A multi-channel interactive customer relationship management system

Abstract

The application discloses a kind of multi-channel interactive customer relationship management systems, it is related to customer relationship management technical field, the application includes dynamic routing decision module, multi-modal data fusion module and intelligent feedback optimization module;Dynamic routing decision module assesses channel load by deep neural network, utilizes reinforcement learning dynamically distributes customer request to optimal node, realizes load balancing and service continuity;Multi-modal data fusion module integrates text, speech, image data, constructs space-time correlation graph, identifies cross-channel behavior mode, and ensures data consistency by multidimensional verification;Intelligent feedback optimization module combines customer satisfaction evaluation and multi-modal sentiment analysis, optimizes service strategy using genetic algorithm, and is synchronized to cross-channel knowledge graph to realize self-adapting iteration;The application effectively solves the problem of unreasonable multi-channel load distribution, insufficient data fusion and consistency verification, and inaccurate service strategy optimization, improves customer service quality and experience.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of customer relationship management, in particular to a multi-channel interactive customer relationship management system. BACKGROUND

[0002] With the rapid development of the Internet and mobile technology, the interaction channels between customers and enterprises are increasingly diversified, covering websites, mobile applications, social media, call centers and other forms. This multi-channel interaction mode provides enterprises with more extensive customer touch points, but at the same time brings many challenges. On the one hand, the customer request volume of different channels fluctuates, which easily leads to some channels being overloaded while others are relatively idle, affecting service efficiency and customer experience. On the other hand, the data types generated by each channel are complex and diverse, including text, voice, images, etc., and there may be contradictions or inconsistencies between the data, making it difficult to effectively integrate and analyze, so as to accurately grasp the actual needs and behavior patterns of customers. In addition, traditional customer relationship management systems have limitations in service strategy optimization, and are difficult to adjust service processes in real time according to customer feedback and emotional changes, and cannot meet the growing demand for personalized and high-quality services.

[0003] In order to solve the above problems, there are some related technologies in the market at present. For example, some systems distribute customer requests to different servers through simple load balancing algorithms, but these algorithms are usually fixed and cannot be dynamically adjusted according to real-time channel load and service capacity, and do not consider the priority of customer requests and service continuity across channels. Some systems try to integrate multi-channel data, but mostly only stay at the data storage level, lack of depth analysis and correlation of data, and cannot effectively identify cross-channel behavior patterns, nor can they solve the problem of data contradictions. In terms of service optimization, although there is a simple feedback mechanism based on customer satisfaction evaluation, it does not fully utilize multi-modal data for sentiment analysis, and cannot accurately capture changes in customer emotions, resulting in insufficient precision of service strategy optimization.

[0004] In order to solve the above defects, a technical solution is provided. SUMMARY

[0005] The purpose of the present application is to solve the problems of unreasonable load distribution, insufficient data fusion and consistency verification, and inaccurate service strategy optimization in the multi-channel interaction of existing customer relationship management systems, and to propose a multi-channel interactive customer relationship management system.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A multi-channel interactive customer relationship management system, comprising:

[0008] A dynamic routing decision module collects multi-channel customer requests in real time, evaluates channel load through a deep neural network, and dynamically allocates requests to optimal nodes using reinforcement learning to achieve load balancing and service continuity.

[0009] A multi-modal data fusion module integrates text, voice, and image data, constructs a spatio-temporal correlation graph to identify cross-channel behavior patterns, and triggers multi-dimensional verification to ensure data consistency when there is contradictory information.

[0010] An intelligent feedback optimization module combines customer satisfaction evaluation and multi-modal sentiment analysis, optimizes service strategies using genetic algorithms, and synchronizes them to the cross-channel knowledge graph for adaptive iteration.

[0011] Further, the execution steps of the dynamic routing decision module are as follows:

[0012] Real-time collection of customer interaction requests from various channels such as websites, mobile applications, social media, and call centers, extraction of request features including interaction channel type, request content semantic features, customer identity code, and request initiation timestamp.

[0013] A channel load evaluation model is constructed based on a deep neural network, which inputs the processing queue length, historical average response time, and current number of online customer service personnel of each service node, and outputs real-time service capability scores of each channel.

[0014] A dynamic routing strategy is established using reinforcement learning algorithm, which calculates routing weight coefficients based on the matching degree of customer request features and channel service capability scores. When the request backlog of a certain channel exceeds the threshold, the cross-channel load balancing mechanism is automatically triggered.

[0015] When generating dynamic routing instructions, high-priority customer requests are preferentially allocated to channels with the highest service scores, and cross-channel session state tracking identifiers are established to ensure service continuity.

[0016] Further, the specific operation steps of triggering the cross-channel load balancing mechanism in the dynamic routing decision module are as follows:

[0017] Set a dynamic threshold based on historical data.

[0018] Real-time monitoring of request backlog and service capability scores of each channel.

[0019] When the backlog of a certain channel exceeds the threshold and the service capability score is lower than the average, trigger the load balancing mechanism.

[0020] From the currently available channels, select candidate channels with a service capability score greater than 0.7 and a load queue length less than 30.

[0021] If multiple channels meet the conditions, the channel with the highest matching degree to the customer's request type is preferred;

[0022] The routing weight coefficient generated by the Q-Learning algorithm is used to calculate the comprehensive score of the candidate channel: comprehensive score = service capability score x 0.6 + routing weight coefficient x 0.4;

[0023] The channel with the highest comprehensive score is selected as the target distribution node;

[0024] The non-urgent requests in the backlog channel are proportionally migrated to the target channel, and the migration proportion is dynamically adjusted according to the backlog degree;

[0025] High-priority requests, including VIP customers and complaint requests, are still retained in the original channel, but are distributed to customer service nodes with scores higher than the preset standard;

[0026] A unique session identifier is generated for each cross-channel migrated request, and real-time synchronization of interaction records is performed through Redis;

[0027] After the target channel customer service takes over, the context is quickly obtained through the session ID to avoid repeated description of the problem by the customer;

[0028] The queue length and response time of each channel after migration are monitored in real time, and if the target channel load decreases within 5 minutes and does not meet the expected value, secondary migration is triggered;

[0029] The decision result of this load balancing is fed back to the Q-Learning model as training data to update the routing strategy weight.

[0030] Further, the execution steps of the multi-modal data fusion module are as follows:

[0031] A standardized data cleaning pipeline is established to extract features from text, speech, and image heterogeneous interaction data. Text data uses the BERT model to extract semantic vectors, speech data uses voiceprint recognition and speech-to-text processing, and image data uses convolutional neural networks to extract visual features;

[0032] A spatio-temporal association graph is constructed to associate interaction events generated by the same customer in different channels according to the timeline, and to identify cross-channel behavior patterns. When contradictory information is detected in multi-channel interactions, the data credibility verification submodule is started;

[0033] When implementing real-time data fusion, the attention mechanism is used to dynamically adjust the data weight of each channel. For the case where the same appeal is generated in three consecutive channels, the feature weight is automatically increased by 1.5 times;

[0034] When generating a unified customer portrait, the forgetting factor algorithm is introduced to automatically perform weight decay processing on historical interaction data that exceeds the set time limit.

[0035] Further, the specific operation steps of the dynamic routing decision module in detecting contradictory information in multi-channel interaction and starting the data credibility verification submodule are as follows:

[0036] Triggering condition: Real-time analysis of space-time correlation graph, finding that the behavior logic or data of the same customer in multiple channels is inconsistent;

[0037] Contradiction determination rule: Time stamp conflict, behavior logic conflict and data value contradiction;

[0038] The verification process is as follows:

[0039] Check the legitimacy of the data source: Verify whether the channel is authenticated by the enterprise;

[0040] Anonymity or temporary session channel is marked as low credibility, and further verification is continued;

[0041] Time window analysis: Set a time difference threshold, if the cross-channel event timestamp exceeds the threshold and the content is contradictory, trigger an alarm;

[0042] Behavior logic verification: Call the business process rule engine to check whether the operation conforms to the preset logic;

[0043] Third-party interface call, logistics status: Call the logistics company API to obtain the actual signing time and signing person information of the package;

[0044] Payment voucher: Connect the Alipay or WeChat payment interface to verify whether the order amount and status are consistent;

[0045] Physical evidence comparison including image or video verification: Use the hash algorithm or target detection model to compare whether the client uploaded voucher and system archive are consistent;

[0046] Historical record retrieval: Analyze the client's interaction records in the past 30 days to detect whether there are abnormal patterns;

[0047] Behavior portrait matching: Compare the current behavior with the regular behavior, if the deviation exceeds the threshold, mark it as high risk;

[0048] Data integrity check including key field verification: Check whether the order number, client ID and contact information are complete, if missing, trigger the automatic completion process;

[0049] Data tampering detection: Verify whether the original data has been tampered with by comparing the hash value;

[0050] The verification result processing and decision process is as follows:

[0051] Automatic ruling: If the external data returns a clear result, automatically correct the contradictory data and update the space-time graph;

[0052] Manual intervention: When the external interface is unavailable or the verification result is ambiguous, a manual review work order is generated and assigned to the customer service team to contact the customer for verification.

[0053] Risk Management: After confirming false information, mark the risk level of the customer's account and restrict their operating privileges;

[0054] Update the knowledge graph to record conflicting events and their resolution.

[0055] Furthermore, the specific operation steps of the intelligent feedback optimization module are as follows:

[0056] Satisfaction evaluation data is collected after each customer service loop, and the emotional change curve during the customer interaction process is automatically analyzed through the sentiment analysis model.

[0057] A dual-channel feedback analysis model is constructed to compare and verify the human evaluation data with the machine sentiment analysis results. When the difference between the two exceeds 30%, a human review mechanism is triggered.

[0058] An adaptive strategy optimization engine is used to dynamically adjust the response priority weights of each channel using a genetic algorithm. Parameter combinations that improve performance by less than 5% over three consecutive optimization cycles are automatically eliminated.

[0059] Establish a cross-channel knowledge graph update mechanism to synchronize the optimized service strategy to the decision database of all access channels, ensuring the real-time nature and consistency of strategy updates.

[0060] Furthermore, the specific steps of the intelligent feedback optimization module in automatically analyzing the emotion change curve during customer interaction through the emotion analysis model are as follows:

[0061] Collect customer interaction data from all channels, including text, voice, and images;

[0062] Noise removal, word segmentation, and stop word filtering;

[0063] Convert audio recordings to text using ASR and add timestamps;

[0064] Convert all words to lowercase and handle abbreviations;

[0065] Using a pre-trained VADER model or a fine-tuned BERT model, an emotional polarity score of -1 to 1 is output for each sentence;

[0066] The sentiment scores are arranged in chronological order of interaction time to generate time series data;

[0067] The data was smoothed using interpolation, and a curve of emotional changes was plotted, with key nodes marked.

[0068] Generate an interactive full-process emotion curve, label the events that trigger emotional fluctuations, associate emotional scores with original interaction data, and store them in MongoDB for subsequent analysis.

[0069] Further, by fusing text, voice, image multi-modal data, the accuracy and context understanding ability of sentiment analysis are improved.

[0070] Further, the specific operation steps of the intelligent feedback optimization module through the fusion of text, voice, image multi-modal data are as follows:

[0071] Use the Librosa library to extract the fundamental frequency, speech rate and energy of the voice signal, train a CNN-based voice emotion model, and output anger, calm and anxiety classification labels;

[0072] Use the ResNet model to classify the emoticons or screenshots sent by the customer, extract the text in the picture through Tesseract, and integrate it into the text analysis process;

[0073] Map the text sentiment score, voice emotion label and image classification result into a unified dimension;

[0074] If the voice anger label and the text negative score are consistent, the weight is increased to 1.2 times;

[0075] If the emoticons and the text are contradictory, conflict detection is triggered, and manual review weight is increased by 50%;

[0076] Extract the keywords of each round of dialogue through TF-IDF or BERT, and associate the emotional fluctuations;

[0077] Input the multi-modal emotion sequence, and predict the subsequent emotional trend.

[0078] Compared with the prior art, the beneficial effects of the present application are:

[0079] (1) The present application, through the dynamic routing decision module, can collect multi-channel customer requests in real time and evaluate the load of each channel, dynamically allocate requests to the optimal node using reinforcement learning algorithm, and realize load balancing and service continuity. This mechanism effectively solves the problem of high load in some channels and idle in other channels in traditional systems, significantly improves resource utilization and service efficiency. At the same time, through the cross-channel load balancing mechanism, high-priority customer requests can be processed preferentially, further improving customer experience;

[0080] (2) The multi-modal data fusion module integrates various data types such as text, voice, and images, constructs a spatio-temporal correlation graph to identify cross-channel behavior patterns, and initiates multi-dimensional verification when detecting contradictory information to ensure data consistency. This deep fusion and verification mechanism not only improves the accuracy and reliability of data, but also enables a more comprehensive understanding of customer behavior and needs. In addition, the introduction of the forgetting factor algorithm for weight decay processing of historical interaction data further enhances the timeliness and accuracy of customer portraits, providing strong support for precise service;

[0081] (3) The intelligent feedback optimization module combines customer satisfaction evaluation and multi-modal sentiment analysis, dynamically optimizes service strategies using a genetic algorithm, and synchronizes the optimization results to the cross-channel knowledge graph to achieve adaptive iteration. This module can capture real-time changes in customer emotions and accurately adjust service strategies to ensure the relevance and effectiveness of services. At the same time, through a double-channel feedback analysis model and an automatic elimination mechanism, the accuracy and efficiency of service strategy optimization are further improved, providing customers with more personalized and high-quality service experiences. BRIEF DESCRIPTION OF DRAWINGS

[0082] For the convenience of those skilled in the art, the present application will be further described below with reference to the accompanying drawings;

[0083] Figure 1 The system block diagram of the present application. DETAILED DESCRIPTION

[0084] The technical solutions of the present application will be described below in conjunction with embodiments, obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0085] It should be understood that the terms "include" and "contain" used in the specification and claims of the present disclosure indicate the presence of the described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0086] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used in this disclosure and the claims, "a," "an," and "the" are intended to include both singular and plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this disclosure and the claims, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0087] As shown in FIG. 1, a multi-channel interactive customer relationship management system includes a dynamic routing decision module, a multi-modal data fusion module, and an intelligent feedback optimization module. Figure 1

[0088] The dynamic routing decision module collects multi-channel customer requests in real time, evaluates channel load through a deep neural network, and dynamically allocates requests to optimal nodes using reinforcement learning to achieve load balancing and service continuity.

[0089] Real-time collection of customer interaction requests from various channels such as websites, mobile applications, social media, call centers, etc., extraction of request features including interaction channel type, request content semantic features, customer identity code, request initiation timestamp.

[0090] Based on a deep neural network, a channel load evaluation model is constructed, which inputs the processing queue length, historical average response time, and current number of online customer service personnel of each service node, and outputs the real-time service capability score of each channel.

[0091] A reinforcement learning algorithm is used to establish a dynamic routing strategy, and the routing weight coefficient is calculated according to the matching degree of customer request features and channel service capability score. When the backlog of a certain channel request exceeds the threshold value, the cross-channel load balancing mechanism is automatically triggered. The specific process is as follows:

[0092] According to historical data, set a dynamic threshold, for example, the real-time request queue length of a certain channel > 50, or the average response time > 5 minutes; real-time monitoring of the request backlog (queue length) and service capability score of each channel; when the backlog of a certain channel exceeds the threshold value and the service capability score is lower than the average value of the system (such as score < 0.6), trigger the load balancing mechanism.

[0093] From the currently available channels, select candidate channels with higher service capability scores (such as > 0.7) and lower loads (queue length < 30); if multiple channels meet the conditions, preferentially select the channel with the highest matching degree to the customer request type (for example, voice requests are preferentially allocated to call centers);

[0094] ​The routing weight coefficient generated by the Q-Learning algorithm (step 3 result) is used to calculate the comprehensive score of the candidate channel: comprehensive score = service capability score × 0.6 + routing weight coefficient × 0.4; the channel with the highest comprehensive score is selected as the target distribution node;

[0095] The non-urgent requests (such as ordinary consultations) in the backlog channel are proportionally migrated to the target channel, and the migration proportion is dynamically adjusted according to the backlog level (such as migrating 30% of the requests when the backlog exceeds 20% of the threshold); high-priority requests (such as VIP customers, complaint requests) are still retained in the original channel, but are distributed to dedicated high-score customer service nodes;

[0096] A unique session identifier (UUID) is generated for each cross-channel migrated request, and real-time synchronization of interaction records (such as information provided by the customer, processing progress) is performed through Redis; after the target channel customer service takes over, the context can be quickly obtained through the session ID, avoiding repeated description of the problem by the customer;

[0097] The queue length and response time of each channel after migration are monitored in real time, and if the target channel load does not significantly decrease (such as a decrease of only 10% in queue length) within 5 minutes, a second migration is triggered; the decision results of this load balancing (such as migrated channels, response time changes) are fed back to the Q-Learning model as training data to update the routing strategy weight.

[0098] When generating dynamic routing instructions, high-priority customer requests are preferentially distributed to the channel with the highest service score, and a cross-channel session state tracking identifier is established to ensure service continuity.

[0099] The multi-modal data fusion module integrates text, speech, and image data to construct a spatio-temporal correlation graph to identify cross-channel behavior patterns, and multi-dimensional verification (external interface call, logic check) is triggered for contradictory information to ensure data consistency;

[0100] A standardized data cleaning pipeline is established to extract features from heterogeneous interaction data such as text, speech, and images; BERT model is used to extract semantic vectors from text data, voiceprint recognition and speech-to-text processing are used for speech data, and convolutional neural networks are used to extract visual features from image data;

[0101] A spatio-temporal correlation graph is constructed to associate interaction events generated by the same customer in different channels according to the timeline, and cross-channel behavior patterns are identified; when contradictory information is detected in multi-channel interactions, a data credibility verification submodule is started, specifically:

[0102] Trigger condition: Real-time analysis of spatiotemporal correlation graph, discovering the same customer's behavior logic or data inconsistency in multiple channels (e.g., App shows "order completed", but phone complaint "no receipt"); Contradiction judgment rules: Time stamp conflict (e.g., channel A records time as 14:00, channel B as 14:05 but content conflict); Behavior logic conflict (e.g., no return application submitted but requiring refund); Data value conflict (e.g., payment amount inconsistency); Verification process includes:

[0103] Check the legitimacy of data sources: Verify whether the channel is through enterprise authentication (e.g., whether the social media account is the official binding account of the enterprise); Anonymous or temporary session channel is marked as "low credibility" and needs further verification; Time window analysis: Set a reasonable time difference threshold (e.g., 5 minutes), if the time stamp of cross-channel events exceeds the threshold and the content is contradictory, trigger an alarm; Behavior logic verification: Call the business process rule engine to check whether the operation conforms to the preset logic (e.g., return requires prior application); Third-party interface call, logistics status: Call the logistics company API (e.g., Shunfeng) to get the actual signing time and signing person information of the package; Payment voucher: Connect with Alipay / WeChat payment interface to verify the order amount and status consistency;

[0104] Physical evidence comparison: Image / video verification: Use hash algorithm or target detection model (e.g., YOLO) to compare whether the client uploaded voucher is consistent with the system archive; Historical record retrieval: Analyze the client's interaction records in the past 30 days to detect whether there are abnormal patterns (e.g., frequent contradictory complaints, high-risk labels); Behavior portrait matching: Compare the current behavior with the regular behavior (e.g., active period, commonly used channels), if the deviation exceeds the threshold (e.g., strange channel complaint in the early morning), mark it as "high risk"; Data integrity check, key field verification: Check whether the order number, client ID, contact information, etc. are complete, if missing, trigger the automatic completion process (e.g., request the client to supplement the information by SMS); Data tampering detection: Verify whether the original data has been tampered with by hash value comparison (e.g., SHA-256); Verification result processing and decision process as follows:

[0105] Automatic ruling: If external data (e.g., logistics API) returns a clear result, the system automatically corrects the contradictory data and updates the spatiotemporal graph; Manual intervention: When external interface is not available or verification result is ambiguous, generate manual review work order and assign to customer service team to contact client for verification; Risk disposal: After confirming the false information, mark the client account risk level and limit the operation permission (e.g., prohibit high-frequency work order submission); Update knowledge graph, record contradictory events and processing results for subsequent AI model training.

[0106] The attention mechanism is used to dynamically adjust the weight of each channel data when implementing real-time data fusion. For the case where the same appeal occurs in three consecutive channels, the feature weight is automatically increased by 1.5 times. The forgetting factor algorithm is introduced when generating a unified customer portrait, and the weight of historical interaction data exceeding the set time limit is automatically decayed.

[0107] The intelligent feedback optimization module combines customer satisfaction evaluation and multi-modal sentiment analysis (text, voice, image), optimizes service strategies using genetic algorithms, and synchronizes to the cross-channel knowledge graph for adaptive iteration.

[0108] After each customer service closed loop, satisfaction evaluation data is collected, and the emotional change curve during customer interaction is automatically analyzed by the sentiment analysis model. The specific process is as follows:

[0109] Collect customer interaction data across all channels, including text (online chat, email), voice (call recording to text), and images (such as emoticon screenshots in customer notes). Remove noise (such as ad links and special symbols), tokenize (use Jieba for Chinese), and filter stop words. Convert audio recordings to text using ASR (such as Ali Cloud Speech Recognition) and label timestamps. Convert all text to lowercase, handle abbreviations (e.g., "pls" to "please"), and use pre-trained VADER models (for social media short texts) or BERT fine-tuned models (for vertical fields such as e-commerce customer service) to output sentiment polarity scores (-1 to 1) for each sentence. For example: "The service is terrible!" → -0.8 (negative); "The problem is solved, thank you!" → 0.7 (positive).

[0110] Arrange the sentiment scores in chronological order to generate time series data. Smooth the data using interpolation methods (such as linear interpolation) and plot the emotional change curve, marking key nodes (such as emotional troughs and peaks). Generate an interactive emotional curve, label events that trigger emotional fluctuations (such as long waiting times or problem resolution), and associate sentiment scores with original interaction data for subsequent analysis in MongoDB.

[0111] Further improve the accuracy and contextual understanding of sentiment analysis by fusing text, voice, and image multi-modal data. The process is as follows:

[0112] Use the Librosa library to extract the pitch, speech rate, and energy of the voice signal, train a CNN-based voice emotion model (such as the RAVDESS dataset pre-training), and output classification labels such as anger, calm, and anxiety. Use the ResNet model to classify the emoticons / images sent by the customer, extract the text in the image (such as customer handwritten feedback) using Tesseract, and integrate it into the text analysis process.

[0113] Map text sentiment scores, speech emotion labels, and image classification results to a unified dimension (e.g., -1~1 scale); if the speech anger label and the text negative score are consistent, the weight is increased to 1.2 times; if the emoji and the text are contradictory (e.g., the word "satisfied" but the sending of an unhappy emoji), trigger conflict detection, manual review weight increase by 50%; extract keywords from each round of conversation (e.g., "refund", "logistics") through TF-IDF or BERT, and associate emotional fluctuations; input multi-modal emotion sequences to predict subsequent emotional trends (e.g., the customer may be angry again in 10 minutes).

[0114] Build a dual-channel feedback analysis model to compare and verify the artificial evaluation data and the machine emotion analysis results, and trigger the artificial review mechanism when the difference between the two exceeds 30%;

[0115] Use adaptive strategy optimization engine to dynamically adjust the response priority weight of each channel using genetic algorithm, and automatically eliminate parameter combinations that have improved less than 5% in three consecutive optimization cycles;

[0116] Establish a cross-channel knowledge graph update mechanism to synchronize the optimized service strategy to the decision database of all access channels to ensure the real-time and consistency of the strategy update.

[0117] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and do not limit the application to specific embodiments. Obviously, many modifications and changes can be made according to the content of the specification. The specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A multi-channel interactive customer relationship management system, characterized by, Comprise: A dynamic routing decision module for real-time collection of multi-channel customer requests, assessment of channel load through deep neural networks, and dynamic allocation of requests to optimal nodes using reinforcement learning to achieve load balancing and service continuity; A multi-modal data fusion module for integrating text, voice, and image data, constructing a spatio-temporal correlation graph to identify cross-channel behavior patterns, and triggering multi-dimensional verification to ensure data consistency in the presence of contradictory information; An intelligent feedback optimization module for combining customer satisfaction evaluations and multi-modal sentiment analysis, using genetic algorithms to optimize service strategies, and synchronizing them to the cross-channel knowledge graph for adaptive iteration; The execution steps of the dynamic routing decision module are as follows: Real-time collection of customer interaction requests from various channels such as websites, mobile applications, social media, and call centers, extracting request features including interaction channel type, request content semantic features, customer identity code, and request initiation timestamp; Building a channel load assessment model based on deep neural networks, inputting the current processing queue length, historical average response time, and number of online customer service personnel of each service node, and outputting the real-time service capability score of each channel; Using reinforcement learning algorithms to establish a dynamic routing strategy, calculating routing weight coefficients based on the matching degree of customer request features and channel service capability scores, and automatically triggering a cross-channel load balancing mechanism when the request backlog of a channel exceeds a threshold value; When generating dynamic routing instructions, high-priority customer requests are preferentially allocated to channels with the highest service scores, and cross-channel session state tracking identifiers are established to ensure service continuity; The specific operation steps for triggering the cross-channel load balancing mechanism in the dynamic routing decision module are as follows: Set a dynamic threshold based on historical data; Real-time monitoring of request backlog and service capability scores of each channel; When the backlog of a channel exceeds the threshold and the service capability score is lower than the average, trigger the load balancing mechanism; Select candidate channels with a service capability score greater than 0.7 and a load queue length less than 30 from the current available channels; If multiple channels meet the conditions, preferentially select the channel with the highest matching degree to the customer request type; Use the routing weight coefficients generated by the Q-Learning algorithm to calculate the comprehensive score of the candidate channels: Comprehensive score = service capability score × 0.6 + routing weight coefficient × 0.4; Select the channel with the highest comprehensive score as the target allocation node; Migrate non-urgent requests from the backlog channel to the target channel in proportion, with the migration proportion dynamically adjusted according to the backlog level; High-priority requests, including VIP customers and complaint requests, remain in the original channel but are allocated to customer service nodes with scores higher than the preset standard; Generate a unique session identifier for each cross-channel migrated request and synchronize the interaction records in real time through Redis; After the target channel customer service takes over, quickly obtain the context through the session ID to avoid customer repeated problem description; Real-time monitoring of the queue length and response time of each channel after migration, if the target channel load decreases within 5 minutes and does not meet the expected value, trigger secondary migration; Feed the decision results of this load balancing back to the Q-Learning model as training data to update the routing strategy weight.

2. The multi-channel interactive customer relationship management system of claim 1, wherein, The execution steps of the multi-modal data fusion module are as follows: A standardized data cleaning pipeline is established to extract features from text, speech, and image heterogeneous interactive data. The BERT model is used to extract semantic vectors from text data, voiceprint recognition and speech-to-text processing are used for speech data, and convolutional neural networks are used to extract visual features from image data. A spatiotemporal correlation graph is constructed to correlate the interaction events of the same customer across different channels over time, and cross-channel behavior patterns are identified. When contradictory information is detected in multi-channel interactions, a data credibility verification submodule is started. When implementing real-time data fusion, the attention mechanism is used to dynamically adjust the data weights of each channel. For cases where the same demand is generated in three consecutive channels, the feature weight is automatically increased to 1.5 times. When generating a unified customer portrait, a forgetting factor algorithm is introduced to automatically decay the weights of historical interaction data that exceed the specified time limit.

3. A multi-channel interactive customer relationship management system according to claim 2, wherein, The specific operation steps of the data credibility verification submodule in the dynamic routing decision module are as follows: Triggering condition: Real-time analysis of the spatiotemporal correlation graph reveals inconsistencies in the behavior logic or data of the same customer across multiple channels. Contradiction determination rules: Time stamp conflict, behavior logic conflict, and data value contradiction. The verification process is as follows: Check the legality of the data source: Verify whether the channel is authenticated by the enterprise. Anonymity or temporary session channels are marked as low credibility and further verification is continued. Time window analysis: Set a time difference threshold. If the time stamps of cross-channel events exceed the threshold and the content is contradictory, an alert is triggered. Behavior logic verification: Call the business process rule engine to check whether the operation conforms to the preset logic. Third-party interface call, logistics status: Call the logistics company API to obtain the actual signing time and signing person information of the package. Payment credentials: Integrate with Alipay or WeChat Pay interfaces to verify the order amount and status consistency. Physical evidence comparison including image or video verification: Use hash algorithms or object detection models to compare whether the client uploaded credentials match the system archive. Historical record retrieval: Analyze the customer's interaction records over the past 30 days to detect any abnormal patterns. Behavior portrait matching: Compare the current behavior with the regular behavior. If the deviation exceeds the threshold, mark it as high risk. Data integrity check including key field verification: Check if the order number, customer ID, and contact information are complete. If missing, trigger the automatic completion process. Data tampering detection: Verify whether the original data has been tampered with by comparing hash values. The verification result processing and decision process is as follows: Automatic ruling: If the external data returns a clear result, automatically correct the contradictory data and update the spatiotemporal graph. Manual intervention: When external interfaces are unavailable or verification results are ambiguous, generate a manual review work order and assign it to the customer service team to contact the customer for verification. Risk disposal: After confirming the false information, mark the customer account risk level and restrict the operation permission. Update the knowledge graph and record the contradictory events and processing results.

4. The multi-channel interactive customer relationship management system of claim 1, wherein, The specific operation steps of the intelligent feedback optimization module are as follows: Collect satisfaction evaluation data after each customer service closed loop, and automatically analyze the emotional change curve in the customer interaction process through a sentiment analysis model. A double-channel feedback analysis model is constructed to compare and verify the artificial evaluation data and the machine sentiment analysis results, and when the difference between the two exceeds 30%, an artificial review mechanism is triggered; An adaptive strategy optimization engine is used to dynamically adjust the response priority weight of each channel by using a genetic algorithm, and for parameter combinations that have improved by less than 5% for three consecutive optimization cycles, automatic elimination is implemented; A cross-channel knowledge graph updating mechanism is established to synchronize the optimized service strategy to the decision database of all access channels to ensure the real-time and consistency of the strategy update.

5. A multi-channel interactive customer relationship management system according to claim 4, wherein, The specific operation steps of the intelligent feedback optimization module for automatically analyzing the emotion change curve in the customer interaction process are as follows: Collect customer interaction data across all channels, including text, voice, and images; Remove noise, tokenize, and filter stop words; Convert audio to text using ASR and label timestamps; Convert to lowercase and handle abbreviations; Use a pre-trained VADER model or BERT fine-tuning model to output sentiment polarity scores for each sentence between -1 and 1; Arrange the sentiment scores in chronological order to generate time series data; Smooth the data using interpolation and plot the emotion change curve, marking key nodes; Generate an interactive full-process emotion curve, label events that trigger emotional fluctuations, associate sentiment scores with original interaction data, and store them in MongoDB for subsequent analysis; Fusion of text, voice, and image multi-modal data can improve the accuracy of sentiment analysis and contextual understanding.

6. A multi-channel interactive customer relationship management system according to claim 5, wherein, The specific operation steps of the intelligent feedback optimization module for fusing text, voice, and image multi-modal data are as follows: Use the Librosa library to extract the fundamental frequency, speech rate, and energy of the voice signal, train a CNN-based voice emotion model, and output anger, calm, and anxiety classification labels; Use the ResNet model to classify the emoticons or screenshots sent by the customer, extract the text in the picture using Tesseract, and integrate it into the text analysis process; Map text sentiment scores, voice emotion labels, and image classification results to a unified dimension; If the voice anger label and the text negative score are consistent, the weight is increased to 1.2 times; If the emoticons and the text are contradictory, conflict detection is triggered, and the artificial review weight is increased by 50%; Extract keywords from each round of dialogue using TF-IDF or BERT and associate them with emotional fluctuations; Input the multi-modal emotion sequence to predict the subsequent emotional trend.

Images