Communication equipment service quality dynamic evaluation and optimization system based on customer follow-up visit

Abstract

The application relates to the technical field of data processing, in particular to a communication equipment service quality dynamic evaluation and optimization system based on customer follow-up visits, which comprises an acquisition module, an abnormality judgment module, a cause analysis module, a feature calculation module, a quality evaluation module, an output module and an adjustment module. The application identifies unexpressed attitudes according to the time sequence entropy value of the emotional score sequence; quantifies the deviation degree of explicit expression and implicit emotion through the KL divergence of IVR keys and voice emotion; cross-verification of the follow-up text semantics and the time consumption of the maintenance process is carried out to realize cause tracing; the information content behind the silence is quantified, comprehensive judgment is realized through the calculation of the quality deviation degree; the threshold is dynamically adjusted according to the fluctuation characteristics of the deviation degree to form a closed loop, and the problems of low real experience recognition accuracy and low service quality abnormality cause detection accuracy caused by excessive dependence on single satisfaction score and the inability to recognize complex emotional guidance are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a dynamic evaluation and optimization system for the service quality of communication equipment based on customer feedback. Background Technology

[0002] Customer follow-up after maintenance services is a crucial link in closed-loop service quality management. By conducting automated or manual follow-ups with customers to collect feedback on aspects such as repair response speed, technical capabilities, and service attitude, service shortcomings can be identified in a timely manner, service processes can be optimized, and customer satisfaction can be improved. However, how to accurately extract customers' true experiences from massive amounts of follow-up data and identify deep-seated service quality issues has become a core technical challenge that urgently needs to be solved in the field of telecommunications service management.

[0003] Currently, mainstream customer follow-up systems primarily employ a technical solution combining automated follow-up based on interactive voice responses with natural language processing (NLP) sentiment analysis. Specifically, the system guides customers to press buttons for rating or record voice feedback through pre-set IVR voice prompts, and then uses NLP technology to analyze the sentiment of the customer's transcribed speech, generating a satisfaction score and basic problem classification tags. Some advanced systems also connect to a backend work order system to obtain basic performance indicators such as maintenance service response time and closed-loop time for auxiliary analysis. These technical solutions, to a certain extent, automate the follow-up process and structure basic data, enabling the identification of obvious service failure cases.

[0004] Therefore, existing technologies suffer from the following problems: The evaluation dimensions are too singular and static, relying excessively on customer-expressed satisfaction scores or single-point emotional values, ignoring the dynamic evolution of customer emotions during follow-up visits; they cannot identify contradictory expressions where explicit customer feedback contradicts true emotions—for example, it's common for customers to select "satisfied" in the IVR but express complaints in voice conversations, but existing systems analyze keypad data and voice data separately, failing to detect inconsistencies and obscuring a large number of customers' true attitudes that contradict explicit feedback; they treat silence as ineffective time and ignore it, failing to obtain valuable clues about customers' unexpressed attitudes; subjective feedback and objective process data are isolated, making it impossible to determine whether customer perception aligns with objective facts, leading to vague or even incorrect identification of anomalies. Summary of the Invention

[0005] To address this, the present invention provides a dynamic evaluation and optimization system for communication equipment service quality based on customer feedback. This system overcomes the problems of low accuracy in identifying genuine experiences and low accuracy in detecting causes of service quality anomalies in existing technologies due to over-reliance on a single satisfaction score and the inability to recognize complex emotional orientations, through multi-dimensional data analysis and dynamic adjustment mechanisms.

[0006] To achieve the above objectives, the present invention provides a dynamic evaluation and optimization system for communication equipment service quality based on customer feedback, comprising: The acquisition module is used to acquire customer follow-up text data, emotion score sequence, key press sequence and silent event sequence, and the time consumption of each test link in the maintenance service process. An anomaly detection module is used to determine the occurrence of an emotional abnormality event based on the emotional fluctuation degree and a preset emotional abnormality threshold, wherein the emotional fluctuation degree is determined based on the temporal distribution characteristics of the emotional score sequence; The cause analysis module is used to determine the feedback problem type based on the emotional abnormality event and the follow-up text data, and to determine the process abnormality type based on the time consumption of all the steps, and to calculate the abnormality cause matching degree based on the matching degree between the feedback problem type and the process abnormality type. The feature calculation module is used to calculate the semantic behavior deviation degree based on the emotional abnormality event, the key sequence and the emotional score sequence, and to calculate the silence index based on the silence duration and the silence position. The quality assessment module is used to determine service quality anomalies based on the numerical distribution characteristics of the quality deviation degree, wherein the quality deviation degree is determined based on the anomaly cause matching degree, the semantic behavior deviation degree, and the silence index. The output module is used to generate a comprehensive service quality report based on the service quality anomalies. The adjustment module is used to adjust the preset emotional abnormality threshold according to the fluctuation characteristics of the quality deviation within a preset observation period.

[0007] Furthermore, the anomaly detection module includes: A volatility calculation unit is used to calculate the emotional volatility based on the probability distribution of the emotional score sequence; An anomaly determination unit, which is connected to the volatility calculation unit, is used to determine that an emotional abnormality event has occurred when the emotional volatility is greater than a preset emotional abnormality threshold.

[0008] Furthermore, the volatility calculation unit includes: The state division subunit is used to divide the emotion score sequence according to a preset time length to obtain several emotion windows; The fluctuation calculation subunit, which is connected to the state division subunit, is used to statistically analyze the distribution probability of the emotional score in each emotional window within a preset emotional score range, and calculate the sample entropy based on the distribution probability to obtain the emotional fluctuation degree.

[0009] Furthermore, the cause analysis module includes: The cause analysis unit is used to identify the feedback problem type based on the feedback text data and the preset BERT model, and to determine the process exception type based on the time consumption of all the steps and the preset process example library. A matching degree calculation unit, which is connected to the feedback analysis unit and the process analysis unit respectively, is used to calculate the matching degree of the abnormality cause based on the matching degree of the feedback problem type and the process abnormality type.

[0010] Furthermore, the cause analysis unit includes: The feedback analysis subunit is used to segment and quantize the returned text data to obtain a word vector sequence, generate a semantic feature vector based on the word vector sequence and a preset BERT model, calculate the similarity matching degree based on the semantic feature vector and several standard question vectors in the preset question type feature library, and determine the question label corresponding to the maximum value among all similarity matching degrees as the feedback question type. The process analysis subunit is used to construct a service process sequence based on the time consumption of all the steps, compare the service process sequence with the standard service process sequence in the preset process example library, calculate the time consumption deviation of each step to be tested, and determine the step to be tested as a bottleneck node when the time consumption deviation is greater than the preset time consumption deviation threshold, and obtain the process anomaly type corresponding to the bottleneck node.

[0011] Furthermore, the matching degree calculation unit includes: The vector determination subunit is used to extract keywords of the feedback problem type and perform vectorization processing to obtain a first feature vector, and to extract keywords of the process exception type and perform vectorization processing to obtain a second feature vector. A matching degree calculation subunit, which is connected to the vector determination subunit, is used to calculate the cosine similarity between the first feature vector and the second feature vector to obtain the anomaly cause matching degree.

[0012] Furthermore, the feature calculation module includes: The first feature calculation unit is used to determine the expected emotional score corresponding to each key in the key sequence according to the emotional score table of the preset mapping rule base and the key sequence, count the occurrence frequency of each expected emotional score, generate the expected emotional distribution, normalize each emotional score in the emotional score sequence, generate the actual emotional distribution, and calculate the KL divergence between the expected emotional distribution and the actual emotional distribution to obtain the semantic behavior deviation degree. The second feature calculation unit is used to map the silence duration and the silence location one-to-one to obtain several silence segments, and calculate the silence index based on the distribution density and duration distribution of each silence segment in the follow-up dialogue.

[0013] Furthermore, the second feature calculation unit includes: The density calculation subunit is used to evenly divide the total dialogue duration into a preset number of time windows, count the cumulative duration of the silent segments in each time window, generate a silence density distribution vector, and calculate the density distribution entropy based on the silence density distribution vector. The duration calculation subunit is used to divide the silence duration of the silence segment according to a preset silence level, count the number of silence segments in each silence level, generate a silence duration distribution vector, and calculate the duration distribution entropy based on the silence duration distribution vector. An index subunit, which is connected to the density subunit and the duration subunit respectively, is used to perform a weighted summation of the density distribution entropy and the duration distribution entropy to obtain the silence index.

[0014] Furthermore, the quality assessment module includes: The deviation calculation unit is used to perform a weighted summation of the abnormal cause matching degree, the semantic behavior deviation degree, the silence index, the preset matching degree weight, the preset deviation weight, and the preset silence weight to obtain the quality deviation degree. A quality assessment unit, connected to the deviation calculation unit, is used to determine that a service quality anomaly has occurred when the quality deviation exceeds a preset deviation threshold.

[0015] Furthermore, the adjustment module includes: The variation calculation unit is used to calculate the standard deviation of the quality deviation within the preset observation period to obtain the deviation fluctuation value; An adjustment unit, connected to the change calculation unit, is used to adjust the preset emotional abnormality threshold based on the comparison result between the deviation fluctuation value and the preset deviation fluctuation threshold.

[0016] Compared with existing technologies, the beneficial effects of this invention lie in its construction of a three-dimensional service quality evaluation system through multi-dimensional data collection and fusion analysis. This system encompasses the correspondence between emotional fluctuations, explicit feedback, and implicit emotions, as well as process efficiency and non-verbal signals. Customers' genuine experiences are often reflected in complex signals such as emotional contradictions, differences between explicit and implicit emotions, and hesitation or silence. A single-dimensional satisfaction score cannot fully capture this. This invention uses the temporal entropy value of the emotional score sequence as an indicator of emotional fluctuations to identify unexpressed attitudes. It also quantifies the degree of deviation between explicit and implicit emotions through the KL divergence of IVR button presses and voice emotions. Finally, it cross-validates the semantics of the follow-up text with the maintenance process time. The system calculates the matching degree between problem types and process bottlenecks to achieve root cause tracing; it also transforms the distribution density and duration distribution of silent events into information entropy to quantify the information content behind the silence; finally, the features of the above four dimensions are weighted and fused to form a quality deviation degree, enabling a comprehensive judgment of service quality; it can analyze customers' surface language and behavior through multi-dimensional data to quantify their true experience, and dynamically adjust the emotional abnormality threshold through the fluctuation characteristics of the deviation degree to form an adaptive optimization closed loop, effectively solving the problems of low accuracy in identifying true experience and low accuracy in detecting the causes of service quality abnormalities caused by over-reliance on a single satisfaction score and the inability to identify complex emotional orientations.

[0017] Furthermore, by analyzing the temporal distribution characteristics of customer emotional score sequences to calculate emotional volatility, the degree of emotional fluctuation of customers during the follow-up process is itself an important abnormal signal. When customers experience drastic emotional fluctuations, it often means that there is an unexpressed attitude or conflicting experience. Traditional assessment methods based on average scores or single extreme values ​​smooth out this key information. By converting the distribution probability of the emotional score sequence into sample entropy, the degree of emotional disorder can be quantified. The higher the entropy value, the more unstable the emotion. Exceeding the preset threshold is judged as an emotional abnormal event. It can identify emotional contradictions that cannot be reflected by static indicators from dynamic time series, providing a basis for judgment for subsequent cause tracing and in-depth analysis.

[0018] Furthermore, by dividing the emotional score sequence into time windows and calculating sample entropy, the volatility of emotions is quantified. Customers' emotional expressions are time-dependent; a single emotional value cannot reflect the dynamic evolution of emotions. The fluctuation pattern of emotions over time is key to identifying anomalies. By dividing the sequence into windows, local characteristics of emotions within specific time periods are identified. By mapping emotional scores within a window to a preset score range and statistically distributing probabilities, continuous emotional values ​​are transformed into discrete probability distributions, preserving the intensity of emotions while eliminating interference from small fluctuations. Finally, the degree of disorder in emotional states within each window is calculated using sample entropy. A higher sample entropy value indicates more disordered and unpredictable changes in emotions within the window, which can be used to identify potential differences in customer experiences or missing information. This allows the system to identify emotional anomalies that static indicators cannot reflect from the perspective of temporal evolution, providing precise triggering evidence for subsequent cause analysis.

[0019] Furthermore, the feedback text is transformed into semantic feature vectors using a pre-set BERT model, and matched with a question type library to obtain the feedback question type. Simultaneously, the time consumption data for each step is compared with a standard process model to identify bottleneck nodes where the time deviation exceeds a threshold, thus determining the process anomaly type. Finally, cross-validation is achieved by calculating the matching degree between the two. A high matching degree indicates a high degree of consistency between customer perception and the objective process; a low matching degree suggests the possibility of other undiscovered anomalies. This allows the system to establish a correlation between customer language and objective data, performing correlation analysis from both subjective feedback and objective performance dimensions, significantly improving the accuracy and reliability of identifying service quality anomalies.

[0020] Furthermore, by segmenting and vectorizing the return text and encoding it into semantic feature vectors using the BERT model, and then performing similarity matching with standard vectors in a pre-defined question type library, the customer's language is transformed into quantifiable question labels. Simultaneously, through the process analysis sub-unit, the time consumption of each step is constructed into a service process sequence, compared with the standard process model, and the deviation of the time consumption of each step is calculated to identify bottleneck nodes that exceed the threshold. The process data is transformed into a localizable anomaly type, enabling the system to simultaneously mine the causes of anomalies from two dimensions: what the customer said and what happened in the process. This provides a data foundation for subjective and objective cross-validation and significantly improves the accuracy and reliability of identifying the causes of service quality anomalies.

[0021] Furthermore, by transforming discrete anomaly types into quantifiable vector space models, the correlation between subjective customer problems and objective process anomalies is measured. Keywords for both anomaly types are extracted and vectorized, mapping discrete text tags to the same semantic space. Then, a matching degree calculation subunit calculates the cosine similarity of the two vectors, quantifying the semantic correlation between types into a matching degree value between 0 and 1. A high matching degree indicates a high consistency between customer perception and objective process; a low matching degree suggests potential customer misunderstanding or incomplete identification of process anomalies, requiring further investigation. This allows the system to cross-validate subjective and objective analysis results, significantly improving the credibility of identifying service quality anomaly causes and the reliability of decision-making.

[0022] Furthermore, by comparing the probability distribution of customers' explicit choices with implicit emotional signals and quantifying the information content of silence events, a quantitative analysis of customers' true intentions is achieved. Customers may exhibit discrepancies between explicit and implicit emotional expressions during follow-up visits. IVR button presses represent explicit choices, while voice emotion reflects implicit emotions. The difference between the two is a key indicator for identifying the consistency of customer feedback. By mapping button sequences to expected emotional distributions and normalizing emotional score sequences to actual emotional distributions, and calculating the KL divergence between the two, the degree of difference between explicit choices and implicit emotions can be quantified. A larger KL divergence indicates a greater difference between explicit choices and implicit emotions. Simultaneously, customer silence in dialogue is not a lack of information, but rather an important carrier of information. The distribution density of silence reflects the concentrated areas of information loss, and the duration distribution of silence reflects the intensity distribution of nonverbal signals. By converting the distribution density and duration distribution of silence segments on the time axis into a silence index, the information content hidden behind silence can be quantified. This allows the system to simultaneously identify the difference signals between explicit choices and implicit emotions, as well as the nonverbal signals reflected by silence, providing multi-dimensional quantitative evidence for analyzing customers' surface responses and assessing their true experience.

[0023] Furthermore, by transforming silence events into spatiotemporal dual-dimensional information entropy, a quantitative assessment of customer nonverbal signals is achieved. Silence itself is not a lack of information, but rather a carrier of nonverbal information. The distribution density of silence on the time axis reflects which dialogue nodes the customer lacks information at, while the distribution of silence duration reflects the intensity characteristics of nonverbal signals. By dividing the total dialogue duration into time windows, statistically analyzing the cumulative duration of silence segments within each window, and calculating the density distribution entropy, the distribution characteristics of silence on the time axis can be quantified. The lower the entropy value, the more concentrated the silence is in a specific area. Furthermore, by classifying silence duration into levels and calculating the duration distribution entropy, the diversity of silence duration can be quantified. The higher the entropy value, the more dispersed the silence duration distribution, potentially reflecting complex information patterns. Finally, by weighting and fusing the two, a silence index that comprehensively reflects the information content of silence is obtained. This index can uncover the deeper meaning of nonverbal signals from two dimensions: when silence occurs and how long silence lasts. It transforms previously overlooked silence into a quantifiable assessment indicator, providing a new data dimension for identifying customers' unexpressed attitudes.

[0024] Furthermore, by weighted and fused together the anomaly cause matching degree, semantic behavior deviation degree, and silence index, a multi-dimensional comprehensive quantitative assessment of service quality is achieved. The true state of customer experience cannot be fully depicted by a single dimension. The anomaly cause matching degree reflects the degree of consistency between subjective problems and objective processes; a low matching degree may correspond to customer misunderstandings or incomplete identification of process anomalies. The semantic behavior deviation degree characterizes the degree of deviation between the customer's explicit expression and implicit emotions; a high deviation degree indicates a significant difference between explicit choices and implicit emotions. The silence index quantifies the information content of the customer's nonverbal signals; a high index corresponds to a concentration of missing information. These three dimensions—"degree of cause correlation," "explicit and implicit differences," and "nonverbal signals"—depict different aspects of service quality and are orthogonal and complementary to each other. Multi-dimensional heterogeneous information is fused into a unified quality deviation index. When the deviation degree exceeds a preset threshold, it is judged as a service quality anomaly. This allows the system to comprehensively consider the consistency between subjective and objective factors, explicit and implicit differences, and nonverbal signals, cross-validating the true state of service quality from multiple perspectives, significantly improving the comprehensiveness and accuracy of anomaly identification.

[0025] Furthermore, by monitoring the temporal fluctuation characteristics of service quality deviation and dynamically adjusting the sentiment anomaly threshold, adaptive optimization of the system evaluation criteria is achieved. The volatility of service quality deviation reflects the stability of the current evaluation criteria. When the deviation fluctuates significantly within the preset observation period, it indicates that the system's judgment results on service quality are inconsistent, and there may be improper threshold settings leading to judgment bias. When the fluctuation value exceeds the preset threshold, it indicates that the current sentiment anomaly threshold setting is too high, resulting in the failure to identify abnormal events. The threshold needs to be lowered to enhance the system's ability to identify sentiment anomalies. This allows the system to automatically correct core parameters based on the stability of historical evaluation results, forming a closed-loop mechanism of "evaluation-feedback-optimization." This ensures that the sentiment anomaly judgment criteria always match the current data distribution characteristics, thereby maintaining the long-term stability and accuracy of the system. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the communication equipment service quality dynamic evaluation and optimization system based on customer feedback in this embodiment; Figure 2 This is a logic diagram of the anomaly detection module in this embodiment for determining the occurrence of an emotional abnormality event; Figure 3 This is the logic diagram for determining the feedback problem type by the cause analysis unit in this embodiment; Figure 4 This is the logic diagram for determining the bottleneck node by the cause analysis unit in this embodiment. Detailed Implementation

[0027] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0028] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0029] Please see Figure 1 As shown, this is a schematic diagram of a communication equipment service quality dynamic evaluation and optimization system based on customer feedback, as described in this embodiment. This embodiment provides a communication equipment service quality dynamic evaluation and optimization system based on customer feedback, including: The acquisition module is used to acquire customer follow-up text data, emotion score sequence, key press sequence and silent event sequence, and the time consumption of each test link in the maintenance service process. An anomaly determination module, which is connected to the acquisition module, is used to determine the occurrence of an emotional anomaly event based on the emotional fluctuation degree and a preset emotional anomaly threshold, wherein the emotional fluctuation degree is determined based on the temporal distribution characteristics of the emotional score sequence; The cause analysis module is connected to the anomaly determination module and the acquisition module respectively. It is used to determine the feedback problem type based on the emotional anomaly event and the follow-up text data, and to determine the process anomaly type based on the time consumption of all the steps. It also calculates the anomaly cause matching degree based on the matching degree between the feedback problem type and the process anomaly type. The feature calculation module is connected to the cause analysis module and the acquisition module respectively, and is used to calculate the semantic behavior deviation degree based on the emotional abnormal event, the key sequence and the emotional score, and calculate the silence index based on the silence duration and the silence position. A quality assessment module, which is connected to the cause analysis module and the feature calculation module respectively, is used to determine service quality anomalies based on the numerical distribution characteristics of the quality deviation degree, wherein the quality deviation degree is determined based on the anomaly cause matching degree, the semantic behavior deviation degree, and the silence index. An output module, which is connected to the quality assessment module, is used to generate a comprehensive service quality report based on the service quality anomalies; An adjustment module, which is connected to the output module and the quality assessment module respectively, is used to adjust the preset emotional abnormality threshold according to the fluctuation characteristics of the quality deviation within a preset observation period.

[0030] In this embodiment, during the automated service quality follow-up process for customers (usually operators or enterprise customers) after communication equipment suppliers or operators have completed maintenance services for critical communication equipment such as base stations, when the system initiates the follow-up, it uses technologies such as IVR interaction, speech recognition and emotion computing, and work order system integration to collect in real time the customer's key selections, voice content and emotional fluctuations, silent segments in the dialogue, and the actual time spent in each stage of the maintenance service, such as reporting, dispatching, spare parts issuance, and on-site repair. Based on this multi-dimensional data, the system automatically identifies whether the customer has unexpressed attitude signals such as emotional fluctuations, explicit or implicit differences, or missing information. Simultaneously, it cross-validates the semantic content of the customer's complaints with bottleneck nodes (such as spare parts storage delays) discovered by the backend process mining. Finally, it generates a comprehensive diagnostic report that integrates the customer's real experience and process efficiency issues, and dynamically adjusts the sensitivity threshold for subsequent follow-ups accordingly to achieve accurate identification and closed-loop optimization of customer service quality issues.

[0031] In this embodiment, the follow-up text data is generated by converting the customer's dialogue with customer service into text content in real time using automatic speech recognition technology, preserving the customer's original words for subsequent semantic analysis; the sentiment score sequence is generated by identifying the sentiment tendency of each sentence in the follow-up dialogue using voice sentiment analysis technology, outputting a sentiment score in the range of [-1, 1] (-1 for negative, 0 for neutral, and 1 for positive), and forming a sequence in chronological order; the key sequence is generated by recording the key trajectory of the customer during the automatic voice navigation process through the IVR system log, for example, "1-3-2" represents the customer pressing keys 1, 3, and 2 in sequence; the silence event sequence is generated by extracting silent segments from the dialogue flow using voice activity detection technology, recording the start time, end time, and duration of each silence, forming a set of silence segments; the time consumption of each test link in the maintenance service process is the process timestamp data extracted from the work order system or service management system, including the repair receipt time, order dispatch time, engineer arrival time, repair completion time, spare parts outbound time, customer acceptance time, etc., and the actual time consumption of each link is calculated by the difference between adjacent times.

[0032] In this embodiment, the output module is the system's result presentation and decision support component. When the quality assessment module determines that a service quality anomaly has occurred, the output module is triggered and performs the following processing: First, it obtains the service quality anomaly determination result and quality deviation value from the quality assessment module; it obtains the detected problem type and process anomaly type and their matching degree from the cause analysis module; and it obtains the semantic behavior deviation degree and silence index from the feature calculation module. Second, it integrates and formats these multi-dimensional analysis results to generate a structured comprehensive service quality report. This report includes at least the following: anomaly identifier (marked as "service quality anomaly"), quality deviation score (e.g., 0.85), and an anomaly cause summary (e.g., "Detected problem type: spare parts response delay; process anomaly type..."). The report analyzes various aspects of a customer's behavior, including: warehouse node delays and anomaly cause matching (e.g., semantic behavior deviation of 0.67, indicating explicit / implicit differences; silence index of 0.73, indicating unexpressed attitudes). Finally, the output module pushes the generated report in a readable format to the business system or customer service management platform for the service improvement team to review issues and optimize processes. For example, the report might include specific optimization suggestions: "It is recommended to check the spare parts outbound process at the warehouse node. The current time deviation is 125%. Also, pay attention to the silence signals of these customers during follow-up calls and proactively follow up to obtain further information." In this way, the output module transforms the system's complex analysis results into actionable information that can directly guide business improvements.

[0033] A preset emotional abnormality threshold is a critical value used to determine whether a customer's emotional state has experienced abnormal fluctuations. It depends on the statistical distribution characteristics of emotional fluctuations in historical follow-up data, such as statistical indicators like the median and percentiles calculated based on a large number of historical samples, as well as the business's sensitivity requirements for customers who have not clearly expressed their attitudes, such as the expected balance between recall and precision. At the same time, the system's tolerance for false positives must also be considered. It is usually set between 0.5 and 0.8. In this embodiment, it is set to 0.65, which can identify potential dissatisfied customers with obvious emotional fluctuations while avoiding misjudging customers with normal emotional fluctuations as abnormal, thus balancing the system's sensitivity and accuracy.

[0034] The preset observation period is the length of the time window used by the system to statistically analyze the fluctuation characteristics of service quality deviation. It depends on the granularity of service quality trend monitoring in the business, such as the daily, weekly, or monthly reporting cycle, as well as the system's adaptive adjustment response speed requirements, such as the need to quickly respond to seasonal fluctuations or smooth long-term trends. It is usually set to 7 to 30 days. In this embodiment, it is set to 7 days, which can identify short-term fluctuation trends in service quality in a timely manner and respond quickly to changes in customer experience, while avoiding instability in standard deviation calculation due to insufficient sample size, thereby ensuring the reliability of threshold adjustment.

[0035] Through multi-dimensional data collection and fusion analysis, a comprehensive service quality evaluation system was constructed, encompassing the correspondence between emotional fluctuations, explicit feedback, and implicit emotions, process efficiency, and nonverbal signals. Customers' genuine experiences are often reflected in complex signals such as emotional contradictions, differences between explicit and implicit emotions, and hesitation or silence. A single-dimensional satisfaction score cannot fully capture this. Therefore, the temporal entropy value of the emotional score sequence was used as an indicator of emotional fluctuations to identify unexpressed attitudes. The KL divergence of IVR button presses and voice emotions was used to quantify the degree of deviation between explicit and implicit emotions. Furthermore, the semantics of the follow-up text and the maintenance process time were cross-validated to calculate the relationship between problem types and flow. The system identifies bottlenecks in the process to trace the root causes; it also transforms the distribution density and duration of silent events into information entropy to quantify the information content behind the silence; finally, the features of the above four dimensions are weighted and fused to form a quality deviation degree, enabling a comprehensive judgment of service quality; it can analyze customers' surface language and behavior through multi-dimensional data to quantify their true experience, and dynamically adjust the emotional abnormality threshold through the fluctuation characteristics of the deviation degree to form an adaptive optimization closed loop, effectively solving the problems of low accuracy in identifying true experience and low accuracy in detecting the causes of service quality abnormalities caused by over-reliance on a single satisfaction score and the inability to identify complex emotional orientations.

[0036] Please see Figure 2 As shown, this is a logic diagram for the anomaly determination module in this embodiment to determine the occurrence of an emotional abnormality event. In this embodiment, the anomaly determination module includes: A volatility calculation unit is used to calculate the emotional volatility based on the probability distribution of the emotional score sequence; An anomaly determination unit, which is connected to the volatility calculation unit, is used to determine that an emotional abnormality event has occurred when the emotional volatility is greater than a preset emotional abnormality threshold.

[0037] By analyzing the temporal distribution characteristics of customer emotional score sequences to calculate emotional volatility, the degree of emotional fluctuations during customer follow-up visits is itself an important abnormal signal. When a customer's emotions fluctuate drastically, it often means that there is an unexpressed attitude or conflicting experience. Traditional assessment methods based on average scores or single extreme values ​​smooth out this key information. By converting the distribution probability of the emotional score sequence into sample entropy, the degree of emotional disorder can be quantified. The higher the entropy value, the more unstable the emotion. Exceeding a preset threshold is judged as an emotional abnormal event. It can identify emotional contradictions that cannot be reflected by static indicators from dynamic time series, providing a basis for judgment for subsequent cause tracing and in-depth analysis.

[0038] Specifically, the volatility calculation unit includes: The state division subunit is used to divide the emotion score sequence according to a preset time length to obtain several emotion windows; The fluctuation calculation subunit, which is connected to the state division subunit, is used to statistically analyze the distribution probability of the emotional score in each emotional window within a preset emotional score range, and calculate the sample entropy based on the distribution probability to obtain the emotional fluctuation degree.

[0039] The preset time length is the time span for dividing the emotion score sequence into windows. It depends on the distribution characteristics of the average duration of a single follow-up call, such as the median or average of historical call durations, as well as the business's time resolution requirements for emotion change recognition, such as whether to identify emotion fluctuations at the second level or at the minute level. At the same time, the density of emotion sampling points within the window must also be considered to ensure statistical significance. Usually, the total duration is divided into 10 to 15 windows of equal length. In this embodiment, the total duration is divided into 12 windows of equal length, which can accommodate enough emotion sampling points to ensure statistical significance, and can also identify the phased changes in customer emotions, avoiding the window being too large to smooth out the details of emotion fluctuations, or the window being too small to introduce random noise.

[0040] Preset sentiment score intervals are several ranges of scores that are discretized from continuous sentiment scores. These intervals depend on the precision requirements of the sentiment analysis task, such as whether it's necessary to distinguish between strong and weak negative sentiment, and the granularity of sentiment state differentiation in the business context, such as whether only three categories (positive / neutral / negative) are needed or five categories (strong positive / weak positive / neutral / weak negative / strong negative). Simultaneously, the natural distribution of sentiment scores in historical data must be considered to ensure a relatively balanced sample proportion across intervals and avoid entropy calculation bias. Normalized sentiment scores are typically... The range [-1, 1] is divided into 3 to 5 intervals, such as negative [-1, -0.3), neutral [-0.3, 0.3], and positive (0.3, 1]. In this embodiment, it is set to three intervals: [-1, -0.4] for negative, [-0.4, 0.4] for neutral, and (0.4, 1] for positive. This not only preserves the distinction between the three basic emotional states of positive, neutral, and negative, but also avoids the subsequent entropy value calculation from being too sensitive to small fluctuations due to overly fine intervals, thereby ensuring the stability of emotional fluctuation.

[0041] This system quantifies emotional fluctuations by dividing the emotional score sequence into time windows and calculating sample entropy. Customer emotional expression is time-dependent; a single emotional value cannot reflect the dynamic evolution of emotions. The key to identifying anomalies lies in the fluctuation pattern of emotions over time. By dividing the sequence into windows, local characteristics of emotions within specific time periods are identified. Mapping emotional scores within a window to a preset score range and statistically distributing probabilities transforms continuous emotional values ​​into discrete probability distributions, preserving the intensity of emotions while eliminating interference from small fluctuations. Finally, sample entropy is used to calculate the degree of disorder in emotional states within each window. A higher sample entropy value indicates more disordered and unpredictable emotional changes within the window, which can be used to identify potential differences in customer experiences or missing information. This allows the system to identify emotional anomalies that static indicators cannot reflect from the perspective of temporal evolution, providing precise triggering evidence for subsequent cause analysis.

[0042] Specifically, the cause analysis module includes: The cause analysis unit is used to identify the feedback problem type based on the feedback text data and the preset BERT model, and to determine the process exception type based on the time consumption of all the steps and the preset process example library. A matching degree calculation unit, which is connected to the feedback analysis unit and the process analysis unit respectively, is used to calculate the matching degree of the abnormality cause based on the matching degree of the feedback problem type and the process abnormality type.

[0043] By using a pre-defined BERT model, the feedback text is transformed into a semantic feature vector, which is then matched with a question type library to obtain the feedback question type. Simultaneously, the time consumption data for each step is compared with a standard process model to identify bottleneck nodes where the time deviation exceeds a threshold, thus determining the process anomaly type. Finally, cross-validation is achieved by calculating the matching degree between the two. A high matching degree indicates a high degree of consistency between customer perception and the objective process; a low matching degree suggests the possibility of other undiscovered anomalies. This allows the system to establish a correlation between customer language and objective data, performing correlation analysis from both subjective feedback and objective performance dimensions, significantly improving the accuracy and reliability of identifying service quality anomalies.

[0044] Please see Figure 3 As shown, this is the logic diagram for determining the feedback problem type by the cause analysis unit in this embodiment. Please refer to... Figure 4 As shown, this is the logic diagram for determining the bottleneck node by the cause analysis unit in this embodiment. In this embodiment, the cause analysis unit includes: The feedback analysis subunit is used to segment and quantize the feedback text data to obtain a word vector sequence, generate a semantic feature vector based on the word vector sequence and the preset BERT model, calculate the cosine similarity between the semantic feature vector and several standard question vectors in the preset question type feature library to obtain several similarity matching degrees, and determine the question label corresponding to the maximum value among all similarity matching degrees as the feedback question type. The process analysis subunit is used to construct a service process sequence based on the time consumption of all the steps, compare the service process sequence with the standard service process sequence in the preset process example library, calculate the time consumption deviation of each step to be tested, and determine the step to be tested as a bottleneck node when the time consumption deviation is greater than the preset time consumption deviation threshold, and obtain the process anomaly type corresponding to the bottleneck node.

[0045] The preset BERT model is a pre-trained language model based on the Transformer architecture. Its basic framework adopts the BERT-base-chinese model, which includes a 12-layer Transformer encoder, 12 attention heads, 768-dimensional hidden layers, and approximately 110 million parameters. The initial parameters of the model are general language representation weights obtained by pre-training on a large-scale Chinese corpus. In this embodiment, the model is fine-tuned using historical customer return text data and their corresponding question type labels. The cross-entropy loss function is used during training, the AdamW optimizer is used, and the learning rate is set to 2e. -5The batch size is 16, and the training rounds are 5. After inputting the word vector sequence, the model performs bidirectional context encoding through a multi-layer Transformer encoder to extract the deep semantic representation of each word. Finally, the output vector at the position marked by [CLS] is taken as the semantic feature vector of the entire text. This vector can identify the core semantic information about service issues in the return text and is used for subsequent question type matching.

[0046] The preset problem type feature library is a pre-built vector retrieval library containing several standard problem vectors and their corresponding problem tags. The standard problem vectors are 768-dimensional semantic feature vectors generated by inputting typical problem description texts (such as "slow spare parts delivery," "incompetent engineers," "long response time," etc.) into a fine-tuned BERT model. Problem tags are the business semantic categories corresponding to each standard problem vector. In this embodiment, the preset problem tags include common maintenance service problem types such as "spare parts response delay," "insufficient technical personnel capability," "excessive response time," "poor service attitude," and "incomplete repair." This feature library collects text descriptions of various typical problems, gathering 5-10 samples with different expressions for each problem type. The fine-tuned BERT model generates semantic feature vectors for each sample, and the center point of the vectors of the same type is taken as the standard problem vector for that type of problem, and stored together with the corresponding problem tags in the feature library.

[0047] In this embodiment, the acquired return visit text data is first processed by Chinese word segmentation, dividing the text into word sequences. Each word is mapped to a 768-dimensional word vector by querying a pre-trained word vector table, resulting in a word vector sequence. Next, this word vector sequence is input into a preset BERT model, where a 12-layer Transformer encoder performs contextual semantic encoding, extracting the output vector at the [CLS] marker position as the semantic feature vector of the entire return visit text. Then, the cosine similarity between this semantic feature vector and all standard question vectors in a preset question type feature library is calculated, yielding several similarity matching scores. Finally, the magnitudes of all similarity matching scores are compared, and the question label associated with the standard question vector corresponding to the maximum value is determined as the feedback question type for the current return visit. For example, if the semantic feature vector generated from the return visit text "Your spare parts took three days to arrive" has the highest similarity to the standard question vector of the "Spare parts response delay" category in the feature library, then "Spare parts response delay" is output as the feedback question type.

[0048] The preset process example library is a pre-built service process benchmark library containing several standard service process sequences and their corresponding service type tags. The standard service process sequences are sets of maintenance service steps arranged chronologically based on historical best practices or business specifications. Each step records a standard step name and standard time value. For example, for the "base station equipment repair" service type, the standard service process sequence is: [Repair request reception (standard time 0.5 hours) - Dispatch scheduling (1.0 hour) - Engineer order acceptance (0.5 hours) - Spare parts dispatch (2.0 hours) - Engineer arrival (1.5 hours) - On-site repair (3.0 hours) - Customer acceptance (0.5 hours)], with a total standard duration of 9.0 hours. The library maintains corresponding standard service process sequences for different fault types (such as "RRU fault", "power supply fault", and "transmission fault"). Each sequence is generated through statistical analysis of best practice data from a large number of historical high-efficiency service cases, taking the median or mean time of each step as the standard value, and is calibrated and confirmed by business experts.

[0049] In this embodiment, firstly, based on the actual time consumption data of each test stage during the maintenance service process, a service process sequence is constructed in chronological order, with each stage recording its name and actual time consumption value. Secondly, according to the fault type or service type of the current service, a corresponding standard service process sequence is retrieved from a preset process example library as a comparison benchmark. Then, the actual service process sequence is compared with the standard service process sequence stage by stage, and the time consumption deviation of each test stage is calculated using the formula: Di = |Ti(actual) - Ti(standard)| / Ti(standard), where Di is the time consumption deviation of the i-th stage. The time deviation is calculated as follows: Ti(actual) is the actual time spent in the i-th step, and Ti(standard) is the standard time spent in the corresponding step. Finally, the time deviation of each step is compared with the preset time deviation threshold. If the time deviation of a certain step is greater than the threshold, the step is determined to be a bottleneck node, and the corresponding process exception type is output according to the step type to which the bottleneck node belongs. For example, if the actual time spent in the "spare parts outbound" step is 4.5 hours, the standard time is 2.0 hours, and the time deviation is 125%>30%, then the step is determined to be a bottleneck node, and "spare parts outbound delay" is output as the process exception type.

[0050] The preset time deviation threshold is a critical value used to determine whether the actual time of each step in the service process deviates significantly from the standard time. It depends on the natural fluctuation range of the time of each step in historical service data, such as the standard deviation or interquartile range of the time of each step, as well as the business's tolerance requirements for service efficiency, such as the SLA standards corresponding to different customer levels. At the same time, the importance of the step type must also be considered. For example, core steps (repair completion) and non-core steps (follow-up confirmation) can use different thresholds. It is usually set to 20% to 50% of the standard time, where the standard time refers to the standard time of each step in the standard service process sequence in the preset process example library. In this embodiment, it is set to 30% of the standard time, which can effectively identify bottleneck steps with efficiency problems and avoid misjudging slight delays within the normal fluctuation range as abnormalities, thereby maintaining a reasonable false alarm rate while ensuring the targeted optimization of the process.

[0051] By segmenting and vectorizing the feedback text and encoding it into semantic feature vectors using the BERT model, and then performing similarity matching with standard vectors in a pre-defined question type library, customer language is transformed into quantifiable question labels. Simultaneously, through process analysis sub-units, the time consumption of each step is constructed into a service process sequence, compared with a standard process model, and the deviation of time consumption for each step is calculated to identify bottleneck nodes that exceed thresholds. This transforms process data into localizable anomaly types, enabling the system to simultaneously mine the causes of anomalies from two dimensions: what the customer said and what happened in the process. This provides a data foundation for subjective and objective cross-validation, significantly improving the accuracy and reliability of identifying the causes of service quality anomalies.

[0052] Specifically, the matching degree calculation unit includes: The vector determination subunit is used to extract keywords of the feedback problem type and perform vectorization processing to obtain a first feature vector, and to extract keywords of the process exception type and perform vectorization processing to obtain a second feature vector. A matching degree calculation subunit, which is connected to the vector determination subunit, is used to calculate the cosine similarity between the first feature vector and the second feature vector to obtain the anomaly cause matching degree.

[0053] By transforming discrete anomaly types into a quantifiable vector space model, the correlation between subjective customer problems and objective process anomalies is measured. Keywords for both anomaly types are extracted and vectorized, mapping discrete text tags to the same semantic space. A matching degree calculation subunit then calculates the cosine similarity between the two vectors, quantifying the semantic correlation between types into a matching degree value between 0 and 1. A high matching degree indicates a high consistency between customer perception and objective process; a low matching degree suggests potential customer misunderstanding or incomplete identification of process anomalies, requiring further investigation. This allows the system to cross-validate subjective and objective analysis results, significantly improving the credibility of identifying service quality anomaly causes and the reliability of decision-making.

[0054] Specifically, the feature calculation module includes: The first feature calculation unit is used to determine the expected emotional score corresponding to each key in the key sequence according to the emotional score table of the preset mapping rule base and the key sequence, count the occurrence frequency of each expected emotional score, generate the expected emotional distribution, normalize each emotional score in the emotional score sequence, generate the actual emotional distribution, and calculate the KL divergence between the expected emotional distribution and the actual emotional distribution to obtain the semantic behavior deviation degree. The second feature calculation unit is used to map the silence duration and the silence location one-to-one to obtain several silence segments, and calculate the silence index based on the distribution density and duration distribution of each silence segment in the follow-up dialogue.

[0055] The preset mapping rule library is a pre-built key-emotion mapping relationship repository, containing an emotion score table and optional mapping rule configuration parameters. The emotion score table is the core mapping basis, recording the correspondence between each numeric key in the IVR system and the expected emotion score. The mapping relationship is determined as follows: First, a large amount of historical follow-up data is collected, showing the correspondence between customer key selections and their subsequent voice emotion analysis results. A mapping function between key selections and emotion scores is established using statistical learning or regression analysis methods. Second, the mapping results are calibrated based on business expert experience to ensure they conform to the standard design of customer satisfaction surveys. Based on the above statistical calibration method and the industry-standard five-level satisfaction scale, key 1 (very satisfied) is mapped to an emotion score of 1.0, key 2 (somewhat satisfied) to 0.8, key 3 (neutral) to 0.5, key 4 (dissatisfied) to 0.2, and key 5 (very dissatisfied) to 0.0. This mapping relationship ensures that key selections reasonably reflect the emotional tendencies the customer expects to express. Furthermore, those skilled in the art can use the same statistical calibration method to re-collect data and establish applicable mapping rules based on different customer group characteristics or business scenarios.

[0056] In this embodiment, firstly, each key in the IVR key sequence is traversed, and the expected emotional score corresponding to each key is found according to the emotional score table in the preset mapping rule base. An expected emotional score sequence is generated, and the frequency of occurrence of each expected emotional score is counted. The frequency is then normalized to obtain the expected emotional distribution P. 按键 For example, the key sequence [1, 1, 3, 2] is mapped to obtain the expected sentiment score [1.0, 1.0, 0.5, 0.8], and the distribution P is obtained after counting the frequency. 按键 =[0.5, 0.25, 0.25] (corresponding to scores of 1.0, 0.8, and 0.5); secondly, obtain each emotional score in the emotional score sequence (the original range is [-1, 1]), map it to the interval [0, 1] through linear normalization, generate the actual emotional score sequence, and calculate the distribution probability of each score in the preset emotional interval to obtain the actual emotional distribution P. 情感 Finally, calculate the expected sentiment distribution P. 按键 Compared with the actual emotional distribution P 情感 The KL divergence between them is calculated using the formula D. KL =∑P 按键 (x)log[P 按键 (x) / P 情感 (x)], where D KL This is the semantic behavior deviation degree. The larger the value, the greater the difference between the apparent satisfaction expressed by the customer through keystrokes and the true emotion revealed through voice. In other words, the higher the degree of deviation between explicit expression and implicit emotion.

[0057] By comparing the probability distribution of customers' explicit choices with implicit emotional signals and quantifying the information content of silence events, a quantitative analysis of customers' true intentions is achieved. Customers may express explicit and implicit emotions differently during follow-up interviews. IVR button presses represent explicit choices, while voice emotion reflects implicit emotions. The difference between the two is a key indicator for identifying the consistency of customer feedback. By mapping button sequences to expected emotion distributions and normalizing emotion score sequences to actual emotion distributions, and calculating the KL divergence between the two, the degree of difference between explicit choices and implicit emotions can be quantified. A larger KL divergence indicates a greater difference between explicit choices and implicit emotions. Furthermore, customer silence in conversations is not a lack of information, but rather an important carrier of information. The distribution density of silence reflects the concentrated areas of information loss, and the duration distribution of silence reflects the intensity distribution of nonverbal signals. By converting the distribution density and duration distribution of silence segments on the time axis into a silence index, the information content hidden behind silence can be quantified. This allows the system to simultaneously identify the difference signals between explicit choices and implicit emotions, as well as the nonverbal signals reflected by silence, providing multi-dimensional quantitative evidence for analyzing customers' surface responses and assessing their true experience.

[0058] Specifically, the second feature calculation unit includes: The density calculation subunit is used to evenly divide the total dialogue duration into a preset number of time windows, count the cumulative duration of the silent segments in each time window, generate a silence density distribution vector, and calculate the density distribution entropy based on the silence density distribution vector. The duration calculation subunit is used to divide the silence duration of the silence segment according to a preset silence level, count the number of silence segments in each silence level, generate a silence duration distribution vector, and calculate the duration distribution entropy based on the silence duration distribution vector. An index subunit, which is connected to the density subunit and the duration subunit respectively, is used to perform a weighted summation of the density distribution entropy, the duration distribution entropy, the preset density weight, and the preset duration weight to obtain the silence index.

[0059] The preset number is the number of time windows used to divide the total dialogue duration. It depends on the average duration distribution of a single callback call, such as the median or mode of historical call durations, and the granularity requirements for identifying the details of the silence distribution in the business, such as whether it is necessary to identify the clustering characteristics of silence in different stages before, during, and after the dialogue. At the same time, it is also necessary to consider the statistical stability of the cumulative silence duration within each window to avoid too many windows causing a sparse distribution that affects the entropy calculation. Typically, the number of windows is set to 5 to 20 based on the total duration, so that each window covers a time range of 3 to 10 seconds. In this embodiment, it is set to 10, which can ensure the spatial resolution of the silence distribution and identify the clustering characteristics of silence in different stages of the dialogue, while avoiding too many windows causing the cumulative silence duration within each window to be too sparse and affecting the stability of the entropy calculation.

[0060] The preset silence levels are several level intervals that discretize the duration of silence segments. They depend on existing research in the field of dialogue analysis on the psychological meanings represented by silences of different lengths, such as the behavioral distinction between short pauses (thinking), medium silences (hesitation), and long silences (avoidance), as well as the degree of attention paid to different types of silence in business, such as whether excessively long silences are regarded as a sign of resistance. At the same time, the natural distribution of silence duration in historical data also needs to be considered to make the samples of each level relatively balanced. Usually, the silence duration is divided into 3 to 5 levels. In this embodiment, four levels are set: [0, 2) seconds for short silence, [2, 5) seconds for medium silence, [5, 10) seconds for long silence, and more than 10 seconds for excessively long silence. This can distinguish silence types with different business meanings and make the sample distribution within each level relatively balanced, which can effectively reflect the diversity of silence duration.

[0061] The preset density weight and preset duration weight are the weighting coefficients of the silence density distribution entropy and the silence duration distribution entropy when calculating the silence index. They depend on the correlation strength between each feature and the customer's true satisfaction in historical data. For example, the predictive ability of each feature is determined through regression analysis or feature importance assessment, as well as the importance that the business attaches to the "spatial distribution characteristics" and "duration intensity characteristics" of silence. Usually, the density weight is set between 0.3 and 0.7, and the duration weight is set between 0.7 and 0.3, with the sum of the two being 1. In this embodiment, the density weight is set to 0.4 and the duration weight is set to 0.6, which can take into account both the location distribution information and duration intensity information of silence in the dialogue, and can also dynamically adjust the contribution ratio of the two according to the actual business effect, so that the silence index can more accurately reflect the hidden true attitude of the customer.

[0062] To verify the effectiveness of the silence index, this embodiment conducted statistical analysis based on historical follow-up data. 1000 follow-up recordings from completed customer satisfaction tracking surveys were selected, with 500 recordings confirming the presence of individuals who did not explicitly express their attitudes, and 500 recordings confirming that their attitudes were explicit. The silence index was calculated for both groups. The results showed that the mean silence index for the group with unclear attitudes was 0.73 with a standard deviation of 0.12; the mean silence index for the group with explicit attitudes was 0.31 with a standard deviation of 0.15. The difference between the two groups was statistically significant. Further analysis revealed that in the samples with unclear attitudes, 82% had a silence index greater than 0.65, while this proportion was only 11% in the samples with explicit attitudes. For example, a customer's feedback during a follow-up call was consistently positive, but their silence index was 0.81, with a density distribution entropy of 0.28, indicating a high concentration of silence when asked about the cause of the malfunction. The duration distribution entropy was 0.35, suggesting prolonged and extremely prolonged silences. Based on this, the system determined there was an unexpressed attitude. Subsequent manual review confirmed that the customer had a negative experience due to unclear explanations from the repair personnel, but this was not explicitly stated. Another customer had a silence index of 0.25, a density distribution entropy of 0.72 indicating evenly distributed silence, and a duration distribution entropy of 0.68 indicating predominantly short to medium-length silences. This was confirmed to be a normal pause for thought, not a service quality issue. These verification data demonstrate that the silence index effectively quantifies the information content of customers' nonverbal signals, and its value is significantly correlated with unexpressed attitudes: a silence index greater than 0.65 indicates a high probability of unexpressed attitudes; between 0.45 and 0.65 indicates a need for attention; and less than 0.45 indicates silence falls within the normal range of communication. Those skilled in the art can use the same verification method to calibrate the threshold for determining the silence index and various preset parameters based on the distribution characteristics of their own business data.

[0063] By transforming silence events into spatiotemporal dual-dimensional information entropy, a quantitative assessment of customer nonverbal signals is achieved. Silence itself is not a lack of information, but rather a carrier of nonverbal information. The distribution density of silence on the time axis reflects which dialogue nodes have missing information, while the distribution of silence duration reflects the intensity characteristics of nonverbal signals. By dividing the total dialogue duration into time windows, statistically analyzing the cumulative duration of silence segments within each window, and calculating the density distribution entropy, the distribution characteristics of silence on the time axis can be quantified. The lower the entropy value, the more concentrated the silence is in a specific area. Furthermore, by classifying silence duration into levels and calculating the duration distribution entropy, the diversity of silence duration can be quantified. The higher the entropy value, the more dispersed the silence duration distribution, potentially reflecting complex information patterns. Finally, by weighting and fusing the two, a silence index that comprehensively reflects the information content of silence is obtained. This index can uncover the deeper meaning of nonverbal signals from two dimensions: when silence occurs and how long silence lasts. It transforms previously overlooked silence into a quantifiable assessment indicator, providing a new data dimension for identifying customers' unexpressed attitudes.

[0064] Specifically, the quality assessment module includes: The deviation calculation unit is used to perform a weighted summation of the normalized anomaly cause matching degree, the normalized semantic behavior deviation degree, the normalized silence index, the preset matching degree weight, the preset deviation degree weight, and the preset silence weight to obtain the quality deviation degree. A quality assessment unit, connected to the deviation calculation unit, is used to determine that a service quality anomaly has occurred when the quality deviation exceeds a preset deviation threshold.

[0065] The preset matching degree weight, preset deviation degree weight, and preset silence weight are the weighting coefficients of the anomaly cause matching degree, semantic behavior deviation degree, and silence index when fusion calculation of quality deviation degree, respectively. They depend on the contribution of each dimension feature in historical data to the prediction of service quality anomalies, such as determined by logistic regression coefficients, random forest feature importance, or principal component analysis load, as well as the degree of importance that the business attaches to different signal sources, such as whether to focus more on the direct features of "explicit and implicit differences" or the indirect features of "silence". At the same time, the correlation between features should also be considered to avoid multicollinearity problems. The sum of the three is 1, usually set between 0.2 and 0.5. In this embodiment, the matching degree weight is set to 0.25, the deviation degree weight is set to 0.45, and the silence weight is set to 0.30. This can highlight the dominant role of the deviation between explicit expression and implicit sentiment in the evaluation, make full use of silent signals to mine implicit information, and correct the evaluation results through cause matching degree, forming a multi-dimensional and complementary three-dimensional evaluation system.

[0066] The preset deviation threshold is a critical value used to determine whether service quality has become abnormal. It depends on the statistical distribution characteristics of the quality deviation in historical follow-up data, such as the overlapping area of ​​the deviation distribution between normal and abnormal samples, as well as the sensitivity requirements of the business to service quality issues, such as the differences in service standards corresponding to different customer levels. It also needs to consider the calibration of the labeling results of typical abnormal cases by business experts. It is usually set between 0.5 and 0.8. In this embodiment, it is set to 0.72, which can effectively identify the service quality abnormal events that really need attention, while keeping the false alarm rate within an acceptable range, ensuring that the optimization suggestions triggered by the system have actual business value.

[0067] By weighted and fused together the anomaly cause matching degree, semantic behavior deviation degree, and silence index, a multi-dimensional comprehensive quantitative assessment of service quality is achieved. The true state of customer experience cannot be fully described by a single dimension. The anomaly cause matching degree reflects the degree of consistency between subjective problems and objective processes. A low matching degree may correspond to customer misunderstandings or incomplete identification of process anomalies. The semantic behavior deviation degree characterizes the degree of deviation between the customer's explicit expression and implicit emotions. A high deviation degree indicates a significant difference between explicit choices and implicit emotions. The silence index quantifies the information content of the customer's nonverbal signals. A high index corresponds to a concentration of missing information. These three dimensions—"degree of cause correlation," "explicit and implicit differences," and "nonverbal signals"—describe different aspects of service quality and are orthogonal and complementary to each other. Multi-dimensional heterogeneous information is fused into a unified quality deviation index. When the deviation degree exceeds a preset threshold, it is judged as a service quality anomaly. This allows the system to comprehensively consider the consistency between subjective and objective factors, explicit and implicit differences, and nonverbal signals, and cross-verify the true state of service quality from multiple perspectives, significantly improving the comprehensiveness and accuracy of anomaly identification.

[0068] Specifically, the adjustment module includes: The variation calculation unit is used to calculate the standard deviation of the quality deviation within the preset observation period to obtain the deviation fluctuation value; An adjustment unit, connected to the change calculation unit, is used to reduce the preset emotional abnormality threshold based on the relative deviation between the deviation fluctuation value and the preset deviation fluctuation threshold and a preset adjustment coefficient when the deviation fluctuation value is greater than the preset deviation fluctuation threshold. Here, H'=H×[1-k×(|R-R'| / R')], H' is the adjusted preset emotional abnormality threshold, H is the preset emotional abnormality threshold, k is the preset adjustment coefficient, R is the deviation fluctuation value, and R' is the preset deviation fluctuation threshold.

[0069] The preset deviation fluctuation threshold is used to determine whether the fluctuation of the quality deviation within the current period reaches the critical value that requires adjustment of the sentiment abnormality threshold. It depends on the normal fluctuation range of the quality deviation in historical data, such as the statistical distribution of the standard deviation of each historical period (median, percentile), as well as the business's tolerance for the frequency of system adjustments, such as whether weekly or monthly adjustments are allowed. At the same time, the risk of system oscillation caused by threshold adjustments must also be considered. It is usually set between 0.1 and 0.2. In this embodiment, it is set to 0.15, which can both identify the period when the service quality fluctuates abnormally and trigger the threshold adjustment, and avoid frequent parameter adjustments during the stable operation of the system, thus ensuring the stability of the system.

[0070] The preset adjustment coefficient is a proportional factor used to control the step size of the emotional abnormality threshold adjustment. It depends on the business requirements for balancing the system's adaptive speed and stability. For example, it needs to converge to a new steady state within several observation periods, and it needs to be able to resist interference from single abnormal fluctuations. At the same time, it also needs to consider the simulation test results of the threshold adjustment effect in historical data. It is usually set between 0.1 and 0.3. In this embodiment, it is set to 0.2, which can ensure the system's response speed to changes in service quality, so that the threshold can adapt to the new data distribution in a timely manner, and also control the adjustment range to prevent the threshold from changing drastically due to a single abnormal fluctuation, thus affecting the continuity of subsequent judgments.

[0071] By monitoring the temporal fluctuation characteristics of service quality deviation and dynamically adjusting the sentiment anomaly threshold, adaptive optimization of the system evaluation criteria is achieved. The volatility of service quality deviation reflects the stability of the current evaluation criteria. When the deviation fluctuates significantly within the preset observation period, it indicates that the consistency of the system's judgment results on service quality is low, and there may be improper threshold settings leading to judgment bias. When the fluctuation value exceeds the preset threshold, it indicates that the current sentiment anomaly threshold setting is too high, resulting in the failure to identify abnormal events. The threshold needs to be lowered to enhance the system's ability to identify sentiment anomalies. This allows the system to automatically correct core parameters based on the stability of historical evaluation results, forming a closed-loop mechanism of "evaluation-feedback-optimization." This ensures that the sentiment anomaly judgment criteria always match the current data distribution characteristics, thereby maintaining the long-term stability and accuracy of the system.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic evaluation and optimization system for communication equipment service quality based on customer feedback, characterized in that, include: The acquisition module is used to acquire customer follow-up text data, emotion score sequence, key press sequence and silent event sequence, and the time consumption of each test link in the maintenance service process. An anomaly detection module is used to determine the occurrence of an emotional abnormality event based on the emotional fluctuation degree and a preset emotional abnormality threshold, wherein the emotional fluctuation degree is determined based on the temporal distribution characteristics of the emotional score sequence; The cause analysis module is used to determine the feedback problem type based on the emotional abnormality event and the follow-up text data, and to determine the process abnormality type based on the time consumption of all the steps, and to calculate the abnormality cause matching degree based on the matching degree between the feedback problem type and the process abnormality type. The feature calculation module is used to calculate the semantic behavior deviation degree based on the emotional abnormality event, the key sequence and the emotional score sequence, and to calculate the silence index based on the silence duration and the silence position. The quality assessment module is used to determine service quality anomalies based on the numerical distribution characteristics of the quality deviation degree, wherein the quality deviation degree is determined based on the anomaly cause matching degree, the semantic behavior deviation degree, and the silence index. The output module is used to generate a comprehensive service quality report based on the service quality anomalies. The adjustment module is used to adjust the preset emotional abnormality threshold according to the fluctuation characteristics of the quality deviation within a preset observation period.

2. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 1, characterized in that, The anomaly detection module includes: A volatility calculation unit is used to calculate the emotional volatility based on the probability distribution of the emotional score sequence; An anomaly determination unit, which is connected to the volatility calculation unit, is used to determine that an emotional abnormality event has occurred when the emotional volatility is greater than a preset emotional abnormality threshold.

3. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 2, characterized in that, The volatility calculation unit includes: The state division subunit is used to divide the emotion score sequence according to a preset time length to obtain several emotion windows; The fluctuation calculation subunit, which is connected to the state division subunit, is used to statistically analyze the distribution probability of the emotional score in each emotional window within a preset emotional score range, and calculate the sample entropy based on the distribution probability to obtain the emotional fluctuation degree.

4. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 3, characterized in that, The cause analysis module includes: The cause analysis unit is used to identify the feedback problem type based on the feedback text data and the preset BERT model, and to determine the process exception type based on the time consumption of all the steps and the preset process example library. A matching degree calculation unit, which is connected to the feedback analysis unit and the process analysis unit respectively, is used to calculate the matching degree of the abnormality cause based on the matching degree of the feedback problem type and the process abnormality type.

5. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 4, characterized in that, The cause analysis unit includes: The feedback analysis subunit is used to segment and quantize the returned text data to obtain a word vector sequence, generate a semantic feature vector based on the word vector sequence and a preset BERT model, calculate the similarity matching degree based on the semantic feature vector and several standard question vectors in the preset question type feature library, and determine the question label corresponding to the maximum value among all similarity matching degrees as the feedback question type. The process analysis subunit is used to construct a service process sequence based on the time consumption of all the steps, compare the service process sequence with the standard service process sequence in the preset process example library, calculate the time consumption deviation of each step to be tested, and determine the step to be tested as a bottleneck node when the time consumption deviation is greater than the preset time consumption deviation threshold, and obtain the process anomaly type corresponding to the bottleneck node.

6. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 5, characterized in that, The matching degree calculation unit includes: The vector determination subunit is used to extract keywords of the feedback problem type and perform vectorization processing to obtain a first feature vector, and to extract keywords of the process exception type and perform vectorization processing to obtain a second feature vector. A matching degree calculation subunit, which is connected to the vector determination subunit, is used to calculate the cosine similarity between the first feature vector and the second feature vector to obtain the anomaly cause matching degree.

7. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 6, characterized in that, The feature calculation module includes: The first feature calculation unit is used to determine the expected emotional score corresponding to each key in the key sequence according to the emotional score table of the preset mapping rule base and the key sequence, count the frequency of occurrence of each expected emotional score, generate the expected emotional distribution, normalize each emotional score in the emotional score sequence, generate the actual emotional distribution, and calculate the KL divergence between the expected emotional distribution and the actual emotional distribution to obtain the semantic behavior deviation degree. The second feature calculation unit is used to map the silence duration and the silence location one-to-one to obtain several silence segments, and calculate the silence index based on the distribution density and duration distribution of each silence segment in the follow-up dialogue.

8. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 7, characterized in that, The second feature calculation unit includes: The density calculation subunit is used to evenly divide the total dialogue duration into a preset number of time windows, count the cumulative duration of the silent segments in each time window, generate a silence density distribution vector, and calculate the density distribution entropy based on the silence density distribution vector. The duration calculation subunit is used to divide the silence duration of the silence segment according to a preset silence level, count the number of silence segments in each silence level, generate a silence duration distribution vector, and calculate the duration distribution entropy based on the silence duration distribution vector. An index subunit, which is connected to the density subunit and the duration subunit respectively, is used to perform a weighted summation of the density distribution entropy and the duration distribution entropy to obtain the silence index.

9. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 8, characterized in that, The quality assessment module includes: The deviation calculation unit is used to perform a weighted summation of the abnormal cause matching degree, the semantic behavior deviation degree, the silence index, the preset matching degree weight, the preset deviation weight, and the preset silence weight to obtain the quality deviation degree. A quality assessment unit, connected to the deviation calculation unit, is used to determine that a service quality anomaly has occurred when the quality deviation exceeds a preset deviation threshold.

10. The communication equipment service quality dynamic evaluation and optimization system based on customer feedback as described in claim 9, characterized in that, The adjustment module includes: The variation calculation unit is used to calculate the standard deviation of the quality deviation within the preset observation period to obtain the deviation fluctuation value. An adjustment unit, connected to the change calculation unit, is used to adjust the preset emotional abnormality threshold based on the comparison result between the deviation fluctuation value and the preset deviation fluctuation threshold.

Images