Search

CN-121278315-B - Information technology data monitoring method and system based on deep learning

CN121278315BCN 121278315 BCN121278315 BCN 121278315BCN-121278315-B

Abstract

The invention relates to the technical field of information technology monitoring, in particular to an information technology data monitoring method and system based on deep learning. The method comprises the following steps of S1, obtaining multi-mode signals and metadata, obtaining multi-mode comprehensive emotion scores, decision pressure measurement and confidence coefficient estimation according to the multi-mode signals and the metadata, S2, forming emotion track sequences under each operation step of a user according to the multi-mode comprehensive emotion scores, and S3, monitoring pressure-confidence coefficient paradox, emotion disjointing, cross-cultural baseline deviation and promise-redemption consistency deviation in a sliding window based on the decision pressure measurement, the confidence coefficient estimation and the multi-mode comprehensive emotion scores. The invention brings monitoring accuracy by monitoring pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and promise-redemption consistency deviation, and realizes anonymous or real-name layering strategy to achieve both sensitivity and false alarm control.

Inventors

  • ZHAO YU
  • REN JIALE
  • HU ZHITAO
  • ZHAO FEI
  • SU YINGYING
  • ZHAO XIAOGUANG

Assignees

  • 葫芦岛盘古信息技术有限公司

Dates

Publication Date
20260512
Application Date
20251208

Claims (6)

  1. 1. The information technology data monitoring method based on deep learning is characterized by comprising the following steps of: The method comprises the steps of S1, acquiring multi-modal signals and metadata, acquiring multi-modal comprehensive emotion scores, decision pressure measurement and confidence level estimation according to the multi-modal signals and the metadata, acquiring the multi-modal signals and the metadata of a user, wherein the multi-modal signals comprise text signals, voice signals and behavior signals, and acquiring the metadata of a session, and the metadata comprises a time stamp, a session identifier, a user authentication category, a region identifier, a language identifier and a function module identifier; S2, forming an emotion track sequence under each operation step of the user according to the multi-mode comprehensive emotion scores; S3, monitoring pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and commitment-redemption consistency deviation based on the decision pressure measurement, confidence estimation and multi-modal comprehensive emotion score in a sliding window; The decision pressure measurement is near real-time behavior/physiological agent quantity of subjective pressure/cognitive load of the user, and the pressure rise reflects tension in the user or hesitation triggered by external information; The pressure-confidence paradox is judged by a relative pressure rise rate and a relative confidence fall rate, and when the relative pressure rise rate is greater than a pressure threshold and the relative confidence fall rate is less than a confidence threshold, the pressure-confidence paradox red alarm is triggered, wherein the relative pressure rise rate and the relative confidence fall rate are obtained by the following formula: ; ; In the formula, Is the relative pressure rise rate; For users In sliding window Upper decision pressure metric Is a running average of (2); The window is the current judgment window; is a baseline window; Is a numerical stability constant for avoiding division by zero; For users At the baseline window Upper decision pressure metric Is a running average of (2); is the relative confidence degradation rate; For users In sliding window Upper confidence estimation Is a running average of (2); For users At the baseline window Upper confidence estimation Is a running average of (2); calculating pearson correlation coefficients of the multi-modal comprehensive emotion scores and expected emotion scores in the sliding window according to the emotion track sequence, and judging that emotion is disjoint when the pearson correlation coefficients are lower than a preset emotion threshold value; The monitoring of pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and commitment-redemption consistency deviation based on the decision pressure metric, confidence estimate and multimodal synthetic emotion score within a sliding window comprises a region-or language-based cultural baseline Calculating the mean value And standard deviation of And based on normalized residual errors Determining cross-cultural baseline deviations, where For users At the point of time The multi-modal comprehensive emotion score is recorded as a cultural discomfort indication when the standardized residual exceeds a preset baseline threshold value, and a localization suggestion is prompted in closed-loop output; Defining a commitment point Redemption points Calculating an emotion slope of a multi-modal composite emotion score, determining a commitment-redemption consistency deviation when the emotion slope is below a preset emotion threshold, and triggering a commitment undersension diagnosis in closed-loop output, wherein the commitment point The moment at which a key decision is made for the user; redemption points The time when the functional value is really delivered to the system or is felt by the user for the first time; And S4, carrying out causal association and correction adjustment according to the type of anomaly monitoring, wherein causal association is carried out on the pressure-confidence paradox, emotion dislocation, cross-cultural baseline deviation and promise-redemption consistency deviation detected by anomaly monitoring and specific interface elements, operation steps or content in the system accurately, and diagnosis hole, risk alarm level, recommended intervention measure and A/B test suggestion are generated.
  2. 2. The method for monitoring information technology data based on deep learning according to claim 1 is characterized in that the method for obtaining multi-modal comprehensive emotion scores according to the multi-modal signals and metadata comprises the steps of adopting a context encoder based on a transducer to output text emotion scores for the text signals, adopting a Mel spectrum feature combined with a sequence encoder to output voice emotion scores for the voice signals, adopting a time sequence encoder to extract behavior features of micro hesitation, repeated check, mouse tremors and stay time for the behavior signals, obtaining behavior emotion scores according to the behavior features, and obtaining multi-modal comprehensive emotion scores according to the text emotion scores, the voice emotion scores and the behavior emotion scores.
  3. 3. The method of deep learning based data monitoring of claim 1 wherein triggering the pressure-confidence paradox red alert when the relative pressure rise rate is greater than a pressure threshold and the relative confidence fall rate is less than a confidence threshold comprises selecting different pressure thresholds and confidence thresholds in the decision based on a user authentication category, wherein the user authentication category comprises real name, pseudonym and anonymous category.
  4. 4. The method for deep learning based information technology data monitoring of claim 1, wherein the monitoring of pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation, and commitment-redemption consistency deviation within a sliding window based on the decision pressure metric, confidence estimate, and multimodal synthetic emotion score comprises wherein the obtaining of pearson correlation coefficients is described as: ; In the formula, For users In sliding window The pearson correlation coefficient of the upper; For users In sliding window A sliding average of intra-multimodal synthetic emotion scores; to be in sliding window A sliding average of intra-expected composite emotion scores; To calculate the pearson correlation coefficient.
  5. 5. The method for deep learning based information technology data monitoring of claim 1, wherein the monitoring of pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and commitment-redemption consistency deviation within a sliding window based on the decision pressure metric, confidence estimate and multimodal synthetic emotion score comprises the steps of: ; In the formula, Is the emotion slope; To be at the point of redemption Time user Multi-modal comprehensive emotion scores; To be at the promise point Time user Multi-modal comprehensive emotion scores of (1).
  6. 6. An information technology data monitoring system based on deep learning, for implementing the information technology data monitoring method based on deep learning as set forth in any one of claims 1 to 5, comprising: The data acquisition module is used for acquiring multi-modal signals and metadata, and acquiring multi-modal comprehensive emotion scores, decision pressure measurement and confidence level estimation according to the multi-modal signals and the metadata; The track construction module is used for constructing an emotion track sequence under each operation step of the user according to the multi-mode comprehensive emotion scores; The anomaly monitoring module is used for monitoring pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and promise-redemption consistency deviation based on the decision pressure measurement, confidence estimation and multi-modal comprehensive emotion score in a sliding window; the pressure monitoring submodule is used for judging a pressure-confidence paradox state through the relative pressure rising rate and the relative confidence coefficient falling rate; the emotion monitoring submodule is used for calculating the pearson correlation coefficient of the multi-mode comprehensive emotion score and the expected emotion score in the sliding window according to the emotion track sequence and judging an emotion disjoint state; The deviation monitoring sub-module is used for calculating emotion slope according to the promised point and the redemption point and judging the promised-redemption consistency deviation state; The culture monitoring submodule is used for calculating a standardized residual error and judging a cross-culture baseline deviation state; And the monitoring adjustment module is used for carrying out causal association and correction adjustment according to the abnormal monitoring type.

Description

Information technology data monitoring method and system based on deep learning Technical Field The invention relates to the technical field of information technology monitoring, in particular to an information technology data monitoring method and system based on deep learning. Background With the development of internet products and services to multi-terminals, immersions and globalization, user interaction scenes have the characteristics of coexistence of multi-modes, multi-culture and high-risk decisions. In such a scenario, the conventional monitoring and analysis method is mainly based on a single modality, such as only text analysis or only behavior log, or only statistics of surface indexes, such as click rate, retention rate and conversion rate, so that the single modality analysis cannot accurately characterize the real emotion and cognitive state of the user, lacks dynamic expectation and cultural difference modeling, easily causes misjudgment or localization failure, has insufficient feature recognition and causal association capability on a high-risk decision scene, lacks an intervention path of closed-loop verification and insufficient differentiation processing of anonymous and real-name users, and in view of the problems, the industry is in need of a multi-modality emotion perception mechanism capable of fusing text, voice, behavior and optional physiological signals, and meanwhile, can dynamically manage expected emotion baselines according to the context of the user. Disclosure of Invention In order to overcome the defect of lack of intelligent monitoring of user experience data, the invention provides an information technology data monitoring method and system based on deep learning. The technical implementation scheme of the invention is that the information technology data monitoring method based on deep learning comprises the following steps: s1, acquiring multi-mode signals and metadata, and acquiring multi-mode comprehensive emotion scores, decision pressure measurement and confidence estimation according to the multi-mode signals and the metadata; S2, forming an emotion track sequence under each operation step of the user according to the multi-mode comprehensive emotion scores; S3, monitoring pressure-confidence paradox, emotion disjoint, cross-cultural baseline deviation and commitment-redemption consistency deviation based on the decision pressure measurement, confidence estimation and multi-modal comprehensive emotion score in a sliding window; And S4, carrying out causal association and correction adjustment according to the abnormal monitoring type. Preferably, the acquiring the multi-modal signals and metadata, and acquiring multi-modal comprehensive emotion scores, decision pressure metrics and confidence level estimates according to the multi-modal signals and metadata comprises acquiring the multi-modal signals and metadata of a user, wherein the multi-modal signals comprise text signals, voice signals and behavior signals, and acquiring metadata of a session, the metadata comprises a time stamp, a session identifier, a user authentication category, a region identifier, a language identifier and a function module identifier, and acquiring the decision pressure metrics and the confidence level estimates according to the multi-modal signals and the metadata. Preferably, the multi-modal comprehensive emotion score is obtained according to the multi-modal signal and metadata, and comprises the steps of adopting a context encoder based on a transducer to output text emotion score for the text signal, adopting a Mel spectrum feature and a sequence encoder to output voice emotion score for the voice signal, adopting a time sequence encoder to extract behavior features of micro hesitation, repeated check, mouse tremor and residence time for the behavior signal, obtaining behavior emotion score according to the behavior features, and obtaining multi-modal comprehensive emotion score according to the text emotion score, the voice emotion score and the behavior emotion score. Preferably, the monitoring of pressure-confidence paradox, emotional dislocation, cross-cultural baseline deviation and promise-redemption consistency deviation based on the decision pressure measurement, confidence estimation and multi-modal comprehensive emotion score in a sliding window comprises judging the pressure-confidence paradox through a relative pressure rising rate and a relative confidence falling rate, and triggering the pressure-confidence paradox red alarm when the relative pressure rising rate is larger than a pressure threshold and the relative confidence falling rate is smaller than a confidence threshold, wherein the relative pressure rising rate and the relative confidence falling rate are obtained through the following formula: ; ; In the formula, Is the relative pressure rise rate; For users In sliding windowUpper decision pressure metricIs a running average of (2); The window is the curr