to obtain the mold price and the corresponding prediction confidence interval.

[0036] 6. If the price fluctuation of raw materials exceeds the preset material price fluctuation range (e.g., 5%), the raw material fluctuation cost will be added to the predicted price to obtain the mold price and the corresponding prediction confidence interval.

[0037] 7. If the average failure rate of this type of mold trial exceeds the preset failure rate threshold (usually set to 20%), then add a processing fee reserve fund (usually 20% of the processing fee will be added to the predicted quotation) to obtain the mold quotation and the corresponding prediction confidence interval.

[0038] Of course, experienced pricing engineers can also manually adjust the prediction results and indicate the reasons for the modification. The modified data will be fed back to the model as negative samples for subsequent iterations and optimizations.

[0039] S3. Obtain multi-dimensional data of the mold to be predicted and input it into the trained mold price prediction model to obtain the corresponding mold price. After obtaining the mold price, suppliers can be selected according to requirements. The company will score each supplier according to its own supplier scoring rules. When selecting suppliers, several suppliers with higher scores are generally selected as candidate suppliers. Then, the final mold price after selecting each candidate supplier is calculated according to the preset compensation mechanism. This compensation mechanism can be set to be executed automatically at the end of the mold price prediction model. After the predicted price is adjusted by the first price adjustment rule to obtain the mold price, the mold price is finally adjusted to obtain the final mold price. Alternatively, the output of the mold price prediction model can be further adjusted through this compensation mechanism. The compensation mechanism is as follows: 1. Calculate the ratio of the number of orders completed on time, with quality and quantity in the supplier's history to the total number of orders as the historical fulfillment rate. If the historical fulfillment rate of the candidate supplier is less than the preset value (usually 80%), then add a risk premium (usually 5%-10% more than the mold price) to the mold price as the final mold price.

[0040] 2. If the number of collaborations with candidate suppliers is less than the preset number of collaborations, a safety margin cost (generally 15% more than the mold quotation) will be added to the mold quotation to form the final mold quotation.

[0041] When training the above model, the AdamW optimizer can be used, and the parameters can be set as follows: the initial learning rate is 1e-4, the loss function is Huber Loss to take into account the advantages of MSE and MAE and improve robustness, the batch size is BatchSize=64, the number of training epochs is 200, an Early Stopping mechanism is introduced, and training is stopped when the validation set loss does not decrease for 10 consecutive epochs, and GPU parallel computing is used to accelerate the training process.

[0042] After training, the model is encapsulated as a lightweight service interface (such as TensorFlow Serving or ONNXRuntime) and deployed on an enterprise internal server or cloud platform. Users can call the API through the "Price Prediction" function module of the SAM system, input new project parameters, and obtain prediction results in real time. The final output results can be displayed in a visual form on the SAM system interface. In addition, to ensure the continuous updating of this model, whenever a mold project is completed and the final transaction price is confirmed, the system can automatically store the actual price and the final mold quotation into the database. Then, an incremental training program can be run monthly to fine-tune the model parameters using the new data. At the same time, an "abnormal case pool" can be set up to store cases with prediction deviations exceeding 20% ​​for expert review and model improvement. It also supports an online learning mode to gradually absorb new knowledge without affecting normal services.

[0043] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0045] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting mold pricing based on multi-dimensional data fusion, characterized in that, The method specifically includes the following steps: S1. Collect historical mold pricing sequences and multi-dimensional feature data for predicting mold pricing at each time point to construct a training dataset; S2. Construct a mold price prediction model that includes a feature encoding module, a Transformer module, an LSTM module, and a fusion output module, and train the model using the training dataset; The feature encoding module is used to encode multi-dimensional feature data and historical mold quotation sequences; The Transformer module is used to process encoded multi-dimensional feature data through multiple self-attention blocks to learn global context features that characterize the importance weights among the features. The LSTM module is used to extract overall temporal features from historical mold quotation sequences; The fusion output module is used to fuse global context features and overall temporal features, and calculate the mold quotation based on the fusion result; S3. Obtain multi-dimensional data of the mold to be predicted and input it into the trained mold price prediction model to obtain the corresponding mold price.

2. The mold pricing prediction method according to claim 1, characterized in that, The feature encoding module is used to construct preset derived features based on multi-dimensional data, divide the multi-dimensional feature data and derived features into multiple data categories that include at least categorical features and continuous features, encode multiple features according to the preset encoding rules of different data categories, and divide the historical mold quotation sequence into sub-time series with fixed time lengths.

3. The mold quotation prediction method according to claim 1, characterized in that, The Transformer module includes multiple self-attention blocks stacked sequentially. Each self-attention block contains a multi-head attention mechanism, a first residual connection module, a feedforward neural network, and a second residual connection module. The multi-head attention mechanism is used to calculate the local context features in the input features that represent the importance weights between the features, and to concatenate and linearly transform the local context features output by each attention head to obtain the attention output features. The first residual connection module is used to add the attention output features to the input features of the multi-head attention mechanism and perform layer normalization to obtain the first normalized features; The feedforward neural network is used to perform a nonlinear transformation on the first normalized feature to obtain a nonlinear transformed feature. The second residual connection module is used to add the first normalized feature and the nonlinear transformation feature, and perform layer normalization to obtain the second normalized feature; The second normalized feature output from the last self-attention block is labeled as a global context feature.

4. The mold pricing prediction method according to claim 1, characterized in that, The Transformer module also includes a pre-processing layer set at the input end of the self-attention block. The pre-processing layer is used to divide the input features into feature categories that include at least static attribute features, process attribute features and time-series cost features according to their physical meaning, and to map the features of different feature categories to the corresponding feature subspaces through different linear transformation layers. Then, the features in each feature subspace are concatenated to form an initial feature matrix for inputting the first multi-head attention mechanism. The first multi-head attention mechanism includes multiple feature attention heads for processing corresponding feature categories. Each feature attention head calculates local context features in the initial feature matrix to represent the attention weights between features of the corresponding feature category. Then, the local context features output by each feature attention head are concatenated and linearly transformed to obtain the attention output features.

5. The mold pricing prediction method according to claim 1, characterized in that, The LSTM module is a bidirectional LSTM network. The specific steps for extracting the overall time-series features from the historical mold quotation sequence are as follows: The forward and backward hidden state sequences are extracted from the input historical mold price quotation sequence through the forward and backward propagation layers of a bidirectional LSTM network, and the tanh function is used as the activation function inside both. By concatenating the forward hidden state and the backward hidden state at each time step, the temporal evolution features are obtained. Extract the feature vector of the last time step in the temporal evolution features to obtain the overall temporal features.

6. The mold quotation prediction method according to claim 1, characterized in that, The specific steps for calculating the mold quotation are as follows: The overall temporal features and global contextual features are concatenated to obtain fused features; The fused features are processed through at least two fully connected network layers, and Dropout is added to prevent overfitting to obtain the predicted quotes and corresponding prediction confidence intervals. The mold price is calculated based on the predicted price and the preset first price adjustment rule, and the mold price and the predicted confidence interval are output.

7. The mold quotation prediction method according to claim 6, characterized in that, The first price adjustment rule includes at least: Calculate the low price threshold below the historical lowest price based on the historical lowest price. If the predicted price is lower than the low price threshold, issue the first alarm signal and output the predicted price as the mold price, while also outputting the corresponding prediction confidence interval. Calculate the high price threshold that is higher than the historical high price based on the highest historical price. If the predicted price is higher than the high price threshold, issue a second alarm signal and output the predicted price as the mold price, and output the corresponding prediction confidence interval. Calculate the similarity between the current mold and historical molds, and mark the mold with a similarity greater than the similarity threshold and the closest time to the current mold quotation as a similar mold. When the error between the predicted quotation of the current mold and the mold quotation of the similar mold is greater than the preset error value, issue a third alarm signal and output the predicted quotation as the mold quotation, and output the corresponding prediction confidence interval. If the current mold order is urgent, an expedited fee will be added to the predicted quote to obtain the mold quote and the corresponding prediction confidence interval. If the current mold involves imported or exported materials, then the tariff will be added to the predicted price to obtain the mold price and the corresponding prediction confidence interval; If the price fluctuation of raw materials exceeds the preset material price fluctuation range, the raw material fluctuation fee will be added to the predicted price to obtain the mold price and the corresponding prediction confidence interval. If the average failure rate of this type of mold trial exceeds the preset failure rate threshold, then a processing fee reserve will be added to the predicted quotation to obtain the mold quotation and the corresponding prediction confidence interval.

8. The mold quotation prediction method according to claim 1, characterized in that, In step S3, after acquiring multi-dimensional data of the mold to be predicted and inputting it into the trained mold price prediction model to obtain the corresponding mold price, the method further includes: obtaining multiple suppliers with the highest comprehensive scores as candidate suppliers according to preset supplier scoring rules, and then calculating the final mold price after selecting each candidate supplier according to a preset compensation mechanism; the compensation mechanism is as follows: If the historical fulfillment rate of the candidate supplier is less than the preset value, a risk premium will be added to the mold quotation to form the final mold quotation. If the number of collaborations with candidate suppliers is less than the preset number of collaborations, a safety margin cost will be added to the mold quotation to form the final mold quotation.

Images