Order management method and system based on big data and artificial intelligence

Abstract

The present invention proposes an order management method and system based on big data and artificial intelligence, and the method includes: obtaining order data collected by an e-commerce platform and preprocessing; extracting features from the order data, building a multi-layer detection model to detect and fuse the features; automatically processing abnormal orders using RPA technology and smart contracts according to the results of abnormal detection; using natural language processing technology to identify and classify questions related to abnormal orders raised by users, matching corresponding abnormal order solutions, and then generating natural reply content; using big data processing technology to collect data from multiple channels and preprocessing, then extracting features from the preprocessed data, and fusing features to obtain feature data, analyzing and modeling based on the fused feature data, and then optimizing and supporting decisions based on the analysis and modeling results. The present invention combines big data and artificial intelligence technology to provide a complete intelligent order management system.

Description

Technical Field

[0001] The present invention belongs to the field of big data and artificial intelligence, and in particular to an order management method and system based on big data and artificial intelligence. Background Art

[0002] With the rapid development of e-commerce, order management and intelligent marketing have become key links for e-commerce platforms to improve operational efficiency and sales performance. Existing order management systems and marketing strategies have many shortcomings in processing large-scale data, realizing intelligent operations, and providing personalized services. Traditional order management systems mainly rely on manual operations and simple rule engines, which are difficult to detect and handle order anomalies in real time, such as fraudulent orders, duplicate orders, etc. At the same time, customer service and after-sales management rely on manual customer service, which is inefficient and costly, and it is difficult to meet the growing needs of users. In addition, existing marketing strategies are mostly based on historical data analysis, lacking real-time and accuracy, and difficult to dynamically adjust to adapt to market changes and user personalized needs.

[0003] In terms of data integration and analysis, existing technologies are often unable to effectively integrate data from different channels, resulting in serious data silos and difficulty in providing comprehensive business insights and decision support. Traditional data analysis methods mainly rely on preset statistical models and cannot fully utilize the advantages of big data and artificial intelligence technologies, making it difficult to achieve accurate sales forecasts and customer segmentation. Summary of the invention

[0004] The purpose of this invention is to propose an order management method and system based on big data and artificial intelligence, combining big data and artificial intelligence technology to provide a complete intelligent order management and marketing system to improve the platform's operational efficiency and sales performance. By integrating the latest cutting-edge technologies such as deep learning, reinforcement learning, and natural language processing, this invention will achieve comprehensive intelligence and automation in order management, customer service, and data analysis.

[0005] In order to achieve the above object, the present invention provides an order management method based on big data and artificial intelligence, the method comprising:

[0006] S1. Obtain order data collected by the e-commerce platform and perform preprocessing, perform data fusion, normalization, data dimension reduction and selection on the preprocessed data, and then classify, label and store the selected data;

[0007] S2. Extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse anomalies of features;

[0008] S3. Automatically process abnormal orders using RPA technology and smart contracts based on the results of abnormal detection;

[0009] S4. Use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match corresponding abnormal order solutions from the knowledge graph based on the classification results, then generate natural reply content and interact with users;

[0010] S5. Use big data processing technology to collect data from multiple channels and perform preprocessing. Then extract features from the preprocessed data and perform feature fusion to obtain feature data. Perform analysis and modeling based on the fused feature data. Then, perform optimization and decision support based on the analysis and modeling results.

[0011] The multi-layer detection model is divided into primary detection, advanced detection and time series detection:

[0012] The primary detection uses a rule-based approach to quickly screen orders and identify obvious abnormal orders, as shown below:

[0013]

[0014] Among them, T j represents the threshold of the jth feature; represents the primary test result of the i-th order, with a value of 0 or 1, 0 represents normal and 1 represents abnormal; x i,j represents the jth feature of the i-th order;

[0015] The advanced detection is performed using a deep learning model based on an autoencoder, as shown below:

[0016] h i =σ(W1X i +b1)

[0017]

[0018] Among them, h i represents the hidden layer representation of the i-th order, with dimension h; represents the reconstructed feature vector of the i-th order; W1 and W2 represent the weight matrices of the encoder and decoder respectively, b1 and b2 represent the bias vectors respectively, and σ is the activation function;

[0019] The loss function L of the deep learning model based on the autoencoder is expressed as follows:

[0020]

[0021] Where n represents the number of samples, ∥·∥ 2 represents the Euclidean distance, i represents the i-th sample; X i The feature vector representing the i-th order, with dimension d;

[0022] The detection discrimination of the advanced detection is expressed as follows:

[0023]

[0024] Among them, D i It represents the advanced detection result of the i-th order, and its value is 0 or 1, 0 represents normal and 1 represents abnormal; ∈ represents the preset threshold;

[0025] The time series detection uses the long short-term memory network LSTM to perform time series analysis on the order data and identify abnormal patterns within a specific time period, as shown below:

[0026] H t =LSTM(X t )

[0027]

[0028] Among them, W3 is the output layer weight matrix; b3 is the bias vector; X t Represents the order feature vector sequence at time t, containing the features of the past k time steps; Indicates that based on H t The predicted order feature vector at time t+1; H t represents the hidden layer state at time t;

[0029] Then perform timing detection and judgment:

[0030]

[0031] Among them, δ is the preset timing detection threshold; Indicates the time series detection result at time t, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal; X t+1 Represents the actual order feature vector at time t+1;

[0032] The results of primary detection, advanced detection, and time series detection are integrated to form the final anomaly detection result, which is shown as follows:

[0033]

[0034] Among them, ∨ represents the logical OR operation, Indicates the final test result of the i-th order, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal.

[0035] Furthermore, in S1, order data collected by the e-commerce platform is obtained, and the order data includes online sales data, offline sales data, social media data and customer behavior data.

[0036] Furthermore, in S1, the data dimension reduction and selection adopts a principal component analysis algorithm to calculate the covariance matrix of the data, perform eigenvalue decomposition, select the first k eigenvectors as principal components, and then select the feature that contributes most to the target variable.

[0037] Furthermore, in step S3, abnormal order features are extracted and abnormal orders are marked based on the abnormal order data detected in step S2, and then corresponding processing rules are defined according to different abnormality types, and then RPA technology and smart contracts are used to execute the predefined processing rules.

[0038] Furthermore, the use of RPA technology and smart contracts to execute predefined processing rules specifically includes:

[0039] Start the RPA robot and prepare to execute processing tasks;

[0040] RPA robot logs into the order management system and obtains abnormal order information i ; Among them, O i =[a i ,t i ,q i ,p i ,h i ], a i Indicates the order amount, t i Indicates the order time, q i Indicates the quantity of goods purchased, p i Indicates payment method, h i Indicates the user's historical purchasing behavior;

[0041] According to the processing rules Execute each step in sequence, and record the results and timestamp of each processing step into the database; i represents the processing flow of the i-th order, Represents a series of processing rules defined based on the anomaly detection results of an order.

[0042] Furthermore, the step S4 specifically includes:

[0043] Extract key features from user questions, use a pre-trained language model to convert user questions into feature vectors, use a classification model based on a graph neural network to classify user questions, and then construct a graph structure, where nodes represent user questions and edges represent the similarity between questions;

[0044] According to the customer problem classification results, the corresponding abnormal order solutions are matched from the knowledge graph. i, searching for corresponding solutions in the knowledge graph; wherein the knowledge graph contains knowledge points and solutions related to abnormal order processing, nodes represent knowledge points and solutions, and edges represent relationships between knowledge points or between knowledge points and solutions;

[0045] Design a dialogue management model to track the state of the dialogue and use a generative model to generate natural responses.

[0046] Collect user feedback on each reply as the basis for improving the dialogue management model, adjust the dialogue management model parameters according to user feedback, and optimize the classification and generative models.

[0047] Furthermore, the similarity Sim is calculated as follows:

[0048]

[0049] Among them, Q i Represents problem i, Q j represents problem j, σ represents activation function, and exp represents exponential function;

[0050] Then classification is performed, and the classification formula is expressed as follows:

[0051] C i =GNN(Q i )

[0052] Among them, GNN is a trained graph neural network model, C i Represents question Q i The classification results.

[0053] Furthermore, according to the classification result of the customer problem, the corresponding abnormal order solution is matched from the knowledge graph, wherein the matching process is as follows:

[0054] From the question node in the knowledge graph To begin, search for solution nodes connected to it;

[0055] Determine the best match based on the strength of the relationship between the problem node and the solution node;

[0056] A graph search algorithm is used to find the optimal solution node.

[0057] Furthermore, in the step S5, analysis and modeling are performed based on the fused feature data, the analysis includes clustering analysis and association rules, and the modeling uses a supervised learning method to perform predictive modeling.

[0058] In another aspect of the present invention, an order management system based on big data and artificial intelligence is provided, the system comprising:

[0059] The data collection module is used to obtain order data collected by the e-commerce platform and perform preprocessing, perform data fusion, normalization, data dimension reduction and selection on the preprocessed data, and then classify, label and store the selected data;

[0060] The order analysis module is used to extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse features; based on the results of anomaly detection, RPA technology and smart contracts are used to automatically process abnormal orders;

[0061] The abnormal order processing module is used to use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match the corresponding abnormal order solutions from the knowledge graph based on the classification results, and then generate natural reply content and interact with users;

[0062] The system update module is used to collect data from multiple channels using big data processing technology, perform preprocessing, extract features from the preprocessed data, fuse the features to obtain feature data, perform analysis and modeling based on the fused feature data, and then perform optimization and decision support based on the analysis and modeling results;

[0063] The multi-layer detection model is divided into primary detection, advanced detection and time series detection:

[0064] The primary detection uses a rule-based approach to quickly screen orders and identify obvious abnormal orders, as shown below:

[0065]

[0066] Among them, T j represents the threshold of the jth feature; represents the primary test result of the i-th order, with a value of 0 or 1, 0 represents normal and 1 represents abnormal; x i,j represents the jth feature of the i-th order;

[0067] The advanced detection is performed using a deep learning model based on an autoencoder, as shown below:

[0068] h i =σ(W1X i +b1)

[0069]

[0070] Among them, h i represents the hidden layer representation of the i-th order, with dimension h; represents the reconstructed feature vector of the i-th order; W1 and W2 represent the weight matrices of the encoder and decoder respectively, b1 and b2 represent the bias vectors respectively, and σ is the activation function;

[0071] The loss function L of the deep learning model based on the autoencoder is expressed as follows:

[0072]

[0073] Where n represents the number of samples, ∥·∥ 2 represents the Euclidean distance, i represents the i-th sample; X i The feature vector representing the i-th order, with dimension d;

[0074] The detection discrimination of the advanced detection is expressed as follows:

[0075]

[0076] Among them, D i It represents the advanced detection result of the i-th order, and its value is 0 or 1, 0 represents normal and 1 represents abnormal; ∈ represents the preset threshold;

[0077] The time series detection uses the long short-term memory network LSTM to perform time series analysis on the order data and identify abnormal patterns within a specific time period, as shown below:

[0078] H t =LSTM(X t )

[0079]

[0080] Among them, W3 is the output layer weight matrix; b3 is the bias vector; X t Represents the order feature vector sequence at time t, containing the features of the past k time steps; Indicates that based on H t The predicted order feature vector at time t+1; H t represents the hidden layer state at time t;

[0081] Then perform timing detection and judgment:

[0082]

[0083] Among them, δ is the preset timing detection threshold; Indicates the time series detection result at time t, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal; X t+1 Represents the actual order feature vector at time t+1;

[0084] The results of primary detection, advanced detection, and time series detection are integrated to form the final anomaly detection result, which is shown as follows:

[0085]

[0086] Among them, ∨ represents the logical OR operation, Indicates the final test result of the i-th order, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal.

[0087] The beneficial technical effects of the present invention are at least as follows:

[0088] (1) This invention uses deep learning-based autoencoders and time series anomaly detection algorithms to detect order anomalies (such as fraudulent orders and duplicate orders) in real time. Combined with RPA (robotic process automation) and smart contract technology, it realizes the automated processing and refund of abnormal orders. It detects and processes order anomalies efficiently and accurately, reduces manual intervention, and improves processing speed and accuracy.

[0089] (2) The present invention uses natural language processing (such as GPT model) and knowledge graph to provide customers with intelligent customer service and automatically answer common questions. Use machine learning models (such as decision trees, SVM) to analyze after-sales service data, optimize after-sales processes, and improve customer satisfaction. Provide 7x24 hours of intelligent customer service to improve customer service efficiency and satisfaction. Optimize after-sales service processes to reduce customer complaints and return rates.

[0090] (3) The present invention uses data lake and ETL technology to integrate online and offline sales data, social media data, and customer behavior data to form a unified data platform. Based on big data analysis and deep learning models, it provides decision support such as sales trend forecasting, customer segmentation, and market analysis. It breaks down data silos, realizes omni-channel data integration and analysis, provides comprehensive business insights and decision support, and improves the accuracy and real-time nature of marketing strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0092] Figure 1 This is a flow chart of the order management method based on big data and artificial intelligence of the present invention.

[0093] Figure 2 This is a framework diagram of the order management system based on big data and artificial intelligence of the present invention. DETAILED DESCRIPTION

[0094] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0095] like Figure 1 As shown, the order management method based on big data and artificial intelligence provided by the embodiment of the present invention includes the following steps S1-S5:

[0096] S1. Obtain order data collected by the e-commerce platform and preprocess it. Perform data fusion, normalization, data dimension reduction and selection on the preprocessed data. Then classify, label and store the selected data.

[0097] Specifically, the first step of data integration and preprocessing is to collect data from multiple channels. For e-commerce platforms, the main data sources include: online sales data (orders, product information, transaction records), offline sales data (POS system data), social media data (user comments, likes, shares), and customer behavior data (browsing history, click-through rate, dwell time).

[0098] Then, the data is extracted, transformed and loaded through ETL (Extract, Transform, Load) tools.

[0099] Furthermore, use API interfaces or database connections to obtain data from various data sources. Schedule tasks regularly (for example, once an hour) to ensure the real-time nature of the data. Formula:

[0100] D i (t) = f API (S i ,t) (1)

[0101] Among them, D i (t) represents the data set extracted from the i-th data source at time t, f API is the extraction function, S i is the data source.

[0102] Furthermore, the data formats from different sources are unified into a standard format for internal use. At the same time, duplicate data is removed, incorrect data is corrected, and missing data is filled. Example conversion steps:

[0103] D i ′=T normalize (D i ) (2)

[0104] D i =T clean(D i ′) (3)

[0105] Among them, T normalize and T clean Denote standardization and cleaning operations, respectively, and D i ′ and D i ″ represent the standardized data and cleaned data, respectively.

[0106] Load the transformed data into the data lake to ensure data consistency and accessibility. Use distributed storage (such as Hadoop HDFS or Amazon S3) to store large-scale data. Formula:

[0107]

[0108] Among them, D lake Represents a comprehensive dataset in a data lake.

[0109] Furthermore, in order to fully utilize the advantages of multi-source data, data fusion is required. The purpose of data fusion is to associate data from different sources to form a unified data view.

[0110] Align data based on timestamps or other key fields (such as user ID, order ID). Use interpolation or sampling methods to synchronize data at different times. Formula:

[0111] D aligned =Align(D lake ,KeyFields) (5)

[0112] Among them, D aligned Represents the aligned dataset, Align is the alignment function, and KeyFields is the key field used for alignment.

[0113] Preferably, association rule mining or graph models are used to associate data from different sources. An association matrix is ​​established to represent the relationship between data. Formula:

[0114] R = Association Matrix (D aligned ) (6)

[0115] Where R represents the association matrix, and AssociationMatrix is ​​the association matrix generation function.

[0116] Furthermore, the purpose of data standardization and normalization is to eliminate data scale differences to facilitate subsequent analysis and processing. Calculate the mean and standard deviation of each feature and perform Z-score standardization. Formula:

[0117]

[0118] Among them, x represents the original data value, μ is the mean, σ is the standard deviation, and z is the standardized value.

[0119] Then, the data is mapped to the [0,1] interval and Min-Max normalization is used.

[0120] formula:

[0121]

[0122] Where x represents the original data value, min(x) and max(x) are the minimum and maximum values ​​of the data, respectively, and x′ is the normalized value.

[0123] Furthermore, in order to reduce data dimensions and improve computational efficiency, data dimensionality reduction and feature selection are required.

[0124] Preferably, the covariance matrix of the data is calculated, eigenvalue decomposition is performed, and the first k eigenvectors are selected as principal components. Formula:

[0125]

[0126] X pca =XW (10)

[0127] Among them, Σ is the covariance matrix, n is the number of samples, x i is the i-th sample, μ is the mean vector, W is the eigenvector matrix, X pca is the data after dimensionality reduction.

[0128] Preferably, based on information gain, chi-square test and other methods, the feature that contributes most to the target variable is selected. Formula:

[0129] I(Y;X i )=H(Y)-H(Y|X i ) (11)

[0130] Among them, I(Y;X i ) represents feature X i The information gain of , H(Y) is the entropy of the target variable Y, H(Y|X i ) is a given feature X i The conditional entropy of Y at .

[0131] Preferably, in order to better perform subsequent analysis and processing, the data needs to be classified and labeled. Use decision trees, random forests and other methods to classify data according to predefined rules. Formula:

[0132] T(X)=DecisionTree(X) (12)

[0133] Among them, T(X) is the classification result, DecisionTree is the decision tree classification model, and X is the input data.

[0134] According to the classification results and business needs, the data is labeled, and the labeled categories include order type, customer type, etc. Formula:

[0135] L(X)=Label(T(X)) (13)

[0136] Among them, L(X) is the labeling result, Label is the labeling function, and T(X) is the classification result.

[0137] Finally, the processed data needs to be stored in an efficient data management system to ensure data accessibility and security.

[0138] Use a distributed database (such as Hadoop HDFS or Amazon S3) to store processed data to ensure high availability and reliability of data. Formula:

[0139] D stored =Store(L(X)) (14)

[0140] Among them, D stored is the stored data, Store is the storage function, and L(X) is the labeled data.

[0141] Use data management tools (such as Apache Hive or AWS Glue) to manage stored data and provide data query and access interfaces. Formula:

[0142] D managed =Manage(D stored ) (15)

[0143] Among them, D managed is the managed data, Manage is the data management function, D stored For stored data.

[0144] Through the above steps, the data is integrated, preprocessed, standardized, dimensionally reduced, classified, labeled, and finally stored and managed. These data will be used for subsequent real-time order anomaly detection, automated processing, intelligent customer service, omni-channel data analysis, and real-time price optimization systems to ensure efficient operation and accurate analysis of the system. The specific solutions for each step are targeted at the special data needs of the e-commerce platform to ensure that the preprocessed data can meet the requirements of intelligent and automated processing, laying the foundation for the realization of the entire patent solution.

[0145] S2. Extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse anomalies of features.

[0146] Specifically, the present invention extracts multiple features from the order data, including but not limited to order amount, order time, quantity of purchased goods, payment method, user's historical purchase behavior, etc. By analyzing these features, abnormal orders can be better identified. Formula:

[0147] X i =[x i,1 ,x i,2 ,…,x i,d ] (16)

[0148] Among them, X i The feature vector representing the i-th order, with dimension d. i,j represents the jth feature of the ith order. d represents the total number of features.

[0149] Specifically, the present invention constructs a multi-stage detection model, which is divided into two stages: primary detection and advanced detection.

[0150] The primary detection stage uses a rule-based method to quickly screen orders and identify obvious abnormal orders. The present invention defines a set of feature thresholds and detects the features of each order.

[0151] Define T j represents the threshold of the j-th feature. Indicates the primary test result of the i-th order, with a value of 0 or 1, 0 indicates normal, and 1 indicates abnormal. Formula:

[0152]

[0153] The advanced detection stage uses a deep learning model based on autoencoders for more sophisticated detection. Autoencoders can capture the nonlinear relationship of order data and are suitable for the detection of complex anomalies.

[0154] Define h i Represents the hidden layer representation of the i-th order, with dimension h. Represents the reconstructed feature vector of the i-th order. Formula:

[0155] h i =σ(W1X i +b1) (18)

[0156]

[0157] Among them, W1 and W2 are the weight matrices of the encoder and decoder respectively, b1 and b2 are the bias vectors respectively, and σ is the activation function.

[0158] Loss function:

[0159]

[0160] Where n is the number of samples, ∥·∥ 2 represents the Euclidean distance.

[0161] Detection and discrimination:

[0162] D i Indicates the advanced detection result of the i-th order, with a value of 0 or 1, 0 indicates normal, and 1 indicates abnormal. Formula:

[0163]

[0164] Among them, ∈ is the preset threshold.

[0165] In order to further improve the accuracy of detection, the present invention adopts a time series anomaly detection algorithm and uses a long short-term memory network (LSTM) to perform time series analysis on order data to identify abnormal patterns within a specific time period.

[0166] Define X t Represents the order feature vector sequence at time t, containing the features of the past k time steps. t Represents the hidden layer state at time t. Formula:

[0167] H t =LSTM(X t ) (twenty two)

[0168]

[0169] Among them, W3 is the output layer weight matrix and b3 is the bias vector.

[0170] Timing detection and judgment:

[0171] Indicates the timing test result at time t, with a value of 0 or 1, 0 for normal and 1 for abnormal. The formula is:

[0172]

[0173] Wherein, δ is a preset timing detection threshold.

[0174] The results of primary detection, advanced detection and time series detection are integrated to form the final anomaly detection result. Indicates the final test result of the i-th order, with a value of 0 or 1, 0 indicates normal, and 1 indicates abnormal. Formula:

[0175]

[0176] Here, ∨ represents a logical OR operation.

[0177] Through the above steps, the real-time order anomaly detection system can efficiently and accurately identify abnormal orders on e-commerce platforms. This solution combines multiple detection methods, using feature extraction, multi-level detection models and time series analysis technology to achieve comprehensive detection of order anomalies, ensure the accuracy and security of order data, and lay a solid foundation for the implementation of the entire patent solution.

[0178] S3. Use RPA technology and smart contracts to automatically process abnormal orders based on the results of anomaly detection.

[0179] Specifically, after detecting abnormal orders, RPA technology and smart contracts are used to automatically process these orders, including refunds, order cancellations, and other operations. This step not only relies on the results of abnormal detection, but also combines intelligent processing procedures to improve processing efficiency and accuracy and reduce manual intervention. Through automated processing, companies can quickly respond to abnormal situations and improve customer satisfaction and trust.

[0180] Furthermore, feature information of abnormal orders is extracted from the anomaly detection system: order amount a i , Order time t i , quantity of goods purchased q i 、Payment method i and user's historical purchase behavior i ,definition:

[0181] O i =[a i ,t i ,q i ,p i ,h i ] (26)

[0182] Furthermore, using the detection results of the previous step, abnormal orders are marked, assuming that the abnormal type has been classified in the previous step:

[0183] F i Indicates the abnormal type flag of the i-th order, with values ​​of 1 (fraudulent order), 2 (duplicate order), etc. Definition:

[0184] F i = Previous classification result (27)

[0185] Define corresponding processing rules according to different exception types. Each exception type corresponds to a set of processing steps, including order cancellation, refund processing, customer notification, etc. Steps:

[0186] Determine the exception type flag F i :Based on the classification result F of the previous stepi Determine the type of exception.

[0187] Define processing rules R k : Processing rules include multiple processing steps, such as order cancellation, refund processing, customer notification, etc. Example:

[0188] For fraudulent orders (F i =1):

[0189] R1=[CancelOrder,ProcessRefund,SendNotification] (28)

[0190] For repeat orders (F i =2):

[0191] R2=[CancelDuplicateOrder,SendNotification] (29)

[0192] Use RPA technology and smart contracts to execute predefined processing rules. Each processing step is completed by RPA robots and smart contracts, including accessing the order system, performing cancellation operations, processing refunds, sending notifications, etc. Steps:

[0193] Initialize the RPA robot: Start the RPA robot and prepare to execute processing tasks;

[0194] Access the order system to obtain order information: The RPA robot logs in to the order management system and obtains abnormal order information. i ;

[0195] Execute each step in the processing rule: Execute each step in sequence and record the result and timestamp of each processing step into the database; i represents the processing flow of the i-th order, Represents a series of processing rules defined based on the abnormal detection results of the order (which will be associated with the definition of processing rules R k ), Example: Cancel order operation:

[0196] CancelOrder( i )→status(O i )=Cancelled (30)

[0197] Processing refund operations:

[0198] ProcessRefund(O i )→refund_status(O i )=Processed (31)

[0199] Inform customer operations:

[0200] SendNotification( i )→notification_status(O i )=Sent (32)

[0201] The processing results of each order are recorded in the system and monitored in real time to ensure the correctness and effectiveness of the processing flow. Steps:

[0202] Record processing results: record the results and timestamp of each processing step in the database, example:

[0203] LogResult(P i )→

[0204] {order_id: i ,status:Cancelled,refund_status:

[0205] Processed,notification_status:Sent,timestamp:t}(33)

[0206] Monitor processing results: Monitor processing results in real time to ensure the correctness and timeliness of the processing process. If an abnormal situation is found (such as processing failure), an alarm will be issued in time and manual intervention will be carried out.

[0207] Through the above steps, the automated abnormal order processing system can efficiently and accurately process abnormal orders on the e-commerce platform. The system combines RPA technology and smart contracts to realize the automation and intelligence of abnormal order processing, ensure the timeliness and accuracy of order processing, and lay a solid foundation for the realization of the entire patent solution.

[0208] S4. Use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match the corresponding abnormal order solutions from the knowledge graph based on the classification results, then generate natural reply content and interact with users.

[0209] Specifically, natural language processing technology is used to identify and classify questions raised by customers related to abnormal orders.

[0210] Furthermore, feature extraction of customer questions: extract key features from the questions asked by customers and use pre-trained language models (such as BERT) to convert customer questions into feature vectors. Definition:

[0211] Q i=BERT(CustomerQuery i ) (34)

[0212] A classification model based on Graph Neural Network (GNN) is used to classify customer questions. A graph structure is constructed, where nodes represent customer questions and edges represent the similarity between questions. The similarity measurement formula is:

[0213]

[0214] Among them, Q i Represents problem i, Q j represents problem j, σ represents activation function, exp represents exponential function, specifically, exponential function with natural logarithm as base;

[0215] Classification formula:

[0216] C i =GNN(Q i ) (36)

[0217] Among them, GNN is a trained graph neural network model, C i Represents question Q i The classification results.

[0218] According to the classification results of customer problems, the corresponding abnormal order solutions are matched from the knowledge graph to build the abnormal order knowledge graph:

[0219] The knowledge graph contains knowledge points and solutions related to abnormal order processing;

[0220] Nodes represent knowledge points and solutions, edges represent the relationship between knowledge points or between knowledge points and solutions. N is defined as the knowledge point set, S as the solution set, and R as the relationship set.

[0221] Categorize the results according to the problem i , find the corresponding solution in the knowledge graph:

[0222] Assume that the problem classification result is C i Mapping to nodes in the knowledge graph Find the solution node corresponding to the shortest path through a graph search algorithm (such as breadth-first search or depth-first search) Specific matching process:

[0223] From the question node in the knowledge graph To begin, search for solution nodes connected to it;

[0224] Determine the best match based on the strength of the relationship between the problem node and the solution node (the weight of the edge);

[0225] Use graph search algorithm to find the optimal solution node, matching formula:

[0226]

[0227] Among them, GraphSearch is a graph search function, Represents the matched solution node.

[0228] Furthermore, natural responses are generated through the dialogue management system to interact with customers.

[0229] Design a dialogue management system to track the state of the dialogue and ensure the coherence of the responses. Define:

[0230] D i =DialogueState(i) (38)

[0231] Use a generative model (such as GPT) to generate natural response content, generating the formula:

[0232]

[0233] Among them, GenerateReply is the function to generate the reply, Response i Represents the generated reply content.

[0234] Continuously improve system performance and service quality through customer feedback. Collect customer feedback on each reply as the basis for system improvement. Customer feedback collection includes customer satisfaction ratings, feedback comments, etc. Definition:

[0235] F i =[SatisfactionScore i ,FeedbackComments i ] (40)

[0236] Among them, SatisfactionScore i is the customer's satisfaction score for the i-th reply, FeedbackComments i It is the customer's feedback review.

[0237] Adjust model parameters based on customer feedback, continuously optimize classification model and generation model, update model loss function including classification loss and generation loss, and optimize based on customer feedback. Update formula:

[0238] L=λ1L classify +λ2L generate +λ3L feedback (41)

[0239] Among them, L classify is the classification loss, L generate is the generation loss, L feedback is the loss based on customer feedback, and λ1,λ2,λ3 are the weights of the loss function.

[0240] Through the above steps, the intelligent customer service system can efficiently and accurately handle customer consultation and after-sales service related to abnormal orders on e-commerce platforms. The system combines natural language processing, knowledge graph and dialogue management technology to achieve automation and intelligence of customer service closely integrated with abnormal order processing, ensuring the timeliness and accuracy of services, and laying a solid foundation for the realization of the entire patent solution.

[0241] S5. Use big data processing technology to collect data from multiple channels and perform preprocessing. Then extract features from the preprocessed data and perform feature fusion to obtain feature data. Perform analysis and modeling based on the fused feature data. Then, perform optimization and decision support based on the analysis and modeling results.

[0242] Specifically, the following steps are included:

[0243] Collect data from multiple channels (such as e-commerce platforms, social media, logistics systems, etc.). The feature vector of each data source is defined as follows: E-commerce platform data: X e =[x e1 ,x e2 ,…,x en ], Social media data: X s =[x s1 ,x s2 ,…,x sn ], Logistics system data: X l =[x l1 ,x l2 ,…,x ln ].

[0244] Furthermore, the collected data is cleaned, normalized and missing values ​​are filled. Normalization formula:

[0245]

[0246] Extract features from data from different channels and perform feature fusion. Steps:

[0247] Use convolutional neural network (CNN) to extract features from image data. The CNN feature extraction formula is:

[0248] F image =CNN(X image ) (43);

[0249] Use natural language processing (NLP) to extract features from text data (such as word embedding). The NLP feature extraction formula is:

[0250] F text =BERT(X text ) (44)

[0251] Furthermore, the features of different channels are fused to obtain a unified feature representation. Feature fusion formula:

[0252] F = Concat(F e ,F s ,F l ) (45)

[0253] Among them, F e 、F s and F l They represent the characteristics of e-commerce platform data, social media data, and logistics system data respectively.

[0254] Furthermore, analysis and modeling are performed based on the fused feature data.

[0255] Specifically, use cluster analysis, association rule mining and other technologies to conduct in-depth analysis of the data. Cluster analysis formula (using K-means algorithm):

[0256]

[0257] Among them, S represents the cluster set, k represents the number of clusters, μ i represents the center of the i-th cluster, and λ is the regularization parameter.

[0258] Use supervised learning methods (such as decision trees, random forests) for predictive modeling. Decision tree formula (with regularization):

[0259] DecisionTree(F)→y (47)

[0260] The objective function of the decision tree is:

[0261]

[0262] Among them, y i is the true value, is the predicted value, α is the regularization parameter, and w j is the leaf node weight of the decision tree.

[0263] Furthermore, optimization and decision support are performed based on the analysis and modeling results.

[0264] Preferably, an optimization algorithm (such as linear programming, genetic algorithm) is used to optimize the decision variables. Linear programming formula (with regularization term):

[0265] minimize c T x+γ||x||1 subject to Ax≤b (49)

[0266] Among them, c is the cost coefficient vector, x is the decision variable vector, γ is the regularization parameter, A is the constraint matrix, and b is the constraint vector.

[0267] Based on the optimization results, decision recommendations are generated and presented to decision makers through visualization tools. Decision recommendation formula:

[0268] Decision = Optimize (F) (50)

[0269] Among them, Optimize represents the optimization function, and F represents the fused features.

[0270] Through the above steps, the omni-channel data analysis and decision support system can efficiently and accurately analyze the multi-channel data on the e-commerce platform and provide strong support for decision-making. The system combines big data processing, machine learning and optimization algorithms to achieve automation and intelligence in data analysis and decision support, ensuring the timeliness and accuracy of decision-making, and laying a solid foundation for the realization of the entire patent solution.

[0271] like Figure 2 As shown, another embodiment of the present invention is provided, an order management system based on big data and artificial intelligence, the system comprising:

[0272] The data collection module 101 is used to obtain order data collected by the e-commerce platform, and perform preprocessing, data fusion, normalization, data dimension reduction and selection on the preprocessed data, and then classify, label and store the selected data;

[0273] The order analysis module 102 is used to extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse anomalies of the features; and automatically process abnormal orders using RPA technology and smart contracts based on the results of anomaly detection;

[0274] The abnormal order processing module 103 is used to use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match corresponding abnormal order solutions from the knowledge graph according to the classification results, and then generate natural reply content and interact with users;

[0275] The system update module 104 is used to collect data from multiple channels using big data processing technology, perform preprocessing, extract features from the preprocessed data, perform feature fusion to obtain feature data, perform analysis and modeling based on the fused feature data, and perform optimization and decision support based on the analysis and modeling results;

[0276] The multi-layer detection model is divided into primary detection, advanced detection and time series detection:

[0277] The primary detection uses a rule-based approach to quickly screen orders and identify obvious abnormal orders, as shown below:

[0278]

[0279] Among them, T j represents the threshold of the jth feature; represents the primary test result of the i-th order, with a value of 0 or 1, 0 represents normal and 1 represents abnormal; x i,j represents the jth feature of the i-th order;

[0280] The advanced detection is performed using a deep learning model based on an autoencoder, as shown below:

[0281] h i =σ(W1X i +b1)

[0282]

[0283] Among them, h i represents the hidden layer representation of the i-th order, with dimension h; represents the reconstructed feature vector of the i-th order; W1 and W2 represent the weight matrices of the encoder and decoder respectively, b1 and b2 represent the bias vectors respectively, and σ is the activation function;

[0284] The loss function L of the deep learning model based on the autoencoder is expressed as follows:

[0285]

[0286] Where n represents the number of samples, ∥·∥ 2 represents the Euclidean distance, i represents the i-th sample; X i The feature vector representing the i-th order, with dimension d;

[0287] The detection discrimination of the advanced detection is expressed as follows:

[0288]

[0289] Among them, Di It represents the advanced detection result of the i-th order, and its value is 0 or 1, 0 represents normal and 1 represents abnormal; ∈ represents the preset threshold;

[0290] The time series detection uses the long short-term memory network LSTM to perform time series analysis on the order data and identify abnormal patterns within a specific time period, as shown below:

[0291] H t =LSTM(X t )

[0292]

[0293] Among them, W3 is the output layer weight matrix; b3 is the bias vector; X t Represents the order feature vector sequence at time t, containing the features of the past k time steps; Indicates that based on H t The predicted order feature vector at time t+1; H t represents the hidden layer state at time t;

[0294] Then perform timing detection and judgment:

[0295]

[0296] Among them, δ is the preset timing detection threshold; Indicates the time series detection result at time t, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal; X t+1 Represents the actual order feature vector at time t+1;

[0297] The results of primary detection, advanced detection, and time series detection are integrated to form the final anomaly detection result, which is shown as follows:

[0298]

[0299] Among them, ∨ represents the logical OR operation, Indicates the final test result of the i-th order, with a value of 0 or 1, 0 indicates normal and 1 indicates abnormal.

[0300] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0301] In addition, for technical details that are not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0302] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0303] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0304] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0305] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An order management method based on big data and artificial intelligence, characterized in that: The method comprises: S1. Obtain order data collected by the e-commerce platform and perform preprocessing, perform data fusion, normalization, data dimension reduction and selection on the preprocessed data, and then classify, label and store the selected data; S2. Extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse anomalies of features; S3. Automatically process abnormal orders using RPA technology and smart contracts based on the results of abnormal detection; S4. Use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match corresponding abnormal order solutions from the knowledge graph based on the classification results, then generate natural reply content and interact with users; S5. Use big data processing technology to collect data from multiple channels and perform preprocessing. Then extract features from the preprocessed data and perform feature fusion to obtain feature data. Perform analysis and modeling based on the fused feature data. Then, perform optimization and decision support based on the analysis and modeling results. The multi-layer detection model is divided into primary detection, advanced detection and time series detection: The primary detection uses a rule-based approach to quickly screen orders and identify obvious abnormal orders, as shown below: ； in, Indicates The threshold of the feature; Indicates The primary test result of an order, with a value of 0 or 1, 0 means normal and 1 means abnormal; Indicates Order No. Features The advanced detection is performed using a deep learning model based on an autoencoder, as shown below: ； ； in, Indicates The hidden layer representation of orders has the dimension ; Indicates The reconstructed feature vector of the orders; and denote the weight matrices of the encoder and decoder respectively, and They represent the bias vectors, is the activation function; The loss function of the autoencoder-based deep learning model It is expressed as follows: ； in, represents the number of samples, represents the Euclidean distance, i represents the i-th sample; Indicates The feature vector of orders, with dimension ; The detection discrimination of the advanced detection is expressed as follows: ； in, Indicates The advanced detection result of an order, with a value of 0 or 1, 0 means normal, and 1 means abnormal; Indicates the preset threshold value; The time series detection uses long short-term memory network Perform time series analysis on order data to identify abnormal patterns within a preset time period, as shown below: ； ； in, is the output layer weight matrix; is the bias vector; Indicates time The order feature vector sequence contains the past Features of time steps; Indicates based on The predicted order feature vector at time t+1; Indicates time The hidden layer state of Then perform timing detection and judgment: ； in, is a preset timing detection threshold; Indicates time The timing test result is 0 or 1, 0 means normal, 1 means abnormal; Represents the actual order feature vector at time t+1; The results of primary detection, advanced detection, and time series detection are integrated to form the final anomaly detection result, which is shown as follows: ； in, Represents a logical OR operation. Indicates The final test result of an order, with a value of 0 or 1, 0 means normal, and 1 means abnormal; Wherein, in S1, order data collected by the e-commerce platform is obtained, and the order data includes online sales data, offline sales data, social media data and customer behavior data; In S3, the abnormal order features are extracted and abnormal orders are marked according to the abnormal order data detected in S2, and then corresponding processing rules are defined according to different abnormal types, and then the predefined processing rules are executed using RPA technology and smart contracts; The use of RPA technology and smart contracts to execute predefined processing rules specifically includes: Start the RPA robot and prepare to execute processing tasks; The RPA robot logs into the order management system and obtains abnormal order information ;in, , Indicates the order amount. Indicates the order time. Indicates the quantity of goods purchased. Indicates the payment method. Indicates the user's historical purchasing behavior; According to the processing rules , execute each step in sequence, and record the results and timestamp of each processing step into the database; represents the processing flow of the i-th order, Represents a series of processing rules defined based on the abnormal detection results of the order; The S4 specifically includes: Extract key features from user questions, use a pre-trained language model to convert user questions into feature vectors, use a classification model based on a graph neural network to classify user questions, and then construct a graph structure, where nodes represent user questions and edges represent the similarity between questions; According to the customer problem classification results, the corresponding abnormal order solutions are matched from the knowledge graph. , searching for corresponding solutions in the knowledge graph; wherein the knowledge graph contains knowledge points and solutions related to abnormal order processing, nodes represent knowledge points and solutions, and edges represent relationships between knowledge points or between knowledge points and solutions; Design a dialogue management model to track the state of the dialogue and use a generative model to generate natural responses. Collect user feedback on each reply as the basis for improving the dialogue management model, adjust the dialogue management model parameters according to user feedback, and optimize the classification and generative models.

2. The order management method based on big data and artificial intelligence according to claim 1 is characterized in that: In S1, the data dimension reduction and selection uses a principal component analysis algorithm to calculate the covariance matrix of the data, perform eigenvalue decomposition, select the first k eigenvectors as principal components, and then select the feature that contributes the most to the target variable.

3. The order management method based on big data and artificial intelligence according to claim 1 is characterized in that: The similarity The calculation is as follows: ; in, Indicates problem i, Represents problem j, represents the activation function, represents the exponential function; Then classification is performed, and the classification formula is expressed as follows: ； in, is a trained graph neural network model, Representation problem The classification results.

4. The order management method based on big data and artificial intelligence according to claim 3 is characterized in that: According to the classification results of customer problems, the corresponding abnormal order solutions are matched from the knowledge graph, wherein the matching process is as follows: From the question node in the knowledge graph To begin, search for solution nodes connected to it; Determine the best match based on the strength of the relationship between the problem node and the solution node; A graph search algorithm is used to find the optimal solution node.

5. The order management method based on big data and artificial intelligence according to claim 4 is characterized in that: In S5, analysis and modeling are performed based on the fused feature data, wherein the analysis includes clustering analysis and association rules, and the modeling uses a supervised learning method to perform predictive modeling.

6. An order management system based on big data and artificial intelligence, characterized by: The system comprises: The data collection module is used to obtain order data collected by the e-commerce platform and perform preprocessing, perform data fusion, normalization, data dimension reduction and selection on the preprocessed data, and then classify, label and store the selected data; The order analysis module is used to extract features from order data and build a multi-layer detection model based on deep learning to detect and fuse features; based on the results of anomaly detection, RPA technology and smart contracts are used to automatically process abnormal orders; The abnormal order processing module is used to use natural language processing technology to identify and classify questions related to abnormal orders raised by users, match the corresponding abnormal order solutions from the knowledge graph based on the classification results, and then generate natural reply content and interact with users; The system update module is used to collect data from multiple channels using big data processing technology, perform preprocessing, extract features from the preprocessed data, fuse the features to obtain feature data, perform analysis and modeling based on the fused feature data, and then perform optimization and decision support based on the analysis and modeling results; The multi-layer detection model is divided into primary detection, advanced detection and time series detection: The primary detection uses a rule-based approach to quickly screen orders and identify obvious abnormal orders, as shown below: ； in, Indicates The threshold of the feature; Indicates The primary test result of an order, with a value of 0 or 1, 0 means normal and 1 means abnormal; Indicates Order No. Features The advanced detection is performed using a deep learning model based on an autoencoder, as shown below: ； ； in, Indicates The hidden layer representation of orders has the dimension ; Indicates The reconstructed feature vector of the orders; and denote the weight matrices of the encoder and decoder respectively, and They represent the bias vectors, is the activation function; The loss function of the autoencoder-based deep learning model It is expressed as follows: ； in, represents the number of samples, represents the Euclidean distance, i represents the i-th sample; Indicates The feature vector of orders, with dimension ; The detection discrimination of the advanced detection is expressed as follows: ； in, Indicates The advanced detection result of an order, with a value of 0 or 1, 0 means normal, and 1 means abnormal; Indicates the preset threshold value; The time series detection uses long short-term memory network Perform time series analysis on order data to identify abnormal patterns within a preset time period, as shown below: ； ； in, is the output layer weight matrix; is the bias vector; Indicates time The order feature vector sequence contains the past Features of time steps; Indicates based on The predicted order feature vector at time t+1; Indicates time The hidden layer state of Then perform timing detection and judgment: ； in, is a preset timing detection threshold; Indicates time The timing test result is 0 or 1, 0 means normal, 1 means abnormal; Represents the actual order feature vector at time t+1; The results of primary detection, advanced detection, and time series detection are integrated to form the final anomaly detection result, which is shown as follows: ； in, Represents a logical OR operation. Indicates The final test result of an order, with a value of 0 or 1, 0 means normal, and 1 means abnormal; Wherein, in the data collection module, order data collected by the e-commerce platform is obtained, and the order data includes online sales data, offline sales data, social media data and customer behavior data; In the order analysis module, abnormal order features are extracted and abnormal orders are marked based on the detected abnormal order data, and then corresponding processing rules are defined according to different abnormal types, and then the predefined processing rules are executed using RPA technology and smart contracts; The use of RPA technology and smart contracts to execute predefined processing rules specifically includes: Start the RPA robot and prepare to execute processing tasks; The RPA robot logs into the order management system and obtains abnormal order information ;in, , Indicates the order amount. Indicates the order time. Indicates the quantity of goods purchased. Indicates the payment method. Indicates the user's historical purchasing behavior; According to the processing rules , execute each step in sequence, and record the results and timestamp of each processing step into the database; represents the processing flow of the i-th order, Represents a series of processing rules defined based on the abnormal detection results of the order; The abnormal order processing module specifically includes: Extract key features from user questions, use a pre-trained language model to convert user questions into feature vectors, use a classification model based on a graph neural network to classify user questions, and then construct a graph structure, where nodes represent user questions and edges represent the similarity between questions; According to the customer problem classification results, the corresponding abnormal order solutions are matched from the knowledge graph. , searching for corresponding solutions in the knowledge graph; wherein the knowledge graph contains knowledge points and solutions related to abnormal order processing, nodes represent knowledge points and solutions, and edges represent relationships between knowledge points or between knowledge points and solutions; Design a dialogue management model to track the state of the dialogue and use a generative model to generate natural responses. Collect user feedback on each reply as the basis for improving the dialogue management model, adjust the dialogue management model parameters according to user feedback, and optimize the classification and generative models.

Images