Project data processing method, device and equipment based on large model

Abstract

The embodiment of the invention provides a project data processing method, device and equipment based on a large model. Comprising the steps of obtaining initial project data and a prediction demand of a target project; based on the large model, performing correction processing on the initial project data to obtain to-be-processed project data; the to-be-processed project data comprises a static attribute field and a dynamic attribute field; based on the large model, the to-be-processed project data and the prediction demand, adjusting a preset initial prediction model to obtain a multi-source feature fusion prediction model matched with the prediction demand; based on the multi-source feature fusion prediction model and the prediction demand, extracting a fusion feature vector matched with the prediction demand from the static attribute field and the dynamic attribute field of the to-be-processed project data; and based on the multi-source feature fusion prediction model and the prediction demand, processing the fusion feature vector to obtain a prediction result corresponding to the prediction demand. The method is used for achieving the effects of improving the project data processing efficiency and reducing the model development cost.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus and device for processing project data based on a large model. Background Technology

[0002] In the field of software development, time-series forecasting based on project data is a core support for ensuring efficient project progress and precise resource allocation. Data such as work hours consumed, resource input, and progress generated during project operation all exhibit significant time-series characteristics. Through time-series forecasting, key information such as future work hour requirements, project cycle milestones, and resource gaps can be predicted in advance.

[0003] In existing technologies, time series forecasting based on project data often involves manually cleaning the project data and then using various deep learning models to perform forecasting on the processed data. However, the above methods are inefficient due to the manual data cleaning process. Furthermore, deploying multiple models to adapt to the various forecasting needs of the same project not only increases the cost of model development and training but also makes it difficult to fully utilize the inherent correlations within the same project data, resulting in poor consistency of forecasting results for multiple needs and insufficient adaptability to different scenarios. Summary of the Invention

[0004] The method, apparatus, and equipment for processing project data based on large models provided in this application are intended to improve the efficiency of project data processing and reduce the cost of model development.

[0005] In a first aspect, embodiments of this application provide a method for processing project data based on a large model, including:

[0006] Obtain initial project data and projected requirements for the target project;

[0007] Based on the large model, the initial project data is corrected to obtain the project data to be processed; wherein, the project data to be processed includes static attribute fields and dynamic attribute fields;

[0008] Based on the large model, the preset initial prediction model is adjusted according to the project data to be processed and the prediction requirements to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0009] Based on the multi-source feature fusion prediction model and the prediction requirements, a fusion feature vector matching the prediction requirements is extracted from the static attribute fields and dynamic attribute fields of the project data to be processed.

[0010] Based on the multi-source feature fusion prediction model and the prediction requirements, the fused feature vector is processed to obtain the prediction result corresponding to the prediction requirements.

[0011] In one possible implementation, the step of correcting the initial project data based on a large model to obtain project data to be processed includes:

[0012] Based on the large model, the fields in the initial project data are classified by field attributes to obtain the project data to be corrected; wherein, each field in the project data to be corrected has field attributes, and the field attributes include static attribute fields and dynamic attribute fields;

[0013] Based on the large model, determine the problem type of the field to be corrected in the project data to be corrected;

[0014] Based on a preset knowledge base, the question type of the field to be corrected, and the field attributes of the field to be corrected, the correction method corresponding to the field to be corrected is determined; wherein, the preset knowledge base includes correction methods corresponding to question types with different field attributes;

[0015] Based on the preset toolchain in the large model, the field to be corrected is corrected according to the correction method corresponding to the field to be corrected, and the data of the project to be processed is obtained; wherein, the preset toolchain has preset data correction logic.

[0016] In one possible implementation, the step of classifying the fields in the initial project data based on a large model to obtain the project data to be corrected includes:

[0017] For the fields to be classified in the initial project data, a query vector is constructed based on the field information of the fields to be classified;

[0018] Based on the preset knowledge base and the query vector, determine the association information that matches the field to be classified from the preset knowledge base;

[0019] Based on the large model, the field attributes of the field to be classified are determined according to the association information; wherein, the project data to be corrected is the initial project data after the field attributes are classified.

[0020] In one possible implementation, the step of extracting a fusion feature vector matching the prediction requirement from the static and dynamic attribute fields of the project data to be processed, based on the multi-source feature fusion prediction model and the prediction requirement, includes:

[0021] Based on the dynamic feature selection module in the multi-source feature fusion prediction model, the dynamic attribute fields of the project data to be processed are vectorized to obtain the initial dynamic attribute feature vectors corresponding to each dynamic attribute field.

[0022] Based on the dynamic feature selection module, the initial dynamic attribute feature vector is filtered according to the prediction requirements to obtain the dynamic attribute feature vector.

[0023] Based on the static attribute injection module in the multi-source feature fusion prediction model, static attribute feature vectors are extracted from the static attribute fields of the project data to be processed; wherein, the static attribute feature vectors have the same dimension as the dynamic attribute feature vectors.

[0024] Based on the encoding module in the multi-source feature fusion prediction model, the static attribute feature vector and the dynamic attribute feature vector are concatenated to obtain the fused feature vector.

[0025] In one possible implementation, the dynamic attribute field includes a dynamic historical attribute field and a dynamic future predictable attribute field; the step of filtering the initial dynamic attribute feature vector based on the dynamic feature selection module according to the prediction requirements to obtain a dynamic attribute feature vector includes:

[0026] Based on the dynamic feature selection module, an attention mechanism is used to calculate the importance weight of each initial dynamic attribute feature vector to the prediction requirement; wherein, the initial dynamic attribute feature vector corresponding to the dynamic historical attribute field and the initial dynamic attribute feature vector corresponding to the dynamic future knowable attribute field are assigned weight calculation preferences during the calculation process.

[0027] Based on the importance weights, the initial dynamic attribute feature vectors are filtered to obtain dynamic attribute feature vectors.

[0028] In one possible implementation, the static attribute fields include categorical static attribute fields and numerical static attribute fields; the static attribute injection module based on the multi-source feature fusion prediction model extracts static attribute feature vectors from the static attribute fields of the project data to be processed, including:

[0029] Based on the static attribute injection module, the categorized static attribute fields in the project data to be processed are embedded and transformed to obtain the first vector.

[0030] Based on the static attribute injection module, the numerical static attribute fields in the project data to be processed are normalized, and the normalized numerical static attribute fields are linearly transformed to obtain the second vector.

[0031] Based on the first vector and the second vector, the static attribute feature vector is obtained.

[0032] In one possible implementation, the fused feature vector includes dynamic attribute feature vectors and static attribute feature vectors; the step of processing the fused feature vectors based on the multi-source feature fusion prediction model and the prediction requirement to obtain a prediction result corresponding to the prediction requirement includes:

[0033] Based on the long sequence modeling module in the multi-source feature fusion prediction model, a preset gated unit network is used to encode a portion of the feature vectors in the fused feature vectors step by step to obtain an encoded feature matrix; wherein, the portion of the feature vectors represents dynamic attribute feature vectors with a time sequence length greater than a preset threshold.

[0034] Based on the dual-axis attention module in the multi-source feature fusion prediction model, attention feature fusion processing is performed on the encoded feature matrix according to the prediction requirements to obtain the dual-axis fused feature matrix.

[0035] Based on the modulation module in the multi-source feature fusion prediction model, the dual-axis fusion feature matrix is ​​modulated according to the prediction requirements to obtain the modulated feature matrix.

[0036] Based on the prediction head in the multi-source feature fusion prediction model, the modulated feature matrix is ​​subjected to prediction processing to obtain the prediction result corresponding to the prediction requirement.

[0037] In one possible implementation, the dual-axis attention module in the multi-source feature fusion prediction model performs attention feature fusion processing on the encoded feature matrix according to the prediction requirements to obtain a dual-axis fused feature matrix, including:

[0038] The encoded feature matrix is ​​preprocessed by decoding to obtain the preprocessed encoded feature matrix;

[0039] Based on the time granularity represented by the prediction requirement and the preprocessed encoding feature matrix, the time axis attention weight is determined; wherein, the time axis attention weight represents the weight between dynamic attribute feature vectors at different time steps in the preprocessed encoding feature matrix;

[0040] The preprocessed encoded feature matrix is ​​weighted based on the time axis attention weights to obtain the time axis attention output feature matrix.

[0041] Based on the preprocessed encoding feature matrix, variable axis attention weights are determined; wherein, the variable axis attention weights characterize the weights between different feature vectors in the preprocessed encoding feature matrix.

[0042] The preprocessed encoded feature matrix is ​​weighted based on the variable axis attention weights to obtain the variable axis attention output feature matrix.

[0043] The time-axis attention output feature matrix and the variable-axis attention output feature matrix are weighted and fused to obtain the dual-axis fused feature matrix.

[0044] In one possible implementation, the modulation module in the multi-source feature fusion prediction model modulates the dual-axis fusion feature matrix according to the prediction requirements to obtain a modulated feature matrix, including:

[0045] Based on the gated residual network in the modulation module, the feature channels of the dual-axis fusion feature matrix are filtered to obtain the filtered feature matrix.

[0046] Based on the scene labels represented by the predicted demand, determine the scaling factor and offset factor corresponding to the scene labels;

[0047] Based on the feature-channel linear modulation module in the modulation module, the filtered feature matrix is ​​modulated based on the scaling factor and the offset factor to obtain the modulated feature matrix.

[0048] In one possible implementation, the step of adjusting a preset initial prediction model based on the large model, according to the project data to be processed and the prediction requirements, to obtain a multi-source feature fusion prediction model that matches the prediction requirements, includes:

[0049] Based on the data of the project to be processed and the prediction requirements, the initial prediction model is trained to obtain an intermediate prediction model;

[0050] Based on the large model, the state vector of the intermediate prediction model is determined; wherein, the state vector represents the model accuracy of the initial prediction model;

[0051] If it is determined that the state vector does not meet the preset conditions, then based on the large model, the parameters of the intermediate prediction model are adjusted according to the state vector until the model accuracy of the intermediate prediction model meets the preset conditions.

[0052] The intermediate prediction model obtained when the preset conditions are met is the multi-source feature fusion prediction model.

[0053] In one possible implementation, the state vector includes one or more of the following: training loss curve shape, gradient stability, model output bias structure information, data state change information, and scene description labels;

[0054] Wherein, the training loss curve shape represents the shape of the loss function curve of the intermediate prediction model; the gradient stability represents the gradient norm of the intermediate prediction model; the model output deviation structure information represents the deviation distribution characteristics between the prediction results and the true values ​​of the intermediate prediction model; the data state change information represents the distribution characteristics of the data to be processed; and the scene description label represents the scene identifier corresponding to the prediction requirement.

[0055] Secondly, embodiments of this application provide a processing apparatus for project data based on a large model, comprising:

[0056] The acquisition module is used to acquire initial project data and predicted requirements for the target project.

[0057] The correction module is used to correct the initial project data based on the large model to obtain project data to be processed; wherein the project data to be processed includes static attribute fields and dynamic attribute fields.

[0058] The training module is used to adjust the preset initial prediction model based on the large model, the project data to be processed, and the prediction requirements, so as to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0059] The extraction module is used to extract a fusion feature vector that matches the prediction requirement from the static attribute fields and dynamic attribute fields of the project data to be processed, based on the multi-source feature fusion prediction model and the prediction requirement.

[0060] The prediction module is used to process the fused feature vector based on the multi-source feature fusion prediction model and the prediction requirement to obtain the prediction result corresponding to the prediction requirement.

[0061] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0062] The memory stores computer-executed instructions;

[0063] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0064] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0065] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0066] The method, apparatus, and equipment for processing project data based on a large model provided in this application first acquire the initial project data and prediction requirements of the target project. Then, the initial project data is corrected using a large model to obtain project data to be processed with clear field attributes and meeting data quality standards. Next, based on the large model, the project data to be processed, and the prediction requirements, a preset initial prediction model is adjusted to obtain a multi-source feature fusion prediction model adapted to the prediction requirements. Subsequently, a fusion feature vector that integrates static and dynamic attributes is extracted from the project data to be processed using this model. Finally, the fusion feature vector is processed by temporal inference to obtain the prediction result. This method automates data correction and model adjustment through a large model, avoiding the inefficiency and errors of manual processing. At the same time, relying on the multi-module collaborative capability of the multi-source feature fusion prediction model, it fully explores the inherent correlation of project data. Only one model architecture is needed to adapt to multiple prediction requirements of the same project, which not only reduces the model development and maintenance costs, but also improves the accuracy, consistency, and scenario adaptability of the prediction results, simplifies the operation process of project temporal prediction, and provides efficient and reliable data support for project management decisions. Attached Figure Description

[0067] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0068] Figure 1 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 1 ;

[0069] Figure 2 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 2 ;

[0070] Figure 3 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 3 ;

[0071] Figure 4 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 4 ;

[0072] Figure 5 A schematic diagram of a prediction curve provided for this application;

[0073] Figure 6A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 5 ;

[0074] Figure 7 A schematic diagram illustrating one application scenario provided in this application;

[0075] Figure 8 A schematic diagram of the structure of the device for processing project data based on a large model provided in this application;

[0076] Figure 9 A schematic diagram of the structure of the electronic device provided in this application.

[0077] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0078] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0079] In related technologies, time series prediction based on project data relies on manual data cleaning and requires the deployment of multiple deep learning models to adapt to various prediction needs of the same project. Manual cleaning is not only inefficient, but also prone to data quality instability due to differences in experience. Parallel development and training of multiple models increases costs and fails to fully explore the inherent correlation of data from the same project, resulting in poor consistency of prediction results and insufficient adaptability to scenarios. Based on this, the inventors of this application further considered that if the semantic understanding and intelligent decision-making capabilities of a large model could be used to replace manual work in automatically correcting project data, it would not only solve the problems of data processing efficiency and quality, but also provide a reliable data foundation for subsequent predictions. At the same time, considering the drawbacks of multi-model adaptation, the inventors conceived of a method that does not require the development of multiple models. Instead, a pre-set initial prediction model is used as a foundation, and a large model is used to combine the prediction requirements input by the user with the corrected data. The large model drives the initial prediction model to make targeted fine-tuning, so that the fine-tuned model has the ability to extract fusion features that adapt to the requirements from dynamic and static attribute fields. Finally, the prediction results are output based on the fusion features. In this way, multiple prediction requirements of the same project can be adapted through a single model architecture, making full use of the inherent correlation of data to ensure consistency of results and scenario adaptability, while also significantly reducing the model development and training costs.

[0080] The execution subject of this application embodiment can be an electronic device with processing capabilities, such as a computer, server, etc., and this application embodiment is not limited thereto.

[0081] It should be noted that the project data involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0082] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0083] Figure 1 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 1 ,like Figure 1 As shown, the method includes:

[0084] S101. Obtain initial project data and predicted requirements for the target project.

[0085] For example, the target project refers to a specific project that requires time-series predictive analysis, including but not limited to R&D projects, operation and maintenance projects, and engineering projects. Initial project data refers to the raw data generated or associated with the target project during its operation, including static attribute-related data (such as project type, scale, and team configuration) and dynamic attribute-related data (such as historical working hours, resource input, and progress node records). The data source can be tabular files, database storage data, etc. This embodiment does not limit the format and volume of the initial project data.

[0086] Forecasting requirements refer to specific time-series forecasting requests made by users based on target projects, including forecasting scenarios (such as work hour forecasting, cycle forecasting), forecasting time granularity (such as weekly, monthly), forecasting duration (such as the next 3 months, the next 6 cycles), and other information.

[0087] In one example, a user can upload initial project data for a target project through the interactive interface of an electronic device, and input predicted requirements through the interface, such as "based on the project's work hour data for the past 12 months, predict the weekly work hour consumption for the next 3 months." It should be noted that the input method for predicted requirements can be text input or voice input, etc., and this embodiment of the application does not limit this. Alternatively, the electronic device can retrieve the initial project data of the target project stored in a preset project data management platform and read the predicted requirement form submitted in advance by the user.

[0088] S102. Based on the large model, the initial project data is corrected to obtain the project data to be processed.

[0089] For example, a large model refers to an artificial intelligence model with natural language understanding, data quality analysis, and rule reasoning capabilities, capable of generating data processing solutions based on domain knowledge and historical experience. It should be noted that this application does not limit the type of large model.

[0090] Correction processing refers to a series of processing operations, such as data cleaning, format standardization, and field classification, to address quality issues (e.g., missing values, outliers, inconsistent formats) and unclear field attributes in the initial project data.

[0091] Project data to be processed refers to project data that has been corrected and whose quality meets the standards and whose field attributes are clear, and can be used for subsequent model training and feature extraction. Field attributes include static attributes and dynamic attributes; in other words, project data to be processed includes both static and dynamic attribute fields.

[0092] Static attribute fields refer to fields in the project data that do not change over time, used to characterize the inherent attributes of the project, such as project type, rated scale, and department. Dynamic attribute fields refer to fields in the project data that change over time, used to characterize the dynamic operating status of the project, such as work hours consumed, resource input, and progress completion rate at different time steps.

[0093] In one example, the large model first classifies the initial project data by field, then performs data quality checks on each field based on the classification results to identify problems in the field; then, based on the preset project data domain knowledge base, it generates targeted correction schemes; finally, it executes the correction operations through the built-in toolchain (such as the pandas data processing tool integrated by langchain), and outputs the project data to be processed, which includes static attribute fields and dynamic attribute fields.

[0094] S103. Based on the large model, adjust the preset initial prediction model according to the data of the project to be processed and the prediction requirements to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0095] For example, the preset initial prediction model refers to a time-series prediction model architecture that is pre-built and has the ability to fuse multiple source features.

[0096] Adjustment refers to adjusting the parameters of the initial prediction model based on the data of the project to be processed and the prediction requirements, so that the model can be adapted to the specific data patterns and prediction scenarios of the target project.

[0097] Multi-source feature fusion prediction model refers to a time-series prediction model that, after adjustment, can accurately capture the correlation between static and dynamic attributes of a target project and adapt to specific prediction needs.

[0098] In one example, based on a large model, the feature distribution of the project data to be processed and the core requirements of the prediction needs are first analyzed; the attention calculation weight of the dynamic feature selection module is adjusted, the embedding vector dimension of the static attribute injection module is optimized, and the learning rate decay coefficient is set; finally, the project data to be processed is divided into training set and validation set, and the initial prediction model is iteratively trained according to the adjustment strategy. The prediction accuracy is monitored through the validation set until the model meets the preset accuracy threshold, and a multi-source feature fusion prediction model matching the current prediction needs is obtained.

[0099] S104. Based on the multi-source feature fusion prediction model and prediction requirements, extract the fusion feature vector that matches the prediction requirements from the static and dynamic attribute fields of the project data to be processed.

[0100] For example, a fused feature vector refers to a high-dimensional vector representation formed by fusing the static and dynamic attribute features of the project data to be processed.

[0101] In one example, the dynamic feature selection module of the multi-source feature fusion prediction model calculates the importance weight of each dynamic attribute field in the project data to be processed according to the prediction requirements, filters out high-contribution dynamic features and converts them into dynamic feature vectors; then, through the static attribute injection module of the multi-source feature fusion prediction model, the static attribute fields are converted into static feature vectors with the same dimensions as the dynamic feature vectors; finally, in the encoding module of the multi-source feature fusion prediction model, the static feature vectors and dynamic feature vectors are concatenated and interactively fused to generate a fused feature vector that matches the prediction requirements.

[0102] S105. Based on the multi-source feature fusion prediction model and prediction requirements, the fused feature vector is processed to obtain the prediction result corresponding to the prediction requirements.

[0103] For example, the prediction result refers to the quantitative output of the future state of the target project calculated by the multi-source feature fusion prediction model based on the fused feature vector and prediction requirements through time-series inference. This includes specific predicted values, such as predicted working hours for each week and the remaining project cycle. The time granularity and output format of the prediction result are consistent with the prediction requirements. Optionally, the prediction result may also include a model specification document of the multi-source feature fusion prediction model, used to characterize the model parameter information of the currently used model.

[0104] In one example, the fused feature vector is first input into the long sequence modeling module of the multi-source feature fusion prediction model. The optimized gating mechanism captures the long temporal dependencies of dynamic features and outputs the encoded feature matrix. Then, the dual-axis attention module of the multi-source feature fusion prediction model calculates the time axis attention weight and variable axis attention weight based on the temporal granularity of the prediction requirement to fuse and strengthen the encoded feature matrix. Subsequently, the modulation module of the multi-source feature fusion prediction model adjusts the feature channel weights based on the scene label of the prediction requirement and modulates the fused encoded feature matrix. Finally, the prediction head of the multi-source feature fusion prediction model performs linear transformation and activation processing on the modulated encoded feature matrix and outputs the prediction result.

[0105] The project data processing method based on a large model provided in this application first obtains the initial project data and prediction requirements of the target project. Then, the initial project data is corrected using a large model to obtain project data to be processed with clear field attributes and meeting data quality standards. Next, based on the large model, the project data to be processed, and the prediction requirements, a preset initial prediction model is adjusted to obtain a multi-source feature fusion prediction model adapted to the prediction requirements. Subsequently, a fusion feature vector that integrates static and dynamic attributes is extracted from the project data to be processed using this model. Finally, the fusion feature vector is processed by time-series inference to obtain the prediction result. This method automates data correction and model adjustment through a large model, avoiding the inefficiency and errors of manual processing. At the same time, relying on the multi-module collaborative capabilities of the multi-source feature fusion prediction model, it fully explores the inherent correlation of project data. Only one model architecture is needed to adapt to multiple prediction requirements of the same project, which not only reduces the model development and maintenance costs, but also improves the accuracy, consistency, and scenario adaptability of the prediction results, simplifies the operation process of project time-series prediction, and provides efficient and reliable data support for project management decisions.

[0106] Figure 2 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, a detailed explanation of the method for processing project data based on large models is provided, which includes:

[0107] S201. Obtain initial project data and predicted requirements for the target project.

[0108] It should be noted that this step is similar to the aforementioned step S101, and will not be repeated here.

[0109] S202. Based on the large model, classify the fields in the initial project data according to their attributes to obtain the project data to be corrected.

[0110] For example, a field refers to the smallest data unit in the initial project data that carries specific information, such as "project type", "work hours in January 2024", and "number of scheduled personnel". Field attribute classification refers to the process of dividing fields into specific attribute categories based on whether the field changes over time, data characteristics, etc.

[0111] The project data to be corrected refers to the initial project data where each field has been clearly labeled with an attribute after being categorized by field attributes. In other words, each field in the project data to be corrected has field attributes, including static attribute fields and dynamic attribute fields.

[0112] In one example, based on a large model, all field names and corresponding example values ​​of the initial project data are first extracted. Then, combined with a pre-set knowledge base, the field characteristics are analyzed to obtain the project data to be corrected, with attribute labels attached to each field. Here, the pre-set knowledge base refers to a pre-built vectorized knowledge base for project data domains. This knowledge base includes information such as the table structure (schema), field meanings, possible roles of time fields, domain rules, and historical cases. All information has been converted into vector form and stored.

[0113] Optionally, for table files, the table header, data type, and time column formats such as yyyy / MM / dd, timestamp, and string date are identified through parsing operations, and this information is written into the knowledge base; for databases, the schema can be read through database connection, and the field types, index fields, and timestamp fields can be automatically obtained and the knowledge base is written; finally, a vector model is used to encode the field descriptions, domain rules, historical cases, and other content in the knowledge base into vectors to form a vector library.

[0114] Optionally, the knowledge base may also include classification rules for field attributes, such as:

[0115] Static attribute fields: fields without timestamp association and whose values ​​do not change over time, such as project size and type.

[0116] Dynamic historical attribute field: Associated with the timestamp of the completed time period + the value is the observed value, such as the work hours that have occurred.

[0117] Dynamic future attribute field: Associated with a future time period timestamp + value is the planning / scheduling value, such as the personnel deployment that has already been scheduled.

[0118] Unknown future attribute field: associated with future time period timestamp + no planned value.

[0119] Specifically, for the fields to be classified in the initial project data, a query vector is constructed based on the field information of the fields to be classified; according to the preset knowledge base and the query vector, the association information matching the fields to be classified is determined from the preset knowledge base; based on the large model, the field attributes of the fields to be classified are determined according to the association information; wherein, the project data to be corrected is the initial project data after the field attributes are classified.

[0120] In this context, "fields to be categorized" refers to fields in the initial project data for which attribute labels have not yet been defined. Field information refers to the core characteristic data of the fields to be categorized, including field name, data type (e.g., numeric, character, time), example value, timestamp, and business scenario description (e.g., field purpose, data collection method).

[0121] A query vector is a high-dimensional vector obtained by encoding the field information of the field to be classified through a pre-trained vector model, which is used to perform similarity matching with vectors in a pre-defined knowledge base.

[0122] Related information refers to knowledge base content retrieved from a preset knowledge base that has a semantic similarity to the query vector of the field to be classified that is higher than a preset threshold. Field attributes refer to the classification results of a field, i.e., static attribute fields or dynamic attribute fields.

[0123] In one example, for the fields to be classified in the initial project data, the field information is first extracted, such as: "Field name: XXX + Data type: XXX + Example value: XXX + Timestamp association: XXX + Prediction scenario: Working hours prediction". This field information is input into a pre-trained BERT model to generate the corresponding query vector. Then, based on the large model, the query vector is used to search in a preset knowledge base, and knowledge base content with similarity greater than a preset threshold is selected as association information. This association information may include, for example, similar field classification cases + domain time rules + field meaning inference results, etc. This association information is then input into the large model. The large model determines the field attributes of the field to be classified by semantic analysis of the association information and the preset field classification rules, and finally obtains the project data to be corrected with attribute labels for each field.

[0124] S203. Based on the large model, determine the problem type of the field to be corrected in the project data.

[0125] For example, a field to be corrected refers to a field in the project data that has data quality defects and cannot be directly used for model training. The problem type refers to the specific category of data quality problems existing in the field to be corrected, including missing value problems, outlier problems (field data exceeds the reasonable business scope or statistical distribution range), inconsistent format problems (such as inconsistent format of time fields or inconsistent numerical units), and data redundancy problems (fields contain duplicate records or invalid information).

[0126] In one example, a field-by-field quality check is performed on the data of the project to be corrected based on a large model. By statistically analyzing the percentage of non-null values ​​in a field, it is determined whether there are missing values. By using preset statistical rules, abnormal data is identified to determine whether there are outlier issues. By parsing the format, it is determined whether the field format is consistent to determine whether there are format inconsistencies, thus obtaining the problem type corresponding to each field to be corrected.

[0127] S204. Based on the preset knowledge base, the problem type of the field to be corrected, and the field attributes of the field to be corrected, determine the correction method corresponding to the field to be corrected.

[0128] For example, the preset knowledge base refers to a pre-built knowledge base for project data correction, containing standard correction rules, historical amendment examples, business constraints, etc., corresponding to different combinations of field attributes and different question types, and the knowledge base has been vectorized. The correction method refers to the specific data correction operation plan formulated based on the question type and field attribute of the field to be corrected.

[0129] In one example, the large model first retrieves matching correction rules and historical cases from a pre-defined knowledge base based on the attributes and question type of the field to be corrected; then, combined with business constraints, it determines the correction method for the field to be corrected. For example, if the missing value is a static attribute field, mode imputation can be used; if the missing value is a dynamic attribute field, preceding imputation or following imputation can be used, etc.

[0130] S205. Based on the preset toolchain in the large model, the field to be corrected is corrected according to the correction method corresponding to the field to be corrected, and the data of the project to be processed is obtained.

[0131] For example, a pre-defined toolchain refers to a set of tools built into a large model that integrates data processing functions. For instance, it can be a tool proxy set up based on langchain. This tool proxy can include pandas data processing tools, numpy numerical calculation tools, time format conversion tools, etc., and has pre-defined specific data processing logic corresponding to various correction methods (such as code scripts for forward filling and algorithm logic for outlier removal).

[0132] For example, a tool call sequence can be preset, such as first performing time standardization → then processing missing values ​​→ finally correcting outliers. Large models can call the corresponding tools according to the tool call sequence, and perform data cleaning and format standardization on the fields to be corrected according to the determined correction method to obtain the data of the project to be processed.

[0133] S206. Based on the data of the project to be processed and the prediction requirements, the initial prediction model is trained to obtain the intermediate prediction model.

[0134] For example, the initial prediction model refers to a pre-built time-series prediction model architecture with multi-source feature fusion capability, which has not been adapted and trained with target project data. It may include core components such as dynamic feature selection module, static attribute injection module, long sequence modeling module, and dual-axis attention module.

[0135] An intermediate prediction model refers to a model that has preliminary prediction capabilities but has not yet reached the preset accuracy requirements, obtained by using the data of the project to be processed as training data and combining the guidance of prediction needs to train the initial prediction model for the first round.

[0136] In one example, based on the prediction requirements, static and dynamic attribute fields related to the prediction requirements in the project data to be processed are filtered out and divided into training and validation sets according to the proportion. The training set data is input into the initial prediction model. The model uses the dynamic feature selection module to filter out dynamic features that are strongly related to the prediction requirements, uses the static attribute injection module to vectorize the static features and fuse them with the dynamic features, and uses the long sequence modeling module to capture the temporal dependencies to complete the first round of iterative training. After training, an intermediate prediction model with the ability to predict prediction requirements is obtained.

[0137] S207. Based on the large model, determine the state vector of the intermediate prediction model.

[0138] For example, a state vector refers to a multi-dimensional data vector used to characterize the training state and prediction accuracy of an intermediate prediction model. Optionally, the state vector may include one or more of the following: the shape of the training loss curve, gradient stability, model output bias structure information, data state change information, and scene description labels;

[0139] Among these, the training loss curve shape represents the shape of the loss function curve of the intermediate prediction model, such as monotonically decreasing, fluctuating upward, or tending to plateau. Gradient stability represents the change in the gradient norm of the intermediate prediction model, used to determine whether the model has converged. Model output bias structure information represents the distribution characteristics of the deviation between the prediction results and the true values ​​of the intermediate prediction model. Data state change information represents the distribution characteristics of the data to be processed. Scene description labels represent the scene identifier corresponding to the prediction requirement.

[0140] In one example, training logs of intermediate prediction models are obtained based on a large model. The morphological features of the training loss curve (such as the loss value not stabilizing after 150 iterations, indicating fluctuations) and gradient norm change data (such as a gradient norm standard deviation of 0.18, which is greater than the preset threshold of 0.1, indicating gradient instability) are extracted. By comparing the predicted values ​​with the true values ​​on the validation set, it is found that the model output has a systematic bias structure of 5% lower. Combining the distribution characteristics of the project data to be processed and the scene labels of the work hour prediction, a multi-dimensional state vector representing the current model state is constructed.

[0141] S208. Determine whether the state vector meets the preset conditions.

[0142] For example, preset conditions refer to pre-set model training standards used to determine whether the state of the intermediate prediction model meets the requirements of prediction accuracy and stability. Specifically, these conditions may include training loss values ​​below a preset threshold, gradient norm standard deviations below a preset threshold, and absolute values ​​of model output deviations below a preset range. Preset conditions can be dynamically adjusted according to different prediction needs and business scenarios.

[0143] In one example, the electronic device can compare the state vector with preset conditions one by one based on a large model to determine whether the current state vector meets the preset requirements. If it does, step S210 is executed; if it does not, step S209 is executed.

[0144] S209. Based on the large model, adjust the parameters of the intermediate prediction model according to the state vector.

[0145] For example, parameter tuning refers to the process by which a large model dynamically corrects the core parameters of an intermediate prediction model based on model problems reflected in the state vector (such as unstable gradients, large output bias, or high loss values) and historical experience and rules from a pre-defined knowledge base. Model parameters include learning rate, feature weight coefficients, attention mechanism parameters, modulation module coefficients, and number of iterations.

[0146] In one example, the preset knowledge base can contain preset model parameter tuning rules and historical adjustment cases. The state vector is used as a retrieval vector and input into the large model. Combined with the tuning rules and historical adjustment cases in the preset knowledge base, a parameter adjustment strategy is generated. The intermediate prediction model is then adjusted based on the parameter adjustment strategy. After the adjustment is completed, the intermediate prediction model is retrained based on the data of the project to be processed. The process of S206-S208 is executed iteratively until the state vector meets the preset conditions.

[0147] S210. The intermediate prediction model obtained when the preset conditions are met is determined as the multi-source feature fusion prediction model.

[0148] For example, a multi-source feature fusion prediction model refers to a mature time series prediction model that, after multiple rounds of training and parameter adjustment, has a state vector that meets preset conditions, achieves the required prediction accuracy and stability, and can accurately capture the correlation between the static and dynamic attributes of the target project and adapt to specific prediction needs.

[0149] S211. Based on the multi-source feature fusion prediction model and prediction requirements, extract the fusion feature vector that matches the prediction requirements from the static and dynamic attribute fields of the project data to be processed.

[0150] It should be noted that this step is similar to the aforementioned step S104, and will not be repeated here.

[0151] S212. Based on the multi-source feature fusion prediction model and prediction requirements, the fused feature vector is processed to obtain the prediction result corresponding to the prediction requirements.

[0152] It should be noted that this step is similar to the aforementioned step S105, and will not be repeated here.

[0153] The method for processing project data based on a large model provided in this application first obtains the initial project data and prediction requirements of the target project. Then, the large model classifies the fields of the initial project data by attribute to obtain the project data to be corrected. Next, the large model identifies the problem types of the fields to be corrected, determines targeted correction methods based on a preset knowledge base and field attributes, and executes the correction process through a preset toolchain to obtain standardized project data to be processed. Subsequently, based on the prediction requirements and the project data to be processed, an intermediate prediction model is trained to obtain an intermediate prediction model. The large model constructs a state vector representing the model's accuracy and state and determines whether it meets preset conditions. If not, the model parameters are adjusted and iterative training is performed; if it does, the intermediate prediction model is determined as a multi-source feature fusion prediction model. Finally, the model extracts fusion feature vectors matching the prediction requirements from the project data to be processed, and the prediction result is obtained after model temporal inference and feature processing. This method, on the one hand, automates the entire process of data field classification, problem identification, and correction through a large-scale model, completely eliminating reliance on manual operation and significantly improving the efficiency and quality of data processing, while avoiding errors caused by differences in human experience. On the other hand, through a model iteration and optimization mechanism driven by a large-scale model, it ensures a deep fit between the multi-source feature fusion prediction model and the target project data patterns and prediction needs, significantly improving the accuracy and stability of prediction results. Furthermore, relying on a single-model architecture to achieve multi-source feature fusion and temporal inference eliminates the need to deploy multiple models for various prediction needs of the same project, fully exploring the inherent correlations in project data. While ensuring consistency of results under different prediction needs, it significantly reduces the costs of model development, training, and maintenance, simplifies the operational process of project temporal prediction, and provides efficient and reliable data support for project management decisions.

[0154] Figure 3 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 3 ,like Figure 3 As shown, in this embodiment... Figure 1 Based on the examples, a detailed explanation of the method for processing project data based on large models is provided, which includes:

[0155] S301. Obtain initial project data and predicted requirements for the target project.

[0156] It should be noted that this step is similar to the aforementioned step S101, and will not be repeated here.

[0157] S302. Based on the large model, the initial project data is corrected to obtain the project data to be processed.

[0158] It should be noted that this step is similar to the aforementioned step S102, and will not be repeated here.

[0159] S303. Based on the large model, the preset initial prediction model is adjusted according to the data of the project to be processed and the prediction requirements to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0160] It should be noted that this step is similar to the aforementioned step S103, and will not be repeated here.

[0161] S304. Based on the dynamic feature selection module in the multi-source feature fusion prediction model, the dynamic attribute fields of the project data to be processed are vectorized to obtain the initial dynamic attribute feature vectors corresponding to each dynamic attribute field.

[0162] For example, the dynamic feature selection module is a core component in the multi-source feature fusion prediction model used for dynamic attribute field processing, and has the dual functions of feature vectorization and dynamic filtering.

[0163] Dynamic attribute fields refer to fields in the project data that change over time, including dynamic historical attribute fields (observed time series data, such as past weekly working hours) and dynamic future-knowable attribute fields (time series data that has not occurred but can be predicted, such as the number of personnel to be deployed in a scheduled manner).

[0164] Feature vectorization refers to the process of converting the original numerical data of dynamic attribute fields into high-dimensional dense vectors that can be recognized by the model.

[0165] The initial dynamic attribute feature vector refers to the original vector generated after each dynamic attribute field has been vectorized.

[0166] In one example, the dynamic feature selection module normalizes the original data of the dynamic attribute fields of the project data to be processed, eliminating the influence of units; then, the time series embedding layer converts the normalized time series data into a high-dimensional vector to obtain the initial dynamic attribute feature vectors corresponding to each dynamic attribute field.

[0167] S305. Based on the dynamic feature selection module, the initial dynamic attribute feature vector is filtered and processed according to the prediction requirements to obtain the dynamic attribute feature vector.

[0168] For example, the filtering process refers to the process by which the dynamic feature selection module, based on the prediction requirements, calculates the importance of each initial dynamic attribute feature vector to the prediction requirements using a specific algorithm (such as an attention mechanism), retaining high-contribution vectors and eliminating redundant or low-relevance vectors. Dynamic attribute feature vectors refer to the set of dynamic feature vectors that, after filtering, are strongly correlated with the prediction requirements and can provide core support for prediction inference.

[0169] In one example, the dynamic feature selection module uses an attention mechanism to transform the prediction requirement into a query vector, and calculates the contribution weight with each initial dynamic attribute feature vector; the initial dynamic attribute feature vector with a contribution weight greater than a preset threshold is used as the dynamic attribute feature vector.

[0170] Specifically, based on the dynamic feature selection module, an attention mechanism is used to calculate the importance weight of each initial dynamic attribute feature vector to the prediction requirement. Among them, the initial dynamic attribute feature vectors corresponding to the dynamic historical attribute fields and the initial dynamic attribute feature vectors corresponding to the dynamic future knowable attribute fields are assigned weight calculation preferences during the calculation process. Based on the importance weights, the initial dynamic attribute feature vectors are filtered to obtain dynamic attribute feature vectors.

[0171] The attention mechanism refers to an algorithm that assigns differentiated weights by calculating the correlation between features and prediction requirements, enabling it to accurately capture features crucial to the prediction target. Importance weights are quantitative values ​​representing the contribution of each initial dynamic attribute feature vector to the prediction requirement; higher weights indicate a stronger correlation between the feature and the prediction target.

[0172] Dynamic historical attribute fields refer to observed dynamic fields that change over time (such as past weekly working hours and historical resource input). Their corresponding initial dynamic attribute feature vectors carry the past operational patterns of the project. Dynamic future knowable attribute fields refer to unforeseen but predictable dynamic fields that change over time (such as the scheduled number of personnel and planned resource allocation). Their corresponding initial dynamic attribute feature vectors carry the project's future planning information.

[0173] Weight calculation preference refers to the setting in the weight calculation process where the calculation parameters of the attention mechanism are adjusted (such as adding a bias term or adjusting the similarity calculation weight) to enhance the contribution of dynamic historical attribute feature vectors and dynamic future knowable attribute feature vectors in the weight result. Filtering refers to the process of retaining high-weight feature vectors and removing low-weight redundant vectors based on a comparison of importance weights with preset thresholds.

[0174] For example, for the prediction requirement of "weekly working hours prediction", the prediction requirement is first converted into a standardized query vector based on the dynamic feature selection module. Then, the attention mechanism is invoked to calculate the semantic similarity between each initial dynamic attribute feature vector and the query vector. At the same time, the weight calculation preference mechanism is activated to multiply the initial vectors corresponding to dynamic historical attribute fields and dynamic future knowable attribute fields by a preset preference coefficient based on the similarity calculation results, thereby strengthening their weight ratio. The importance weight of each initial vector is obtained through this mechanism. All initial dynamic attribute feature vectors are screened, retaining the initial dynamic attribute feature vectors with importance weights higher than a preset threshold, and eliminating redundant vectors with weights lower than the threshold, finally obtaining the dynamic attribute feature vector.

[0175] S306. Based on the static attribute injection module in the multi-source feature fusion prediction model, extract static attribute feature vectors from the static attribute fields of the project data to be processed.

[0176] For example, the static attribute injection module refers to the core component in the multi-source feature fusion prediction model used for processing static attribute fields, and has the functions of static feature transformation and dimension adaptation.

[0177] Static attribute fields refer to fields in the project data that do not change over time and represent the inherent attributes of the project. These include categorical static attribute fields (such as project type and team configuration level) and numerical static attribute fields (such as project rated size and team member base).

[0178] Static attribute feature vectors refer to converting static attribute fields into high-dimensional vectors with the same dimensions as dynamic attribute feature vectors, which are recognizable by the model, ensuring subsequent feature fusion. Specifically, the static attribute feature vectors and dynamic attribute feature vectors have the same dimensions.

[0179] Specifically, based on the static attribute injection module, the categorical static attribute fields in the project data to be processed are embedded and transformed to obtain the first vector; based on the static attribute injection module, the numerical static attribute fields in the project data to be processed are normalized, and the normalized numerical static attribute fields are linearly transformed to obtain the second vector; based on the first vector and the second vector, the static attribute feature vector is obtained.

[0180] Among them, categorical static attribute fields refer to fields in the project data that do not change over time and represent the inherent category attributes of the project. Their values ​​are discrete category labels (e.g., project type: software development, hardware integration, operation and maintenance services; team configuration level: level A, level B, level C). These fields cannot be directly processed by the model and need to be converted into vector form through embedding. The first vector refers to the high-dimensional dense vector obtained by mapping the category labels of the categorical static attribute fields through the embedding layer. Numerical static attribute fields refer to fields in the project data that do not change over time and represent the inherent quantitative attributes of the project. Their values ​​are continuous or discrete numerical values ​​(e.g., project rated size: 120 people). Monthly, 80 people (Month; Team size: 20 or 35 people) These fields have different units of measurement and need to be normalized first. Normalization is a standardization operation that maps the original values ​​of numerical static attribute fields to a preset range to eliminate the impact of unit of measurement differences on model training and ensure fair weighting of different numerical fields.

[0181] Linear transformation refers to the operation of linearly mapping the normalized numerical static attribute field to a high-dimensional vector with the same dimension as the first vector, in order to meet the dimensionality requirements of feature fusion. The second vector is the high-dimensional vector obtained after normalization and linear transformation of the numerical static attribute field, whose dimension is exactly the same as the first vector.

[0182] Static attribute feature vectors refer to a unified high-dimensional vector obtained by combining the first vector and the second vector through a preset integration method (such as concatenation, weighted summation, attention fusion).

[0183] For example, the static attribute fields of the project data to be processed include categorical fields and numeric fields, such as "Project Type", with example values: software development, hardware integration; and "Project Rated Size", with example value: 100 people. The static attribute injection module first converts the data of categorical fields into a first vector of dimension 256 through an embedding layer. Then, it normalizes the data of numerical fields and maps them to a second vector of dimension 256 through a linear transformation. Finally, it concatenates and integrates the first and second vectors to obtain a static attribute feature vector of dimension 256, which maintains the same dimension as the dynamic attribute feature vector. The linear transformation can be, for example, y = Wx + b, where W is a 256×1 weight matrix and b is a 256×1 bias term.

[0184] S307. Based on the encoding module in the multi-source feature fusion prediction model, the static attribute feature vector and the dynamic attribute feature vector are concatenated to obtain the fused feature vector.

[0185] For example, the encoding module refers to the core component used for feature fusion in a multi-source feature fusion prediction model, which has the functions of vector concatenation and feature interaction enhancement. The concatenation process refers to the process of combining static attribute feature vectors and dynamic attribute feature vectors with the same dimension according to a preset dimension to form a unified vector representation containing multi-source information.

[0186] A fused feature vector is a high-dimensional vector that, after being spliced ​​together, contains both the inherent attributes (static) and dynamic operating rules of a project (dynamic), providing comprehensive and three-dimensional feature support for model prediction.

[0187] In one example, the encoding module first performs feature alignment verification on the static attribute feature vector and the dynamic attribute feature vector to ensure that they are completely consistent in data format and dimension; then, the dimensions are concatenated in the order of "static feature vector + dynamic feature vector" to form a fused feature vector with dimension 512.

[0188] S308. Based on the multi-source feature fusion prediction model and prediction requirements, the fused feature vector is processed to obtain the prediction result corresponding to the prediction requirements.

[0189] It should be noted that this step is similar to the aforementioned step S105, and will not be repeated here.

[0190] The project data processing method based on a large model provided in this application first obtains the initial project data and prediction requirements of the target project. Then, the large model automatically corrects the initial project data to obtain project data with clear field attributes and meeting data quality standards. Next, the large model, combined with the project data and prediction requirements, makes targeted adjustments to the preset initial prediction model to obtain a multi-source feature fusion prediction model adapted to the prediction requirements. Subsequently, the dynamic feature selection module of this model completes the vectorization and filtering of dynamic attribute fields, and the static attribute injection module extracts static attribute feature vectors, which are then concatenated by the encoding module to form a fused feature vector. Finally, based on the multi-source feature fusion prediction model and prediction requirements, temporal inference and feature processing are performed on the fused feature vector to obtain accurate prediction results. This method relies on the modular design of the multi-source feature fusion prediction model to achieve refined processing and deep fusion of static and dynamic features, fully exploring the inherent correlation of project data and improving the accuracy and reliability of prediction results.

[0191] Figure 4 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 4 ,like Figure 4 As shown, in this embodiment... Figure 1 Based on the examples, a detailed explanation of the method for processing project data based on large models is provided, which includes:

[0192] S401. Obtain initial project data and forecasted requirements for the target project.

[0193] It should be noted that this step is similar to the aforementioned step S101, and will not be repeated here.

[0194] S402. Based on the large model, the initial project data is corrected to obtain the project data to be processed.

[0195] It should be noted that this step is similar to the aforementioned step S102, and will not be repeated here.

[0196] S403. Based on the large model, the preset initial prediction model is adjusted according to the data of the project to be processed and the prediction requirements to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0197] It should be noted that this step is similar to the aforementioned step S103, and will not be repeated here.

[0198] S404. Based on the multi-source feature fusion prediction model and prediction requirements, extract the fusion feature vector that matches the prediction requirements from the static and dynamic attribute fields of the project data to be processed.

[0199] It should be noted that this step is similar to the aforementioned step S104, and will not be repeated here.

[0200] S405. Based on the long sequence modeling module in the multi-source feature fusion prediction model, a preset gated unit network is used to encode part of the feature vectors in the fused feature vector step by step to obtain the encoded feature matrix.

[0201] For example, the long sequence modeling module is a core component in the multi-source feature fusion prediction model used to process long-term dynamic features. It can capture the long-term dependencies of dynamic features and alleviate the gradient decay problem in training long sequence data.

[0202] The pre-defined gated unit network refers to the network structure built into the long sequence modeling module that has the ability to remember and forget time-series information. For example, it can be a QGRU (Gated Recurrent Unit) network, which improves the modeling ability of long time-series data by improving the coordination mechanism of reset gate and update gate.

[0203] Partial feature vectors refer to dynamic attribute feature vectors in the fused feature vectors whose time series length is greater than a preset threshold (such as 12 months). These features contain richer historical operational patterns. Time-step encoding processing refers to the process by which the gated unit network calculates long-time dynamic feature vectors sequentially according to time order, gradually capturing the dependencies between features at different time steps.

[0204] The encoded feature matrix is ​​a two-dimensional matrix generated after time-step encoding. The row dimension corresponds to the time step, and the column dimension corresponds to the feature dimension. Each element in the matrix represents the feature encoding result of the corresponding time step and contains long-term temporal dependency information.

[0205] In one example, the preset threshold is set to 12 months. The feature vectors that meet the preset threshold from the fused feature vectors are input into the long sequence modeling module. The QGRU network in this module encodes the feature vectors step by step in chronological order. The optimized reset gate filters key historical information, and the update gate integrates the current time step features with historical information to alleviate the gradient decay problem. After the encoding is completed, the encoded feature matrix is ​​obtained.

[0206] S406. Based on the dual-axis attention module in the multi-source feature fusion prediction model, attention feature fusion processing is performed on the encoded feature matrix according to the prediction requirements to obtain the dual-axis fused feature matrix.

[0207] For example, the dual-axis attention module is a core component in the multi-source feature fusion prediction model used for feature fusion enhancement. It has dual computational capabilities of time axis attention and variable axis attention, and can capture related information from both the time dimension and the feature dimension.

[0208] Attention feature fusion processing refers to the process of calculating two types of attention weights based on prediction requirements and applying them to the encoded feature matrix to strengthen key information and weaken redundant information.

[0209] Time-axis attention refers to calculating the association weights between features at different time steps in the encoded feature matrix based on the time granularity of the prediction demand (such as monthly or weekly), which can capture the dependencies in the time-series dimension. Variable-axis attention refers to calculating the association weights between different features in the encoded feature matrix, which can capture the interaction relationships in the feature dimension.

[0210] The dual-axis fusion feature matrix refers to the matrix obtained by fusing the time-axis attention output feature matrix and the variable-axis attention output feature matrix in a preset manner, which contains both temporal dependency information and feature interaction information.

[0211] In one example, for the time granularity of the predicted demand, the dual-axis attention module first analyzes the similarity of features at each time step in the encoded feature matrix to obtain the time axis attention weights, and then applies these time axis attention weights to the encoded feature matrix to obtain the time axis output matrix; next, it analyzes the similarity between features in the encoded feature matrix to calculate the variable axis attention weights, and then applies them to the encoded feature matrix to obtain the variable axis output matrix; finally, it uses a weighted summation method to fuse the two output matrices to obtain the dual-axis fused feature matrix.

[0212] Specifically, this step may include the following sub-steps:

[0213] S4061. Perform decoding preprocessing on the encoded feature matrix to obtain the preprocessed encoded feature matrix.

[0214] For example, decoding preprocessing refers to a series of adaptation operations performed on the characteristics of the encoded feature matrix, such as dimensional adjustment, numerical normalization, and redundant information filtering, in order to optimize the format and distribution of the feature matrix and improve the accuracy of attention weight calculation.

[0215] The preprocessed encoded feature matrix refers to the feature matrix that has been preprocessed by decoding, has uniform dimensions, balanced numerical distribution, and has had redundant information removed.

[0216] In one example, based on the dual-axis attention module, the encoded feature matrix is ​​preprocessed by performing decoding preprocessing such as temporal calibration, noise reduction, dimensionality optimization, and numerical normalization to obtain the preprocessed encoded feature matrix.

[0217] S4062. Determine the time axis attention weights based on the temporal granularity represented by the prediction requirements and the preprocessed encoded feature matrix.

[0218] For example, the time granularity represented by the prediction demand refers to the time unit (e.g., month, week, day) of the prediction result expected by the user. This is the core guideline for calculating the time axis attention weight and determines the criteria for judging the correlation strength of features at different time steps. The time axis attention weight refers to the numerical value of the correlation strength between the dynamic attribute feature vectors of any two time steps in the quantized preprocessed encoded feature matrix. The higher the weight, the stronger the feature correlation between the corresponding two time steps, and the greater the contribution to the prediction result. It can capture the dependency relationship in the temporal dimension and strengthen the feature contribution of key time steps.

[0219] In one example, the time granularity for predicting demand is "monthly," and the preprocessed encoded feature matrix has a dimension of 24×246 (corresponding to features for 24 consecutive months). Based on this time granularity, a "self-attention mechanism" is used to calculate the time axis attention weights: the feature vector of each time step is used as the query vector, and the feature vectors of all time steps are used as key vectors and value vectors. The semantic similarity between the query vector and each key vector is calculated. Finally, a 24×246 time axis attention weight matrix is ​​generated, where each element represents the association weight between two corresponding time steps.

[0220] S4063. The preprocessed encoded feature matrix is ​​weighted based on the time axis attention weight to obtain the time axis attention output feature matrix.

[0221] For example, weighted processing refers to performing matrix multiplication on the preprocessed encoded feature matrix and the time axis attention weight matrix, so that the feature vector of each time step is adjusted by the weight of its associated time step, so as to strengthen the feature contribution of high-weight time steps and weaken the influence of low-weight time steps.

[0222] The time-axis attention output feature matrix refers to the feature matrix generated after weighted processing. The row dimension is still the time step, and the column dimension is still the feature dimension. However, the feature vector of each time step has been integrated with the effective information of the associated time steps, which more accurately represents the temporal dependency relationship.

[0223] S4064. Based on the preprocessed encoded feature matrix, determine the attention weights of the variable axes.

[0224] For example, the variable axis attention weight refers to the numerical value of the correlation strength between any two feature vectors in the quantized preprocessed encoded feature matrix. The higher the weight, the stronger the correlation between the two corresponding features, the greater the collaborative contribution to the prediction target, the more it can capture the interaction relationship in the feature dimension, and strengthen the collaborative contribution of key features.

[0225] In one example, each feature vector is used as the query vector, and all feature vectors are used as key vectors and value vectors. The cosine similarity between the query vector and each key vector is calculated to obtain the variable axis attention weight matrix. Each element in the matrix represents the association weight between the corresponding two feature vectors.

[0226] S4065. The preprocessed encoded feature matrix is ​​weighted based on the variable axis attention weights to obtain the variable axis attention output feature matrix.

[0227] For example, weighted processing refers to performing matrix multiplication on the preprocessed encoded feature matrix and the variable axis attention weight matrix, so that each feature vector is adjusted by the weight of its associated feature vectors, in order to strengthen the collaborative contribution of high-weight features and weaken the interference of low-weight features.

[0228] The variable axis attention output feature matrix refers to the feature matrix generated after weighted processing. The row dimension is still the time step, and the column dimension is still the feature dimension, but each feature vector has incorporated the effective information of the associated features, which more accurately represents the interaction relationship between features.

[0229] S4066. Perform weighted fusion processing on the time axis attention output feature matrix and the variable axis attention output feature matrix to obtain the dual-axis fusion feature matrix.

[0230] For example, weighted fusion processing refers to the process of element-wise weighted summation of two matrices based on preset fusion coefficients to balance the contributions of temporal dependency information and feature interaction information. The dual-axis fusion feature matrix refers to the feature matrix generated after weighted fusion, with the row dimension representing the time step and the column dimension representing the feature dimension, containing comprehensive information on both temporal dependency and feature interaction relationships.

[0231] In one example, weighted fusion coefficients corresponding to different prediction requirements can be preset, and then the current weighted fusion coefficients can be determined based on the current prediction requirements. Furthermore, based on the current weighted fusion coefficients, element-wise weighted summation is performed on the time axis attention output feature matrix and the variable axis attention output feature matrix to obtain the dual-axis fusion feature matrix.

[0232] S407. Based on the modulation module in the multi-source feature fusion prediction model, the dual-axis fusion feature matrix is ​​modulated according to the prediction requirements to obtain the modulated feature matrix.

[0233] For example, the modulation module is a core component in the multi-source feature fusion prediction model used for feature channel selection and scene adaptation, including two sub-modules: Gated Residual Network (GRN) and Feature-wise Linear Modulation (FiLM).

[0234] Modulation processing refers to the process of first filtering effective feature channels through GRN, and then dynamically adjusting feature weights according to prediction requirements through FiLM to achieve accurate adaptation of features to the scene.

[0235] GRN is used to filter the feature channels of the dual-axis fused feature matrix through a gating mechanism, retaining features that are effective for the prediction requirements and filtering out redundant channel information. FiLM is used to generate scaling factors (γ) and offset factors (β) for each feature channel based on the scene label corresponding to the prediction requirements (such as period prediction, work hour prediction). Through a linear transformation of "eigenvalue × scaling factor + offset factor", the weight of each feature channel is dynamically adjusted to enhance the contribution of scene-related features.

[0236] The modulated feature matrix refers to the feature matrix obtained after channel filtering and scenario-based modulation, retaining only the effective features and adapting the weights of each channel to the current prediction requirements.

[0237] In one example, for the prediction scenario represented by the prediction demand, the modulation module first performs gated filtering on each feature channel of the dual-axis fusion feature matrix through GRN, eliminating redundant channels with a correlation of less than a preset threshold with the prediction scenario of the prediction demand, and retaining valid channels; then, based on the scenario label of the prediction demand, it generates scaling and offset coefficients corresponding to the filtered feature channels; finally, it performs channel-by-channel linear modulation on the filtered feature matrix through FiLM to obtain the modulated feature matrix adapted to the resource demand prediction scenario.

[0238] Specifically, this step may include the following sub-steps:

[0239] S4071. Based on the gated residual network in the modulation module, the feature channels of the dual-axis fusion feature matrix are filtered to obtain the filtered feature matrix.

[0240] For example, a feature channel refers to the feature dimension corresponding to the column dimension of the feature matrix, with each channel corresponding to a type of feature. The filtering process refers to the GRN calculating the importance score of each feature channel using gating units, retaining valid channels with scores above a preset threshold, and removing redundant channels with scores below the threshold. The filtered feature matrix refers to the matrix that retains only the feature channels strongly correlated with the prediction requirements after removing redundant channels.

[0241] S4072. Based on the scene labels represented by the predicted demand, determine the scaling factor and offset factor corresponding to the scene labels.

[0242] For example, as mentioned above, the prediction requirement may include a prediction scenario. The scenario label refers to a contextualized identifier for the prediction scenario, used to determine the adaptation weight of each feature channel. The scaling factor (γ) is a magnification or reduction factor individually set for each feature channel in the filtered feature matrix, used to adjust the contribution strength of the channel feature (a factor greater than 1 strengthens the channel, and a factor less than 1 weakens the channel). The offset factor (β) is an offset individually set for each feature channel, used to adjust the numerical distribution of the channel feature to adapt to the contextualized prediction requirement.

[0243] For example, a pre-defined correspondence between scene labels and coefficients can be established, and the scaling coefficient and offset coefficient corresponding to the scene label can be determined based on this correspondence.

[0244] S4073. Based on the feature-channel linear modulation module in the modulation module, the filtered feature matrix is ​​modulated based on the scaling factor and the offset factor to obtain the modulated feature matrix.

[0245] For example, modulation processing refers to the FiLM module performing an element-wise linear operation of "eigenvalue × scaling factor + offset factor" on each element in the filtered feature matrix according to the scaling factor and offset factor of its respective feature channel, so as to enhance the feature contribution of the scene-related core channels and adjust the numerical distribution of non-core channels, making the feature matrix more suitable for the current prediction needs.

[0246] S408. Based on the prediction head in the multi-source feature fusion prediction model, the modulated feature matrix is ​​processed for prediction to obtain the prediction result corresponding to the prediction requirements.

[0247] For example, the prediction head is the core component in a multi-source feature fusion prediction model used to output the prediction result. Essentially, it is a mapping network composed of fully connected layers, capable of converting a high-dimensional feature matrix into a quantized result of the prediction target. Prediction processing refers to a series of processes performed by the prediction head on the modulated feature matrix, such as linear transformations and activation function operations, mapping the feature information into a quantized output that matches the prediction requirements.

[0248] In one example, the prediction head first performs global pooling on the modulated feature matrix, then performs a linear transformation through two fully connected layers to map the feature vector to the output vector; finally, it maps the output value to a reasonable numerical range through a preset activation function, calculates the confidence interval corresponding to each predicted value, and finally outputs the prediction result.

[0249] Optionally, based on existing prediction models and the multi-source feature fusion prediction model provided in this application, predictions for the same project data and the same prediction target are performed. The quantile losses of the existing prediction models and the multi-source feature fusion prediction model are 0.055 and 0.046, respectively, significantly better than the existing prediction models, indicating that the multi-source feature fusion prediction model provided in this application can improve the prediction accuracy of the prediction results. Meanwhile, refer to... Figure 5 As shown, Figure 5 This is a schematic diagram of a prediction curve provided in this application, where the horizontal axis is the time index, with negative numbers (-175 to 0) representing historical observation periods, 0 corresponding to the current time, and positive numbers (0 to 25) representing future prediction periods. The left vertical axis corresponds to the value of the target variable, and the right vertical axis corresponds to the attention weight. Figure 5 In the two sample data tests shown, the blue curve in the figure represents the actual observed values ​​for historical periods, the orange curve represents the future predicted values ​​output by the model, and the gray curve represents the attention weights of the model at different time steps. The orange curve has a high degree of overlap with the actual data in the future period, indicating that the model has high prediction accuracy.

[0250] The project data processing method based on a large model provided in this application first obtains the initial project data and prediction requirements of the target project. Then, the large model automatically corrects the initial project data to obtain project data with clear field attributes and meeting data quality standards. Next, the large model, combined with the project data and prediction requirements, makes targeted adjustments to the preset initial prediction model to obtain a multi-source feature fusion prediction model that matches the prediction requirements. Subsequently, this model extracts fusion feature vectors matching the prediction requirements from the project data. A long sequence modeling module encodes the long-time-series dynamic features step-by-step to obtain an encoded feature matrix. A dual-axis attention module then performs attention fusion from both time and feature dimensions to obtain a dual-axis fusion feature matrix. A modulation module performs channel filtering and scenario-based modulation on this matrix to obtain a modulated feature matrix that matches the prediction requirements. Finally, a prediction head processes the modulated feature matrix to output accurate prediction results. This method relies on the modular design of a multi-source feature fusion prediction model. Through a long sequence modeling module, it effectively captures long-term temporal dependencies; a dual-axis attention module strengthens multi-dimensional associations; and a modulation module achieves precise scene adaptation. It fully explores the intrinsic value of project data and significantly improves the accuracy and reliability of prediction results. At the same time, the single model architecture can be adapted to multiple prediction needs of the same project through fine-tuning, without the need to deploy multiple models. This greatly reduces the cost of model development, training, and maintenance, simplifies the operation process of project time series prediction, and provides efficient and accurate data support for project management decisions.

[0251] Figure 6 A flowchart illustrating the method for processing project data based on a large model provided in this application. Figure 5 , Figure 7 A schematic diagram illustrating an application scenario provided in this application is shown below. Figure 6 and Figure 7 As shown, based on a large model, the system receives the user's input prediction requirements and initial project data for the target project. Then, based on the large model and a pre-set knowledge base, it first classifies the field attributes of each field in the initial project data. Then, it generates a query vector by combining field attributes and field content, and retrieves the correction method for the field from the pre-set knowledge base. Based on the pre-set toolchain in the large model, it uses the correction method corresponding to the field to correct the initial project data, thus obtaining the project data to be processed.

[0252] Furthermore, based on the large model, the preset initial prediction model is fine-tuned and trained using the project data to be processed to obtain an intermediate prediction model and a state vector. Based on the large model, a parameter adjustment strategy is determined according to the state vector and a preset knowledge base. After adjusting the intermediate prediction model based on the parameter adjustment strategy, iterative training is performed using the project data to be processed until an intermediate prediction model whose state vector meets the preset conditions is obtained as a multi-source feature fusion prediction model. A model description document including model parameters is then generated.

[0253] Furthermore, based on the dynamic feature selection module in the trained multi-source feature fusion prediction model, initial dynamic attribute feature vectors can be extracted from the dynamic attribute fields of the project to be processed. The weights of each initial dynamic attribute feature vector are determined according to the prediction requirements, and the initial dynamic attribute feature vectors are filtered based on these weights to obtain dynamic attribute feature vectors. Then, based on the static attribute injection module in the multi-source feature fusion prediction model, static attribute feature vectors are extracted from the static attribute fields of the project to be processed. The two are then concatenated to obtain a fused feature vector. Subsequently, based on the long sequence modeling module in the multi-source feature fusion prediction model, a QGRU network is used to encode the long-time sequence portion of the fused feature vector step-by-step in chronological order to obtain an encoded feature matrix. Based on the dual-axis attention module in the multi-source feature fusion prediction model, the time axis attention weights and variable axis attention weights are calculated separately according to the prediction requirements, thereby obtaining the time axis attention output feature matrix and the variable axis attention output feature matrix, which are then fused to obtain the dual-axis fused feature matrix. Finally, based on GRN and FiLM in the multi-source feature fusion prediction model, the feature channels in the dual-axis fusion feature matrix are modulated according to the scene labels in the prediction requirements before prediction processing is performed to obtain the prediction results.

[0254] In summary, the method for processing project data based on large models provided in this application has the following advantages:

[0255] (1) The data field classification, problem identification and correction are completed automatically through the large model, which completely eliminates the dependence on manual operation, greatly improves the efficiency and quality of data processing, and avoids errors caused by differences in human experience.

[0256] (2) Through the model iteration and optimization mechanism driven by the large model, the parameter adjustment strategy can be automatically determined based on the preset knowledge base without manual judgment, which ensures the deep adaptation of the multi-source feature fusion prediction model to the target project data patterns and prediction needs, and significantly improves the accuracy and stability of the prediction results.

[0257] (3) Based on the single model architecture, multi-source feature fusion and time series reasoning are realized. There is no need to deploy multiple models for multiple prediction needs of the same project. The inherent relationship of project data is fully explored. While ensuring the consistency of results under different prediction needs, the model development, training and maintenance costs are greatly reduced, the operation process of project time series prediction is simplified, and efficient and reliable data support is provided for project management decision-making.

[0258] (4) Based on the modular design of the multi-source feature fusion prediction model, the long sequence modeling module effectively captures long temporal dependencies, the dual-axis attention module strengthens multi-dimensional associations, and the modulation module achieves accurate scene adaptation, fully exploring the intrinsic value of project data and significantly improving the accuracy and reliability of prediction results.

[0259] Figure 8 A schematic diagram of the structure of the device for processing project data based on a large model provided in this application is shown below. Figure 8 As shown, the project data processing device 500 based on a large model provided in this embodiment includes:

[0260] Module 501 is used to acquire initial project data and predicted requirements for the target project.

[0261] The correction module 502 is used to correct the initial project data based on the large model to obtain the project data to be processed; the project data to be processed includes static attribute fields and dynamic attribute fields.

[0262] Training module 503 is used to adjust the preset initial prediction model based on the large model, according to the data of the project to be processed and the prediction requirements, to obtain a multi-source feature fusion prediction model that matches the prediction requirements.

[0263] The extraction module 504 is used to extract the fusion feature vector that matches the prediction requirements from the static attribute fields and dynamic attribute fields of the project data to be processed, based on the multi-source feature fusion prediction model and prediction requirements.

[0264] The prediction module 505 is used to process the fused feature vector based on the multi-source feature fusion prediction model and prediction requirements to obtain the prediction result corresponding to the prediction requirements.

[0265] In one possible implementation, the correction module 502 is used for:

[0266] Based on the large model, the fields in the initial project data are classified by field attributes to obtain the project data to be corrected; each field in the project data to be corrected has field attributes, including static attribute fields and dynamic attribute fields.

[0267] Based on the large model, determine the problem types of the fields to be corrected in the project data;

[0268] Based on the preset knowledge base, the question type of the field to be corrected, and the field attributes of the field to be corrected, the correction method corresponding to the field to be corrected is determined; wherein, the preset knowledge base includes the correction methods corresponding to question types with different field attributes;

[0269] Based on the pre-set toolchain in the large model, the field to be corrected is corrected according to the correction method corresponding to the field to be corrected, and the data of the project to be processed is obtained; the pre-set toolchain has pre-set data correction logic.

[0270] In one possible implementation, the correction module 502 is used for:

[0271] For the fields to be classified in the initial project data, a query vector is constructed based on the field information of the fields to be classified;

[0272] Based on the preset knowledge base and query vector, determine the associated information that matches the field to be classified from the preset knowledge base;

[0273] Based on the large model and the associated information, the field attributes of the fields to be classified are determined; among them, the project data to be corrected is the initial project data after the field attributes are classified.

[0274] In one possible implementation, the extraction module 504 is used for:

[0275] Based on the dynamic feature selection module in the multi-source feature fusion prediction model, the dynamic attribute fields of the project data to be processed are vectorized to obtain the initial dynamic attribute feature vectors corresponding to each dynamic attribute field.

[0276] Based on the dynamic feature selection module, the initial dynamic attribute feature vector is filtered according to the prediction requirements to obtain the dynamic attribute feature vector.

[0277] Based on the static attribute injection module in the multi-source feature fusion prediction model, static attribute feature vectors are extracted from the static attribute fields of the project data to be processed; wherein, the dimensions of the static attribute feature vectors are consistent with those of the dynamic attribute feature vectors.

[0278] Based on the encoding module in the multi-source feature fusion prediction model, the static attribute feature vector and the dynamic attribute feature vector are concatenated to obtain the fused feature vector.

[0279] In one possible implementation, the dynamic attribute field includes a dynamic historical attribute field and a dynamic future-knowable attribute field; the extraction module 504 is used for:

[0280] Based on the dynamic feature selection module, an attention mechanism is used to calculate the importance weight of each initial dynamic attribute feature vector to the prediction requirement; among them, the initial dynamic attribute feature vectors corresponding to the dynamic historical attribute fields and the initial dynamic attribute feature vectors corresponding to the dynamic future knowable attribute fields are assigned weight calculation preferences during the calculation process.

[0281] Based on importance weights, the initial dynamic attribute feature vectors are filtered to obtain dynamic attribute feature vectors.

[0282] In one possible implementation, the static attribute fields include categorical static attribute fields and numeric static attribute fields; the extraction module 504 is used for:

[0283] Based on the static attribute injection module, the categorical static attribute fields in the project data to be processed are embedded and transformed to obtain the first vector;

[0284] Based on the static attribute injection module, the numerical static attribute fields in the project data to be processed are normalized, and the normalized numerical static attribute fields are linearly transformed to obtain the second vector.

[0285] Based on the first and second vectors, the static attribute feature vector is obtained.

[0286] In one possible implementation, the fused feature vector includes dynamic attribute feature vectors and static attribute feature vectors; the prediction module 505 is used for:

[0287] Based on the long sequence modeling module in the multi-source feature fusion prediction model, a pre-defined gated unit network is used to encode a portion of the feature vectors in the fused feature vectors step by step to obtain the encoded feature matrix; among them, the portion of the feature vectors represents dynamic attribute feature vectors with a time sequence length greater than a preset threshold.

[0288] Based on the dual-axis attention module in the multi-source feature fusion prediction model, attention feature fusion processing is performed on the encoded feature matrix according to the prediction requirements to obtain the dual-axis fused feature matrix;

[0289] Based on the modulation module in the multi-source feature fusion prediction model, the dual-axis fusion feature matrix is ​​modulated according to the prediction requirements to obtain the modulated feature matrix.

[0290] Based on the prediction head in the multi-source feature fusion prediction model, the modulated feature matrix is ​​processed for prediction to obtain the prediction result corresponding to the prediction requirements.

[0291] In one possible implementation, the prediction module 505 is used for:

[0292] The encoded feature matrix is ​​preprocessed by decoding to obtain the preprocessed encoded feature matrix;

[0293] Based on the temporal granularity represented by the prediction demand and the preprocessed encoded feature matrix, the temporal attention weights are determined; where the temporal attention weights represent the weights between dynamic attribute feature vectors at different time steps in the preprocessed encoded feature matrix.

[0294] The preprocessed encoded feature matrix is ​​weighted based on the time axis attention weights to obtain the time axis attention output feature matrix.

[0295] Based on the preprocessed encoded feature matrix, the variable axis attention weights are determined; whereby the variable axis attention weights represent the weights between different feature vectors in the preprocessed encoded feature matrix.

[0296] The preprocessed encoded feature matrix is ​​weighted based on the variable axis attention weights to obtain the variable axis attention output feature matrix.

[0297] The time-axis attention output feature matrix and the variable-axis attention output feature matrix are weighted and fused to obtain the dual-axis fused feature matrix.

[0298] In one possible implementation, the prediction module 505 is used for:

[0299] Based on the gated residual network in the modulation module, the feature channels of the dual-axis fusion feature matrix are filtered to obtain the filtered feature matrix.

[0300] Based on the scene labels represented by the predicted demand, determine the scaling factor and offset factor corresponding to the scene labels;

[0301] Based on the feature-channel linear modulation module in the modulation module, the filtered feature matrix is ​​modulated using scaling and offset coefficients to obtain the modulated feature matrix.

[0302] In one possible implementation, the training module 503 is used for:

[0303] Based on the data of the project to be processed and the prediction requirements, the initial prediction model is trained to obtain the intermediate prediction model;

[0304] Based on the large model, the state vector of the intermediate prediction model is determined; where the state vector represents the model accuracy of the initial prediction model.

[0305] If it is determined that the state vector does not meet the preset conditions, then based on the large model, the parameters of the intermediate prediction model are adjusted according to the state vector until the model accuracy of the intermediate prediction model meets the preset conditions.

[0306] Among them, the intermediate prediction model obtained when the preset conditions are met is the multi-source feature fusion prediction model.

[0307] In one possible implementation, the state vector includes one or more of the following: training loss curve shape, gradient stability, model output bias structure information, data state change information, and scene description labels.

[0308] Among them, the training loss curve shape represents the shape of the loss function curve of the intermediate prediction model; gradient stability represents the gradient norm of the intermediate prediction model; model output deviation structure information represents the deviation distribution characteristics between the prediction results and the true values ​​of the intermediate prediction model; data state change information represents the distribution characteristics of the data to be processed; and scene description label represents the scene identifier corresponding to the prediction requirement.

[0309] The processing device for project data based on a large model provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0310] Figure 9 A schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 600 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the electronic device 600 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus.

[0311] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0312] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0313] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0314] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0315] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0316] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0317] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0318] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0319] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0320] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0321] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0322] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0323] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0324] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0325] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for processing project data based on a large model, characterized in that, The method includes: Obtain initial project data and projected requirements for the target project; Based on the large model, the initial project data is corrected to obtain the project data to be processed; wherein, the project data to be processed includes static attribute fields and dynamic attribute fields; Based on the large model, the preset initial prediction model is adjusted according to the project data to be processed and the prediction requirements to obtain a multi-source feature fusion prediction model that matches the prediction requirements. Based on the multi-source feature fusion prediction model and the prediction requirements, a fusion feature vector matching the prediction requirements is extracted from the static attribute fields and dynamic attribute fields of the project data to be processed. Based on the multi-source feature fusion prediction model and the prediction requirements, the fused feature vector is processed to obtain the prediction result corresponding to the prediction requirements.

2. The method according to claim 1, characterized in that, The process of correcting the initial project data based on the large model to obtain the project data to be processed includes: Based on the large model, the fields in the initial project data are classified by field attributes to obtain the project data to be corrected; wherein, each field in the project data to be corrected has field attributes, and the field attributes include static attribute fields and dynamic attribute fields; Based on the large model, determine the problem type of the field to be corrected in the project data to be corrected; Based on a preset knowledge base, the question type of the field to be corrected, and the field attributes of the field to be corrected, the correction method corresponding to the field to be corrected is determined; wherein, the preset knowledge base includes correction methods corresponding to question types with different field attributes; Based on the preset toolchain in the large model, the field to be corrected is corrected according to the correction method corresponding to the field to be corrected, and the data of the project to be processed is obtained; wherein, the preset toolchain has preset data correction logic.

3. The method according to claim 2, characterized in that, Based on a large model, the fields in the initial project data are classified according to their attributes to obtain the project data to be corrected, including: For the fields to be classified in the initial project data, a query vector is constructed based on the field information of the fields to be classified; Based on the preset knowledge base and the query vector, determine the association information that matches the field to be classified from the preset knowledge base; Based on the large model, the field attributes of the field to be classified are determined according to the association information; wherein, the project data to be corrected is the initial project data after the field attributes are classified.

4. The method according to claim 1, characterized in that, The step of extracting a fusion feature vector matching the prediction requirement from the static and dynamic attribute fields of the project data to be processed, based on the multi-source feature fusion prediction model and the prediction requirement, includes: Based on the dynamic feature selection module in the multi-source feature fusion prediction model, the dynamic attribute fields of the project data to be processed are vectorized to obtain the initial dynamic attribute feature vectors corresponding to each dynamic attribute field. Based on the dynamic feature selection module, the initial dynamic attribute feature vector is filtered according to the prediction requirements to obtain the dynamic attribute feature vector. Based on the static attribute injection module in the multi-source feature fusion prediction model, static attribute feature vectors are extracted from the static attribute fields of the project data to be processed; wherein, the static attribute feature vectors have the same dimension as the dynamic attribute feature vectors. Based on the encoding module in the multi-source feature fusion prediction model, the static attribute feature vector and the dynamic attribute feature vector are concatenated to obtain the fused feature vector.

5. The method according to claim 4, characterized in that, The dynamic attribute fields include dynamic historical attribute fields and dynamic future predictable attribute fields; the dynamic feature selection module filters the initial dynamic attribute feature vector according to the prediction requirements to obtain a dynamic attribute feature vector, including: Based on the dynamic feature selection module, an attention mechanism is used to calculate the importance weight of each initial dynamic attribute feature vector to the prediction requirement; wherein, the initial dynamic attribute feature vector corresponding to the dynamic historical attribute field and the initial dynamic attribute feature vector corresponding to the dynamic future knowable attribute field are assigned weight calculation preferences during the calculation process. Based on the importance weights, the initial dynamic attribute feature vectors are filtered to obtain dynamic attribute feature vectors.

6. The method according to claim 4, characterized in that, The static attribute fields include categorical static attribute fields and numerical static attribute fields; the static attribute injection module based on the multi-source feature fusion prediction model extracts static attribute feature vectors from the static attribute fields of the project data to be processed, including: Based on the static attribute injection module, the categorized static attribute fields in the project data to be processed are embedded and transformed to obtain the first vector. Based on the static attribute injection module, the numerical static attribute fields in the project data to be processed are normalized, and the normalized numerical static attribute fields are linearly transformed to obtain the second vector. Based on the first vector and the second vector, the static attribute feature vector is obtained.

7. The method according to claim 1, characterized in that, The fused feature vector includes dynamic attribute feature vectors and static attribute feature vectors; the process of processing the fused feature vectors based on the multi-source feature fusion prediction model and the prediction requirement to obtain the prediction result corresponding to the prediction requirement includes: Based on the long sequence modeling module in the multi-source feature fusion prediction model, a preset gated unit network is used to encode a portion of the feature vectors in the fused feature vectors step by step to obtain an encoded feature matrix; wherein, the portion of the feature vectors represents dynamic attribute feature vectors with a time sequence length greater than a preset threshold. Based on the dual-axis attention module in the multi-source feature fusion prediction model, attention feature fusion processing is performed on the encoded feature matrix according to the prediction requirements to obtain the dual-axis fused feature matrix. Based on the modulation module in the multi-source feature fusion prediction model, the dual-axis fusion feature matrix is ​​modulated according to the prediction requirements to obtain the modulated feature matrix. Based on the prediction head in the multi-source feature fusion prediction model, the modulated feature matrix is ​​subjected to prediction processing to obtain the prediction result corresponding to the prediction requirement.

8. The method according to claim 7, characterized in that, The dual-axis attention module in the multi-source feature fusion prediction model performs attention feature fusion processing on the encoded feature matrix according to the prediction requirements to obtain a dual-axis fused feature matrix, including: The encoded feature matrix is ​​preprocessed by decoding to obtain the preprocessed encoded feature matrix; Based on the time granularity represented by the prediction requirement and the preprocessed encoding feature matrix, the time axis attention weight is determined; wherein, the time axis attention weight represents the weight between dynamic attribute feature vectors at different time steps in the preprocessed encoding feature matrix; The preprocessed encoded feature matrix is ​​weighted based on the time axis attention weights to obtain the time axis attention output feature matrix. Based on the preprocessed encoding feature matrix, variable axis attention weights are determined; wherein, the variable axis attention weights characterize the weights between different feature vectors in the preprocessed encoding feature matrix. The preprocessed encoded feature matrix is ​​weighted based on the variable axis attention weights to obtain the variable axis attention output feature matrix. The time-axis attention output feature matrix and the variable-axis attention output feature matrix are weighted and fused to obtain the dual-axis fused feature matrix.

9. The method according to claim 7, characterized in that, The modulation module in the multi-source feature fusion prediction model modulates the dual-axis fusion feature matrix according to the prediction requirements to obtain a modulated feature matrix, including: Based on the gated residual network in the modulation module, the feature channels of the dual-axis fusion feature matrix are filtered to obtain the filtered feature matrix. Based on the scene labels represented by the predicted demand, determine the scaling factor and offset factor corresponding to the scene labels; Based on the feature-channel linear modulation module in the modulation module, the filtered feature matrix is ​​modulated based on the scaling factor and the offset factor to obtain the modulated feature matrix.

10. The method according to any one of claims 1-9, characterized in that, Based on the large model, and according to the project data to be processed and the prediction requirements, the preset initial prediction model is adjusted to obtain a multi-source feature fusion prediction model that matches the prediction requirements, including: Based on the data of the project to be processed and the prediction requirements, the initial prediction model is trained to obtain an intermediate prediction model; Based on the large model, the state vector of the intermediate prediction model is determined; wherein, the state vector represents the model accuracy of the initial prediction model; If it is determined that the state vector does not meet the preset conditions, then based on the large model, the parameters of the intermediate prediction model are adjusted according to the state vector until the model accuracy of the intermediate prediction model meets the preset conditions. The intermediate prediction model obtained when the preset conditions are met is the multi-source feature fusion prediction model.

11. The method according to claim 10, characterized in that, The state vector includes one or more of the following: training loss curve shape, gradient stability, model output bias structure information, data state change information, and scene description labels; Wherein, the training loss curve shape represents the shape of the loss function curve of the intermediate prediction model; the gradient stability represents the gradient norm of the intermediate prediction model; the model output deviation structure information represents the deviation distribution characteristics between the prediction results and the true values ​​of the intermediate prediction model; the data state change information represents the distribution characteristics of the data to be processed; and the scene description label represents the scene identifier corresponding to the prediction requirement.

12. A device for processing project data based on a large model, characterized in that, include: The acquisition module is used to acquire initial project data and predicted requirements for the target project. The correction module is used to correct the initial project data based on the large model to obtain project data to be processed; wherein the project data to be processed includes static attribute fields and dynamic attribute fields. The training module is used to adjust the preset initial prediction model based on the large model, the project data to be processed, and the prediction requirements, so as to obtain a multi-source feature fusion prediction model that matches the prediction requirements. The extraction module is used to extract a fusion feature vector that matches the prediction requirement from the static attribute fields and dynamic attribute fields of the project data to be processed, based on the multi-source feature fusion prediction model and the prediction requirement. The prediction module is used to process the fused feature vector based on the multi-source feature fusion prediction model and the prediction requirement to obtain the prediction result corresponding to the prediction requirement.

13. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-11.

Images