Intelligent customer service self-learning method and system

Abstract

The invention relates to an intelligent customer service self-learning method and system, and the method comprises the steps: collecting the interaction data of a user and an intelligent customer service, and forming a multi-dimensional data set; based on a PID (Proportion Integration Differentiation) controller, processing the performance indexes in the multi-dimensional data set, calculating a current error signal, and generating a corresponding control instruction according to the current error signal so as to adjust the response behavior of the intelligent customer service system in real time; evaluating the system performance data adjusted by the control instruction to obtain evaluation feedback; and according to the evaluation feedback, dynamically adjusting the parameters of the PID controller through a self-adaptive control strategy, and feeding back the adjusted parameters to the PID controller in the step S2. According to the invention, by introducing a closed-loop feedback mechanism based on the PID controller and a parameter adaptive optimization strategy, real-time regulation and control of the response behavior of the intelligent customer service system are realized, and the stability, the response speed and the user satisfaction of the system are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent customer service, more particularly to an intelligent customer service self-learning method and system. BACKGROUND

[0002] Although the existing methods such as supervised learning and reinforcement learning improve the performance of intelligent customer service to some extent, there are still some problems. For example, supervised learning relies on a large amount of labeled data, and the cost of data collection and labeling is high; reinforcement learning requires a large number of trial and error, and the training process is time-consuming and difficult to guarantee convergence to the optimal strategy; transfer learning has challenges in domain adaptability, and knowledge transfer between different domains may have semantic bias; online learning has limited processing capacity for real-time data, and it is difficult to adapt to changes in user demand. Therefore, there is an urgent need for an intelligent customer service self-learning solution that can continuously optimize. SUMMARY

[0003] The technical problem to be solved by the present application is to provide an intelligent customer service self-learning method and system to overcome the above-mentioned defects of the prior art.

[0004] The technical solution adopted by the present application to solve the technical problem is to construct an intelligent customer service self-learning method, which comprises the following steps:

[0005] Step S1, collecting interaction data between users and intelligent customer service to form a multi-dimensional data set;

[0006] Step S2, based on a PID controller, processing performance indicators in the multi-dimensional data set, calculating a current error signal, and generating a corresponding control instruction according to the current error signal to adjust the response behavior of the intelligent customer service system in real time;

[0007] Step S3, evaluating the system performance data adjusted by the control instruction to obtain evaluation feedback;

[0008] Step S4, dynamically adjusting the parameters of the PID controller through an adaptive control strategy according to the evaluation feedback, and updating the parameters of the PID controller.

[0009] Further, the interaction data includes interaction process data, performance indicator data, and environment and state data, and the step S1 comprises:

[0010] Obtaining the interaction process data, the performance indicator data, and the environment and state data, and performing cleaning and structured processing to form a standardized multi-dimensional data set.

[0011] Further, the performance indicators include: an answer accuracy rate, a response time, and a user satisfaction degree; in the step S2, the PID controller processes the performance indicators in the multi-dimensional data set, and calculating the current error signal includes:

[0012] obtaining a preset target accuracy rate, a preset target response time, and a preset target user satisfaction degree;

[0013] calculating an accuracy rate deviation according to the answer accuracy rate and the preset target accuracy rate;

[0014] calculating a response time deviation according to the response time and the preset target response time;

[0015] calculating a user satisfaction degree deviation according to the user satisfaction degree and the preset target user satisfaction degree;

[0016] weighting and fusing the deviation values to generate a comprehensive current error signal as an input of the PID controller.

[0017] Further, in the step S2, the generating a corresponding control instruction according to the current error signal includes:

[0018] inputting the current error signal into the PID controller to respectively perform proportional operation, integral operation, and differential operation to obtain a proportional adjustment component, an integral correction component, and a differential prediction component;

[0019] superimposing and fusing the proportional adjustment component, the integral correction component, and the differential prediction component to generate a total control signal;

[0020] generating a system control instruction according to the total control signal;

[0021] issuing the system control instruction to an execution unit of the intelligent customer service system.

[0022] Further, the step S3 includes:

[0023] based on Lyapunov stability theory, constructing a system energy function according to the system performance data, and calculating a rate of change of the energy function over time in real time;

[0024] judging system stability according to the rate of change to generate a stability analysis result;

[0025] calculating a system performance indicator based on the system performance data;

[0026] integrating the system performance indicator and the stability analysis result to generate the evaluation feedback report.

[0027] Further, in the step S4, the dynamic parameter adjustment by the adaptive control strategy comprises:

[0028] According to the stability analysis result in the evaluation feedback report, the proportional coefficient, the integral coefficient or the differential coefficient of the PID controller is adjusted, and the adjusted parameter is fed back to the PID controller in the step S2, so that the system energy function decreases over time.

[0029] Further, the dynamic parameter adjustment by the adaptive control strategy further comprises:

[0030] The dynamic model parameters of the intelligent customer service system are estimated in real time and online by a self-correcting control strategy;

[0031] The dynamic model parameters are used to represent the change of the mapping relationship between the system input and the control output, the system input includes the knowledge base matching weight or the model generation parameter, and the control output includes the response accuracy or the response speed.

[0032] Further, the dynamic parameter adjustment by the adaptive control strategy further comprises:

[0033] Based on the change of the estimated dynamic model parameters, the drift degree of the system characteristics is judged;

[0034] According to the drift degree, the gain parameter of the PID controller is dynamically adjusted to make the performance index of the intelligent customer service system track the preset optimal performance target;

[0035] The gain parameter includes one or more combinations of the proportional coefficient, the integral coefficient and the differential coefficient.

[0036] Further, the dynamic parameter adjustment by the adaptive control strategy further comprises:

[0037] A model predictive control strategy is adopted to predict the system performance change trend in a future period of time based on the current system state and the dynamic model parameters;

[0038] The control sequence in the future limited time domain is rollingly optimized with the long-term performance index as the optimal target, and the optimal control action at the current time is executed;

[0039] The long-term performance index includes the user satisfaction cumulative value or the monthly problem solving rate.

[0040] The application also provides an intelligent customer service self-learning system, which comprises a processor and a memory storing a computer program, and the processor realizes the steps of the method according to any one of the above embodiments when executing the computer program.

[0041] The beneficial effects of this invention are as follows: the steps of this invention include: collecting user interaction data with intelligent customer service to form a multi-dimensional dataset; processing the performance indicators in the multi-dimensional dataset based on a PID controller, calculating the current error signal, and generating corresponding control commands based on the current error signal to adjust the response behavior of the intelligent customer service system in real time; evaluating the system performance data after adjustment by the control commands to obtain evaluation feedback; and dynamically adjusting the parameters of the PID controller through an adaptive control strategy based on the evaluation feedback, and feeding the adjusted parameters back to the PID controller in step S2. This application, by introducing a closed-loop feedback mechanism based on a PID controller and a parameter adaptive optimization strategy, achieves real-time control of the response behavior of the intelligent customer service system, significantly improving system stability, response speed, and user satisfaction. Attached Figure Description

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0043] Figure 1 This is a schematic diagram of the self-learning method for inventing intelligent customer service. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to the present invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0045] like Figure 1 As shown, Figure 1 This is a schematic diagram of the intelligent customer service self-learning method process.

[0046] The technical solution adopted by this invention to solve its technical problem is: to provide an intelligent customer service self-learning method, the method comprising the following steps:

[0047] Step S1: Collect user interaction data with the intelligent customer service to form a multi-dimensional dataset;

[0048] Step S2, based on the PID controller, the performance indicators in the multi-dimensional data set are processed, the current error signal is calculated, and the corresponding control instruction is generated according to the current error signal, so as to adjust the response behavior of the intelligent customer service system in real time.

[0049] In step S3, the system performance data adjusted by the control instruction is evaluated to obtain evaluation feedback.

[0050] Step S4, according to the evaluation feedback, the parameters of the PID controller are dynamically adjusted through the adaptive control strategy, and the parameters of the PID controller are updated.

[0051] In this step, it should be noted that the adaptive control strategy includes: a) dynamically adjusting the PID controller parameters based on self-tuning control (STC); b) rolling optimization of long-term performance targets based on model predictive control (MPC); c) adjusting the training strategy based on model reference adaptive control (MRAC).

[0052] Step S1, collecting user interaction data with intelligent customer service to form a multi-dimensional data set;

[0053] Further, the interaction data includes interaction process data, performance indicator data and environment and state data, and step S1 includes: obtaining interaction process data, performance indicator data and environment and state data, and performing cleaning and structured processing to form a standardized multi-dimensional data set. Further, the performance indicators include: answer accuracy, response time and user satisfaction.

[0054] Specifically, the interaction data includes: interaction process data: including user input (text, speech to text, expression, etc.), customer reply content, complete dialogue chain (such as user asking "refund process"→customer reply A→user asking "accounting time"), dialogue round, response time, etc. Performance indicator data: real-time statistics of core indicators, including: answer accuracy: calculated by manual annotation or user confirmation "whether the problem is solved". Response speed: the time consumption from user input to system reply. User satisfaction: obtained by star rating after the end of the dialogue or semantic sentiment analysis on user feedback text. Tools: use ELKStack and other log systems to record structured data, and use NLP tools to analyze unstructured text. Environment and state data: including system load (such as concurrent dialogue number, model inference time consumption), user portrait (such as new and old users, historical problem types), time / scene label (such as workday peak, promotion period).

[0055] Data processing process: After obtaining the above multi-source data, clean it (remove noise, outliers), integrate and structure it to form a standardized multi-dimensional data set. This data set usually includes: original dialogue log (stored in structured JSON format); real-time performance indicator dashboard (accuracy, response speed, etc. time series data); user and system state feature library. The output of this step provides a high-quality data foundation for subsequent error calculation and control decision-making.

[0056] Step S2, based on the PID controller, the performance indicators in the multi-dimensional data set are processed, the current error signal is calculated, and the corresponding control instruction is generated according to the current error signal to adjust the response behavior of the intelligent customer service system in real time;

[0057] Further, in step S2, generating a corresponding control instruction according to the current error signal includes: inputting the current error signal into the PID controller, respectively performing proportional operation, integral operation and differential operation to obtain proportional adjustment component, integral correction component and differential prediction component; superimpose and fuse the proportional adjustment component, the integral correction component and the differential prediction component to generate a total control signal; generate a system control instruction according to the total control signal; the system control instruction is sent to the execution unit of the intelligent customer service system.

[0058] This step is responsible for converting performance gaps into specific optimization actions.

[0059] S2.1: Calculate the current error signal: First, the system obtains the preset performance target value, including: preset target accuracy (such as 95%) preset target response time (such as 2 seconds) preset target user satisfaction (such as 4.5 points), then calculate the deviation of each indicator according to the actual performance data collected in S1: accuracy deviation = preset target accuracy - actual accuracy; response time deviation = preset target response time - actual response time; user satisfaction deviation = preset target user satisfaction - actual user satisfaction; finally, according to the system optimization focus, the weighted fusion of each deviation value is performed to generate a comprehensive current error signal e(t) as the input of the PID controller. For example: e(t) = 0.5*accuracy deviation + 0.3*user satisfaction deviation + 0.2*response time deviation.

[0060] According to the PID principle, a feedback controller is designed:

[0061] Proportional control (P): directly adjust the model output according to the current error (such as dynamically adjust the confidence threshold of the reply), the proportional control adjusts the output according to the current error (i.e. the difference between the set value and the actual value), the controller output is proportional to the error, and the proportional coefficient is Kp. The advantage of proportional control is that it can quickly respond to errors, but it may not completely eliminate steady-state errors.

[0062] Integral control (I): accumulates historical errors to correct long-term systematic biases (e.g., persistent misclassification triggers model retraining), adjusts output based on the accumulation of errors (i.e., the integral of error over time), controller output is proportional to the integral of error, integral coefficient is Ki. The advantage of integral control is that it can eliminate steady-state error, the disadvantage is that it can cause the system response to slow down, even cause oscillation.

[0063] Derivative control (D): predicts error trend, adjusts strategy in advance (e.g., switch to human customer service transfer when user question complexity suddenly increases), adjusts output based on the rate of change of error (i.e., the derivative of error), controller output is proportional to the rate of change of error, derivative coefficient is Kd. The advantage of derivative control is that it can predict future error trends, reduce overshoot and oscillation, the disadvantage is sensitive to noise, may cause control output unstable.

[0064] S2.2: Generate control instructions: input the integrated error signal e(t) into the PID controller, and perform proportional (P), integral (I), and derivative (D) operations: proportional operation (P): u_P(t) = Kp * e(t). Its output is proportional to the current error, used to achieve fast and immediate adjustment. For example, if the accuracy error is large, the control instruction can be mapped to "increase the matching weight of the priority knowledge base" or "increase the strictness of keyword matching"; if the response speed error is large, it is mapped to "simplify the redundant content of the answer template". Integral operation (I): u_I(t) = Ki * ∫e(t)dt. Its output is proportional to the historical cumulative error, used to eliminate long-term steady-state error. For example, if the accuracy of the "refund process" type of question is consistently low, the control instruction can be mapped to "trigger knowledge base special update for this type of question" or "increase the number of human-labeled samples for this type of question". Derivative operation (D): u_D(t) = Kd * de(t) / dt. Its output is proportional to the rate of change of error, used to predict trends and prevent bias from expanding. For example, if user satisfaction drops sharply within 10 minutes, the control instruction can be mapped to "temporarily trigger the human customer service intervention mechanism"; if the response speed error deteriorates rapidly, it is mapped to "temporarily roll back part of the function update". Superimpose the above three components to get the total control signal: u(t) = u_P(t) + u_I(t) + u_D(t).

[0065] Control logic intervention: introduce PID control logic to correct model output: if historical data shows that "the accuracy of this type of question is low" (integral term accumulates error), force call "high confidence answer template"; if it is detected that "the current response speed is slow" (proportional term error), trigger "simplify answer mode" (reduce redundant content).

[0066] Decision output mechanism: final reply = large model generated content + control strategy correction (such as truncating too long replies, supplementing guiding language), and recording the decision process (for subsequent feedback analysis) simultaneously. The output results are as follows: (1) real-time response content of the customer service system (text, button options, etc.); (2) decision process log (such as "PID correction was called, with a response speed error of 0.8 seconds").

[0067] Finally, the total control signal is converted into specific system control instructions and delivered to the execution unit of the intelligent customer service system. The execution unit adjusts the system behavior according to the instructions, such as adjusting the knowledge base matching weight, calling high-confidence answer templates, triggering knowledge base update processes, modifying model generation parameters (such as temperature coefficients), adjusting the manual transfer threshold, etc.

[0068] Step S3: Evaluate the system performance data after adjusting the control instructions to obtain evaluation feedback.

[0069] Further, in step S3, based on Lyapunov stability theory, a system energy function is constructed according to the system performance data, and the rate of change of the energy function over time is calculated in real time; the system stability is judged according to the rate of change, and a stability analysis result is generated; the system performance indicators are calculated based on the system performance data; and the evaluation feedback report is generated by combining the system performance indicators and the stability analysis result.

[0070] It should be noted that the energy function is a Lyapunov function, which is a key indicator of system stability. When its value decreases monotonically and approaches 0 over time, it indicates that the system is stable. By calculating the rate of change of this function, it can be determined whether the system is stable. If the rate of change is negative, it means that the system energy is being consumed and the system is tending towards stability; otherwise, if the rate of change is non-negative, there may be a risk of instability.

[0071] In this step, this step is used to quantitatively evaluate the execution effect of the control instructions in S2. The evaluation not only includes performance indicators, but more importantly, the Lyapunov stability theory is introduced to analyze the system stability:

[0072] 1) Error calculation: define the error of the key indicators: accuracy error = target accuracy (e.g. 90%) - actual accuracy (e.g. 75%), satisfaction error = target satisfaction (4.5 points) - actual satisfaction (3.2 points). Comprehensive error: a single feedback signal is obtained by weighted summation (e.g. accuracy weight 0.6 + satisfaction weight 0.4).

[0073] 2) Multi-source feedback fusion: Explicit feedback: User direct rating (e.g. "Did you solve the problem" button, star rating); Implicit feedback: Conversation behavior data (e.g. whether user repeats the question, conversation is terminated prematurely, whether click "transfer to human"); Human verification: Sample annotation of high error conversations (e.g. whether the reply is accurate by the quality inspection team).

[0074] The output results are as follows: (1) Real-time error indicators (single indicator error + comprehensive error); (2) Error traceability report (e.g. "Error contribution of a certain type of problem (refund) accounts for 60%").

[0075] Construct system energy function: According to system performance data (e.g. accuracy error, error accumulation, parameter update gradient norm), construct a Lyapunov function (V function), for example: V(x) = 0.5k1x1 2 (t) + 0.5k2x2 2 (t) + 0.5k3x3 2 (t), where x1(t) is the accuracy error, x2(t) is the integral error, and x3(t) is the gradient norm. The function represents the "energy" of the system, and the smaller the value, the closer the system is to a stable state.

[0076] Calculate the rate of change of energy function: Real-time calculation of the rate of change of the V function with time dV / dt.

[0077] Judge stability: According to whether dV / dt is continuously less than zero to judge whether the system is stable (i.e. whether the energy is continuously dissipated and the performance is converging positively). If dV / dt ≥ 0, it indicates that the system has a risk of divergence.

[0078] Calculate system performance indicators based on system performance data: Real-time statistical and aggregate calculation of the system performance data to obtain quantitative system performance indicators including answer accuracy, average response time, user satisfaction and human intervention rate; wherein the answer accuracy is calculated by the percentage of the number of correct answers to the total number of conversations, the average response time is calculated by the arithmetic mean of all response times, the user satisfaction is calculated by the weighted average of user ratings, and the human intervention rate is calculated by the percentage of conversations that need human intervention to the total number of conversations.

[0079] Generate evaluation feedback report: Comprehensive performance indicators (accuracy, response speed, etc.) real-time values and stability analysis results to generate a comprehensive evaluation feedback report. The report not only contains the error size, but also contains the error source analysis (e.g. a certain type of problem error contribution accounts for 60%) and the conclusion of system stability.

[0080] The steps to control the model not to diverge using Lyapunov stability theory are as follows:

[0081] Step 1: Define system state and stability criterion;

[0082] 1) Selection of system state vector: Select key variables that reflect the self-learning process to form the state x(t): x1(t): current performance error (e.g., target accuracy - actual accuracy); x2(t): error accumulation (integral term, reflecting historical deviation); x3(t): gradient norm of model parameter update (reflecting the degree of parameter adjustment). State vector: x(t) = [x1(t), x2(t), x3(t)] T .

[0083] 2) Stability definition: Equilibrium point: ideal stable state x* = [0, 0, 0] T (error is 0, there is no cumulative deviation, and there is no drastic adjustment of parameters). Asymptotic stability: if x(t) → x as t → ∞, the system is stable (performance converges to the target); if ||x(t)|| → ∞, the performance diverges (unstable).

[0084] Step 2: Construct Lyapunov function (V function);

[0085] V function needs to meet 3 conditions: (1) positive definite: V(x) > 0 when x ≠ x, V(x*) = 0 (when deviating from the equilibrium point, the "energy" is positive, and at the equilibrium point it is 0); (2) continuous and differentiable: ensure that the derivative can be analyzed for change trend; (3) derivative negative: dV / dt < 0 when x ≠ x* ("energy" monotonically decreases over time).

[0086] For the V function design of intelligent customer service, combined with the system state, construct the following quadratic V function (physical meaning: "total energy" of error and parameter fluctuation):

[0087] V(x) = 1 / 2k1x1 2 (t) + 1 / 2k2x2 2 (t) + 1 / 2k3x3 2 (t);

[0088] where k1, k2, k3 > 0 are weight coefficients (can be adjusted according to the importance of performance indicators, such as higher weight for accuracy error).

[0089] Step 3: Analyze the derivative of V function to ensure stability;

[0090] By calculating dV / dt and proving its negativity, determine whether the system is stable.

[0091] 1) Derivation process, according to chain rule:

[0092] dV / dt = k1x1(t)·dx1 / dt + k2x2(t)·dx2 / dt + k3x3(t)·dx3 / dt;

[0093] where the derivative of each state needs to be defined in combination with the self-learning mechanism:

[0094] dx1 / dt: rate of change of error (e.g. d(accuracy error) / dt, related to model output correction);

[0095] dx2 / dt: rate of change of error accumulation (i.e. current error x1(t), since x2(t) = ∫x1(τ)dτ);

[0096] dx3 / dt: rate of change of parameter update gradient (related to model fine-tuning learning rate, gradient clipping strategy).

[0097] 2) Combine self-learning rule constraints to make derivatives negative, to make dV / dt < 0, the self-learning strategy of intelligent customer service needs to be constrained:

[0098] Error correction rule: if x1(t) > 0 (accuracy lower than target), then dx1 / dt < 0 (reduce error through PID proportional term), ensure k1x1*dx1 / dt < 0.

[0099] Integral term control: limit the growth of x2(t) (e.g. set a cumulative error upper limit), to avoid integral saturation leading to dx2 / dt out of control.

[0100] Parameter update constraint: clip the gradient x3(t) of model fine-tuning (e.g. ||x3(t)|| ≤ G_max), to ensure that dx3 / dt does not fluctuate violently, maintaining k3x3*dx3 / dt < 0.

[0101] Step 4: Design a stabilization mechanism to avoid performance divergence;

[0102] If analysis finds that dV / dt may be non-negative (there is a risk of instability), the following mechanisms need to be added:

[0103] 1) Dynamically adjust learning intensity: when V(x) increases (system energy rises), automatically reduce the learning rate of model fine-tuning or the gain of PID controller (e.g. reduce Kp, Ki), to avoid aggressive parameter updates.

[0104] 2) Introduce "safety boundary": set a threshold for state variables (e.g. x1(t) ≤ ε_max, ε_max is the maximum acceptable error), when approaching the threshold, trigger the protection mechanism: suspend self-learning, enable historical optimal parameters; increase the proportion of manual review samples to correct error sources.

[0105] 3) Periodic reset and calibration: periodically (e.g. every week) calculate V(x), if its value has not decreased for a long time (the system is trapped in local instability), reset part of the state (e.g. clear the cumulative error x2(t)), and reinitialize the learning process.

[0106] The application effect is to ensure the stability of the self-learning process, including:

[0107] 1) Prevent performance divergence: through the negative definiteness constraint of the derivative of the V function, it is ensured that the error and parameter fluctuations will not increase indefinitely (such as the accuracy rate will not drop from 90% to 50%).

[0108] 2) Accelerate convergence: the monotone decreasing of the V function guarantees the continuous consumption of system energy, which promotes the rapid convergence of performance indicators to the target (such as the error gradually converges from 20% to within 5%).

[0109] 3) Robustness improvement: even if there is a sudden disturbance (such as a malicious user input causing a short-term increase in error), the system can return to the right track through the stabilization mechanism to avoid the "butterfly effect".

[0110] Step S4, according to the evaluation feedback, dynamically adjust the parameters of the PID controller through the adaptive control strategy, and feed back the adjusted parameters to the PID controller in step S2.

[0111] In this step, it needs to be explained that the adaptive control strategy includes: a) dynamically adjusting the parameters of the PID controller based on self-tuning control (STC); b) rolling optimization of long-term performance targets based on model predictive control (MPC); c) adjusting the training strategy based on model reference adaptive control (MRAC).

[0112] This step builds the "actuator" of the control system. The core goal is to adjust system parameters and models based on feedback error signals to reduce error (similar to the "execution layer" in a control system).

[0113] Further, in step S4, dynamically adjusting the parameters through the adaptive control strategy includes: adjusting the proportional coefficient, integral coefficient or derivative coefficient of the PID controller according to the stability analysis results in the evaluation feedback report, and feeding back the adjusted parameters to the PID controller in step S2, so that the system energy function decreases over time.

[0114] 4.1: PID parameter adjustment based on stability results: according to the stability analysis results in the evaluation feedback report generated in S3, dynamically adjust the parameters (Kp, Ki, Kd) of the PID controller, the goal is to make the Lyapunov function value decrease over time, ensure the stable convergence of the system. For example: if the system response is slow but stable, Kp can be appropriately increased. If there is overshoot or oscillation, Kp can be appropriately reduced or Kd can be appropriately increased. If there is a steady-state error, Ki can be appropriately increased. The adjusted parameters are immediately fed back to the PID controller in S2 to realize online adaptive parameters.

[0115] Further, the dynamic parameter adjustment by the adaptive control strategy further comprises: by the self-tuning control strategy, real-time online estimation of dynamic model parameters of the intelligent customer service system; wherein the dynamic model parameters are used to represent changes in the mapping relationship between system inputs and control outputs, the system inputs include knowledge base matching weights or model generation parameters, and the control outputs include response accuracy or response speed.

[0116] S4.2: Self-Tuning Control (STC): To further improve adaptability, the Self-Tuning Control (STC) strategy is adopted.

[0117] Self-Tuning Control (STC) is an adaptive control strategy that aims to maintain optimal performance of the control system during operation by online estimation of system parameters and automatic adjustment of controller parameters. STC is particularly suitable for situations where system parameters change over time or there is uncertainty.

[0118] Core framework of STC in intelligent customer service: The classic structure of self-tuning control includes "parameter estimator" and "controller parameter adjuster", which maps to the intelligent customer service system as: user interaction → data collection → system parameter estimation → controller parameter correction → optimized decision output → user interaction (closed loop). Core goal: When system characteristics change (such as sudden change in user consultation type, model performance decline), automatically adjust the control strategy to maintain accuracy, response speed, and other indicators in the optimal range.

[0119] Steps to maintain the optimal performance of intelligent customer service using STC:

[0120] Step 1: Define system model and performance indicators;

[0121] 1) Establish a dynamic system model for intelligent customer service, model the input-output relationship of the customer service system as a recognizable mathematical model, for example:

[0122] Input (u(t)): Adjusting action of the controller (such as knowledge base matching weight, model generation parameter α);

[0123] Output (y(t)): System performance indicators (such as actual accuracy P(t), response speed T(t));

[0124] Model equation: Use a linear regression model to approximate the system dynamics (adjustable according to actual complexity):

[0125] y(t) = a1·y(t-1) + b1·u(t-1) + ξ(t);

[0126] Where a1, b1 are system parameters to be estimated, and ξ(t) is random disturbance (such as sudden user demand).

[0127] 2) Set optimal performance benchmark: define the quantitative target of "best performance", e.g.: accuracy P* = 95% (±2% is acceptable interval); response speed T* = 1.5s (±0.3s is acceptable interval); comprehensive performance index J = 0.7P(t) + 0.3(1 / T(t) (maximize this value).

[0128] Step 2: Online parameter estimation (identify system dynamic changes);

[0129] The parameter estimator analyzes the system input and output data in real time, identifies the changes in model parameters (such as a1, b1), and reflects the drift of system characteristics (such as changes in user demand, model aging).

[0130] 1) Data window design: use sliding window to collect the latest interaction data (such as the last 1000 conversations), including: controller output u(t-1) (such as "knowledge base weight = 0.8"); system output y(t) (such as "actual accuracy = 92%"). The window size needs to balance real-time performance and estimation accuracy (too small window is easily disturbed by noise, too large window is lagging).

[0131] 2) Estimation algorithm selection:

[0132] Least squares method (RLS): suitable for stationary scenarios, recursively update parameter estimation values:

[0133]

[0134] where is the parameter estimation vector, is the data vector, K(t) is the gain matrix.

[0135] Extended Kalman filter (EKF): suitable for nonlinear scenarios (such as nonlinear changes in large model generation quality).

[0136] 3) Output: Real-time system parameters: continuously output reflect the dynamic characteristics of the current system (such as increase the impact of controller adjustment on performance enhancement).

[0137] Step 3: Controller parameter self-correction (adjust and optimize strategy);

[0138] Based on the estimated system parameters, the controller parameter adjuster adjusts the control strategy in real time to make the system output track the optimal target.

[0139] 1) Controller structure design: use "parameter adaptive PID" as the core controller, its output (adjustment action) is:

[0140] u(t) = Kp(t) e(t) + Ki(t) ∫e(τ)dτ + Kd(t) de(t) / dt;

[0141] where e(t) = target value y* - actual value y(t) is the performance error, Kp(t), Ki(t), Kd(t) are real-time corrected PID parameters.

[0142] 2) Parameter correction rule: according to the estimated system parameters Dynamic adjustment of PID parameters:

[0143] When the system response is slow (such as Increase: It means that the historical performance has a stronger impact on the current situation: increase Kp(t) (speed up the current error correction) and decrease Ki(t) (avoid overaccumulation).

[0144] When the controller adjustment effect is weakened (such as Decrease: Simultaneously increase Kp(t), Ki(t), and Kd(t) (enhance the adjustment strength).

[0145] When the disturbance noise increases (such as the variance of ξ(t) rising): decrease Kd(t) (reduce the sensitivity to high-frequency noise).

[0146] 3) Correction strategy for special scenarios:

[0147] User demand mutation (such as promotion activities leading to changes in consultation types): trigger parameter reset mechanism, quickly re-estimate parameters based on new data (shorten window size to 200 dialogues), and temporarily increase Kp(t) to quickly adapt.

[0148] Model performance degradation (such as continuous accuracy decline): increase Ki(t) weight and accumulate adjustment through integral term (such as strengthening knowledge base update and increasing fine-tuning frequency).

[0149] Step 4: Closed-loop verification and stability guarantee;

[0150] 1) Performance monitoring and feedback: Real-time calculation of the deviation between the corrected system output y(t) and the target y*: if the deviation continuously exceeds the acceptable interval (such as accuracy < 93% for 30 minutes), then: trigger the "forgetting factor" adjustment of the parameter estimator (increase the forgetting factor to accelerate learning on new data); enable the backup control strategy (such as temporarily switching to the artificially optimized parameter group).

[0151] 2) Stability constraint: Combine Lyapunov stability theory to design the "safety boundary" of the performance index: set the upper limit of the parameter estimation error (such as ); When the control quantity u(t) exceeds the safety range (such as knowledge base weight > 0.95 may lead to overfitting), force it to the boundary value.

[0152] Application effect: maintain the best performance of intelligent customer service: 1) adaptive dynamic change: when the user consultation changes from "after-sales problem" to "new product consultation", STC can complete parameter adjustment within 1 hour, and the accuracy rate is maintained at 94% or more (without STC, it may drop to 85%). 2) Resist model decay: when the performance of the large model decreases after long-term use, STC adjusts the integral by continuously increasing the integral (Ki), promotes knowledge base iteration and model fine-tuning, and reduces the accuracy decay rate by 60%. 3) Reduce manual intervention: the system can automatically respond to 80% of performance fluctuation scenarios (such as time-specific consultation peak), and the frequency of manual parameter adjustment is reduced from 3 times a week to 1 time a month.

[0153] Further, the dynamic parameter adjustment by the adaptive control strategy further includes: judging the degree of drift of system characteristics based on the estimated change of dynamic model parameters; dynamically adjusting the gain parameters of the PID controller according to the degree of drift, so that the performance indicators of the intelligent customer service system track the preset optimal performance target; wherein the gain parameters include one or more combinations of proportional coefficient, integral coefficient and differential coefficient.

[0154] Further, the dynamic parameter adjustment by the adaptive control strategy further includes: using a model predictive control strategy to predict the system performance trend in the future period of time based on the current system state and the dynamic model parameters; taking the long-term performance indicator as the optimal target, rolling optimization of the control sequence in the future limited time domain, and executing the optimal control action at the current time; wherein the long-term performance indicator includes user satisfaction cumulative value or monthly problem solving rate.

[0155] S4.3: Model predictive control (MPC): for optimizing long-term goals, using model predictive control (MPC) strategy.

[0156] Model predictive control (MPC) is an advanced control strategy widely used in industrial process control and other complex systems. The core idea of MPC is to use the mathematical model of the system to predict the future output, and on this basis, optimize the control action to make the system as close as possible to the set target value in the future period of time.

[0157] Main features of MPC: Model-based: MPC uses a mathematical model of the system to predict future behavior. This model can be linear or nonlinear, depending on the characteristics of the system. Optimal control: MPC determines the optimal control sequence by solving an optimization problem. The optimization objective is usually to minimize some performance metric, such as tracking error, control effort, energy consumption, etc. Rolling optimization: MPC adopts a rolling optimization (also known as receding horizon optimization) approach. At each control decision, only the first control action of the optimized sequence is executed, and then the optimization is re-performed at the next sampling time, forming a continuous rolling process. Constraint handling: MPC can handle various constraints of the system, such as input limits, output limits, state variable limits, etc., ensuring that the control actions are within the feasible range. Multivariable control: MPC can handle multiple input and output variables simultaneously, suitable for the control of multivariable systems.

[0158] In the intelligent customer service system, the core of applying model predictive control (MPC) is to build a system dynamic model to predict the performance trend (such as user satisfaction, problem solving rate) in the future period of time, and optimize the control strategy in advance to achieve the long-term goal (such as monthly user satisfaction ≥ 90%).

[0159] The core framework of MPC in intelligent customer service: The classic structure of MPC includes "prediction model", "rolling optimization" and "feedback correction", which maps to the intelligent customer service system as: historical data → build prediction model → predict future performance → optimize control sequence → execute current control → collect new data (closed loop). Core goal: While meeting real-time constraints (such as response speed), avoid short-term decisions that harm long-term goals (such as excessive simplification of replies to improve speed but reduce satisfaction).

[0160] Steps to achieve self-learning of intelligent customer service using MPC to optimize long-term goals:

[0161] Step 1: Build a prediction model for intelligent customer service;

[0162] The prediction model is the basis of MPC, which describes the dynamic relationship between system inputs (control actions) and outputs (performance indicators), and predicts the system state at the next N time points.

[0163] 1) Define system state and variables:

[0164] State variables (x(t)): Key indicators reflecting the current state of the system, such as: cumulative user satisfaction score s(t), knowledge base matching accuracy a(t); unsolved problem backlog b(t), system load rate l(t) (such as the number of concurrent conversations / maximum capacity).

[0165] Control variables (u(t)): adjustable optimization actions of the system, such as: knowledge base update frequency u1(t) (times / day), large model generation parameters u2(t) (such as temperature coefficient, control reply creativity), artificial intervention threshold u3(t) (such as when satisfaction <3 points, transfer to artificial).

[0166] Output variables (y(t)): performance indicators to be optimized, such as: user satisfaction at future time t y1(t) = f(s(t), u1(t), u2(t)), problem solving rate at future time t y2(t) = g(a(t), b(t), u3(t)).

[0167] 2) Establish a prediction model equation: train a data-driven model (such as LSTM, random forest) based on historical data to describe the dynamic relationship between state and control variable:

[0168] x(t+1) = F(x(t), u(t) # state transition equation;

[0169] y(t) = H(x(t) # output equation;

[0170] Example: if x(t) = [s(t), a(t)], u(t) = [u1(t), u2(t)], then the model can be represented as:

[0171] s(t+1) = 0.8s(t) + 0.1u1(t) - 0.05u2(t) + ξ(t);

[0172] Physical meaning: current satisfaction has the greatest impact on the future, knowledge base update positively improves satisfaction, excessive pursuit of reply creativity may reduce satisfaction, and ξ(t) is random noise.

[0173] 3) Set the prediction time domain and control time domain: prediction time domain (Np): predict the number of steps in the future (such as Np = 7 days, predict the performance of the next week). Control time domain (Nc): optimization control sequence length (such as Nc = 3 days, optimize future 3-day control actions, and use the last control amount for the remaining days). Selection principle: Np needs to cover the main dynamic changes of the system (such as the cycle of user demand), and Nc ≤ Np to balance the amount of calculation.

[0174] Step 2: Rolling optimization to solve the future optimal control sequence;

[0175] At each decision-making time (such as every day), based on the prediction model and the current state, solve the control sequence for the next Nc steps, so that the output for the next Np steps is as close to the target as possible.

[0176] 1) Define the optimization objective function: the objective function needs to integrate short-term performance and long-term goals, for example:

[0177] min J = Σt＝1 np [w1(y1(t)-y1) 2 +w2(y2(t)-y2)]+Σ t＝1 nc [r1u1(t) 2 +r2u2(t) 2 ];

[0178] where Np is prediction horizon, Nc is control horizon; y1(t) is predicted user satisfaction at time t, y1 is target value of user satisfaction (e.g. 90%); y2(t) is predicted problem resolution rate at time t, y2 is target value of problem resolution rate (e.g. 95%); w1, w2 are weight coefficients of tracking error term; u1(t), u2(t) are control variables at time t; r1, r2 are weight coefficients of control cost term. The first term is tracking error term, which is used to minimize the deviation between performance index and target value. The second term is control cost term, which is used to avoid too aggressive control actions (e.g. r1 limits the resource consumption of high frequency updates of knowledge base).

[0179] 2) Constraint setting: control variables and output variables need to satisfy actual constraints:

[0180] Control constraints: u1_min≤u1(t)≤u1_max (e.g. knowledge base is updated no more than 5 times per day); u2(t)∈[0.1, 0.8] (temperature coefficient range).

[0181] Output constraints: y1(t)≥85% (satisfaction baseline); l(t)≤0.8 (system load no more than 80%).

[0182] 3) Solve optimization problem: use numerical optimization algorithm (e.g. interior point method, particle swarm optimization) to solve the constrained objective function, get the optimal control sequence U* = [u(1), u(2),..., u*(Nc)].

[0183] Step 3: Feedback correction - execute current control and update model;

[0184] The core of MPC is "rolling optimization + feedback correction", only execute the first control action of the optimal sequence, re-optimize based on new data at next time to offset prediction error.

[0185] 1) Execute current control: at current time t, only apply u*(1) (e.g. "update knowledge base 3 times today, temperature coefficient is set to 0.5"), the remaining control variables u(2)...u(Nc) are only used as prediction reference.

[0186] 2) Collect real-time data and update state: Obtain the actual output y_actual(t+1) at time t+1 (e.g., actual satisfaction 92%), compare with the predicted output y_pred(t+1), and calculate the prediction error: e(t+1) = y_actual(t+1) - y_pred(t+1).

[0187] 3) Correct the prediction model: If the error is small (e.g., |e| < 2%): fine-tune the model parameters (e.g., update the coefficients of F(x, u) using recursive least squares). If the error is large (e.g., a sudden promotion causes a sharp drop in satisfaction): trigger model retraining (update the prediction model with the latest 3 days of data) and expand the prediction horizon for the next time (e.g., Np increases from 7 days to 10 days).

[0188] MPC optimization example for intelligent customer service scenario: Suppose the system goal is "monthly satisfaction ≥ 90%", and the predicted satisfaction trend for the next 7 days is as follows: 1) If only short-term optimization is performed (e.g., speed up response), the satisfaction may rise to 92% in the next 3 days, but then drop to 86% due to incomplete problem resolution (long-term substandard). 2) MPC, by predicting this trend, will adjust the control sequence: Days 6-8: reduce the response speed weight (u2 decreases), increase the knowledge base depth update (u1 increases), and allow the satisfaction to temporarily drop to 90%; Days 9-12: based on the optimized knowledge base, gradually increase the response efficiency, and keep the satisfaction at 91-93%, ensuring that the monthly target is met.

[0189] Application effect: Achieve long-term optimization of intelligent customer service: 1) Balance short-term and long-term goals: avoid short-sighted behavior of "sacrificing problem-solving quality to improve immediate response speed", and increase the monthly satisfaction target achievement rate by 25%. 2) Respond to periodic fluctuations: by predicting the weekly / monthly cycles of user demand (e.g., a surge in weekend inquiries), adjust resource allocation in advance (e.g., pre-update the knowledge base), and reduce load fluctuations by 40%. 3) Enhance anti-interference capability: when a sudden event (e.g., system failure) causes short-term performance to decline, MPC can develop recovery strategies by extending the prediction horizon, reducing the performance recovery period by 50%.

[0190] c) Adjust the training strategy based on Model Reference Adaptive Control (MRAC):

[0191] Model Reference Adaptive Control (MRAC) is an advanced control strategy that aims to make the response of the controlled system as close as possible to the response of an ideal reference model. MRAC is particularly suitable for situations where system parameters are unknown or change over time, and adjusts the controller parameters online to adapt to the dynamic changes of the system, thereby achieving stable control performance.

[0192] Using model reference adaptive control, set the reference model (such as the ideal customer service response accuracy rate, response speed), real-time comparison of the actual model and the reference model output difference, automatic adjustment of training strategy (such as dynamic allocation of priority of labeled data, adjustment of training batch size).

[0193] The specific method of strategy optimization is as follows:

[0194] 1) Parameter adjustment (based on PID logic): proportional adjustment (immediate correction): if the accuracy error of a certain problem is > 15%, immediately increase the keyword matching weight of this type of problem; integral adjustment (long-term optimization): if the cumulative error of a certain type of problem is > 10% for 1 week, trigger knowledge base update (such as supplementing high-quality samples of this type of problem); differential adjustment (trend prevention): if the satisfaction error increases by 5% within 1 hour, temporarily increase the threshold for human intervention (such as "automatically transfer to human when the user's emotion is negative").

[0195] 2) Model iteration: use high-error dialogue data to construct fine-tuning data sets (such as filtering "wrong reply" samples), regularly fine-tune large models (such as training 3-5 epochs with new data every week); dynamically adjust model hyperparameters: such as when the system load is too high, reduce the maximum length of generated replies (reduce reasoning time consumption).

[0196] 3) Strategy upgrade: optimize rule engine based on error patterns: such as finding that "users pay more attention to efficiency at night", automatically switch to night mode (reply more concisely); A / B test verification: small-scale test (such as 5% of users) of optimization strategies, confirm effectiveness before full-scale promotion.

[0197] The output results are as follows: (1) optimized system parameters (such as matching weight, human intervention threshold); (2) iterated model version (fine-tuned large model); (3) strategy update log (such as "2024-10-01 update knowledge base 30 items for refund issues").

[0198] Through the above S1 to S4 loop execution, the intelligent customer service system forms a complete, "data collection - model decision - effect feedback - strategy optimization" closed-loop self-learning system, which can continuously learn from interaction, dynamically adjust its own behavior, and ultimately achieve continuous performance improvement and stable user experience.

[0199] Finally, the final effect of the closed-loop operation is as follows:

[0200] 1) System performance continues to improve: accuracy, user satisfaction and other indicators gradually approach target values (such as accuracy rate from 75% to 92% in 3 months); error fluctuation range decreases (such as satisfaction error from ±1.5 points to ±0.3 points).

[0201] 2) Adaptive capability enhancement: automatically cope with scenario changes (e.g., dynamically optimize response speed during promotional period consultation surge); accelerate learning speed for new types of problems (e.g., system update after function consultation) (error convergence time shortened by 50%).

[0202] 3) Reduced human intervention: improved automatic repair rate for high error scenarios (e.g., 80% of repeated problem errors can be solved through system self-optimization), reducing human operation cost.

[0203] Through this closed loop, the intelligent customer service system realizes the transformation from "passive response" to "active optimization", combining the "dynamic balance" idea in control theory with the learning ability of AI large model, and finally achieving the control goal of "error minimization".

[0204] System implementation and effect: the system of the embodiment preferably adopts a "control brain" architecture. The control brain is a lightweight module independent of the AI large model, adopting a hierarchical structure of "perception-decision-execution-feedback", and interacting with the large model through API interface (such as sending parameters and obtaining feedback), so as to realize resource isolation and efficient decision-making.

[0205] System architecture: build a closed-loop system of "control theory + large model";

[0206] 1) Define system input, output and state variables;

[0207] Controlled object: intelligent customer service based on AI large model (such as dialogue generation model, intent classification model).

[0208] Input signal: user query, historical dialogue context, business knowledge base.

[0209] Output signal: customer service response text, intent recognition result, user satisfaction score.

[0210] State variable: model parameter (such as large model weight), dialogue history state (such as multi-round dialogue context vector), performance indicator (accuracy, response time, user complaint rate).

[0211] 2) Core control module;

[0212] Sensor (data acquisition): real-time acquisition of user interaction data (dialogue log, click behavior, human customer service result, NPS score).

[0213] Controller (decision unit): calculate "control quantity" (such as model fine-tuning strategy, data enhancement scheme, knowledge graph update rule) according to feedback signal.

[0214] Actuator (execution unit): converts control quantities into specific operations (such as triggering model incremental training, adjusting dialogue strategy, updating knowledge base).

[0215] Develop intelligent customer service self-learning function based on automatic control theory:

[0216] I. Implementation steps of developing intelligent customer service self-learning function based on automatic control theory:

[0217] 1) Data layer: build high-quality feedback dataset;

[0218] Collect multi-dimensional data: explicit feedback includes user ratings, complaint labels, manual correction records, etc.; implicit feedback includes conversation duration, click link behavior, repeated question rate, etc.

[0219] Data preprocessing: through noise filtering technology in control theory (such as Kalman filtering), remove abnormal feedback (such as malicious ratings), retain effective error signals.

[0220] 2) Model layer: large model combined with control algorithm;

[0221] Lightweight control module: build an independent "control brain" (such as rule-based or small model controller) outside the large model, calculate control signals in real time (avoid high cost of frequent full training of large model).

[0222] Efficient fine-tuning technology: combine gradient constraints of control theory (such as limiting parameter update step), use LoRA, QLoRA and other low-rank adaptation methods to quickly adapt to new data while maintaining model stability.

[0223] 3) Control layer: design dynamic optimization strategy;

[0224] Hierarchical control architecture: strategy layer, set long-term goals (such as 10% quarterly satisfaction improvement), develop macro learning strategy (such as periodic full training, real-time incremental fine-tuning); execution layer, based on real-time error signal, dynamically adjust micro parameters (such as learning rate, training data sampling rate), use PID algorithm to achieve rapid response. The mapping execution logic of intelligent customer service based on PID input and output is shown in Table 1.

[0225] Table 1 Mapping execution logic of intelligent customer service based on PID input and output;

[0226]

[0227]

[0228] Exception handling mechanism: when the system detects a sudden performance drop (such as error exceeding threshold), trigger "emergency brake" (suspend automatic learning, switch to manual review) to avoid out-of-control.

[0229] 4) Evaluation layer: Multi-index closed-loop verification;

[0230] Define control theory-related evaluation indexes: response speed (rise time), stability (oscillation amplitude), tracking accuracy (intention recognition accuracy), robustness (cross-domain dialogue performance retention rate).

[0231] Periodically perform frequency domain analysis, such as using Bode plots and Nyquist criteria, to evaluate the response characteristics of the system under different user demand frequencies (such as stability when dealing with sudden high-frequency problems).

[0232] Combination of large model and control algorithm;

[0233] Core architecture of control brain; control brain adopts a hierarchical structure of "perception-decision-execution-feedback", independent of large model but forms a closed-loop interaction with it.

[0234] 1) Perception layer: data preprocessing and feature extraction;

[0235] Function: Extract key features required for control from raw data to reduce the computational cost of the decision layer.

[0236] Input data: real-time interaction data (user question type, dialogue turn, response time); performance indicators (accuracy, satisfaction, real-time statistics of transfer to human rate); system status (large model inference delay, knowledge base matching confidence).

[0237] Processing method: Use rule engine to filter high-priority features (such as "repeat question frequency ≥ 3" marked as high urgency); use lightweight model (such as logistic regression) to convert unstructured text into structured features (such as user emotion label: positive / negative).

[0238] 2) Decision layer: core control algorithm (control strategy generation);

[0239] This is the core of the control brain, based on the features of the perception layer and the preset goals, it calculates the control signal in real time (such as adjusting the large model parameters, triggering knowledge base update).

[0240] Core algorithm combination: basic controller: PID controller (handle linear, high-frequency fluctuation scenarios); advanced controller: model predictive control (MPC, handle long-term optimization scenarios); adaptive mechanism: self-tuning control (STC, dynamically adjust controller parameters).

[0241] 3) Execution layer: transformation and delivery of control signals;

[0242] Function: Convert abstract control signals output by the decision layer (such as "increase matching weight") into specific instructions executable by the large model or system.

[0243] Output instruction types: Large model parameter adjustment (e.g., generate temperature coefficient from 0.7→0.3, reduce reply randomness); knowledge base operation (e.g., "Add this question to the high-frequency library, match weight +20%"); process switching (e.g., "Switch to manual intervention threshold from 3 stars→4 stars, intervene high-risk conversations in advance").

[0244] 4) Evaluation layer: Analysis of feedback data and error calculation;

[0245] Function: Quantitative control of instruction execution effect, providing error signals for decision-making layer.

[0246] Core indicators: Tracking error (e.g., "Actual accuracy - target accuracy"); stability indicators (e.g., error fluctuation variance of consecutive 5 interactions); cost indicators (e.g., additional computational resource consumption caused by control instructions).

[0247] Core algorithm design of control brain;

[0248] Control brain needs to balance real-time performance and optimization effect, adopts "multi-algorithm collaboration" strategy, and dynamically calls for different scenarios:

[0249] 1) Basic control: PID controller (real-time error correction);

[0250] Applicable scenarios: high-frequency, linear performance fluctuations (e.g., sudden response speed slowdown, short-term accuracy decline).

[0251] Algorithm logic: Proportional term (P): directly adjust according to current error (e.g., "response delay = 1.2 seconds") (e.g., "simplify large model generation length"); integral term (I): accumulate historical errors (e.g., "10 consecutive incorrect answers to certain questions"), trigger knowledge base update; derivative term (D): predict error trend (e.g., "satisfaction drops 1.5 stars in 5 minutes"), start manual intervention in advance.

[0252] Lightweight design: use lookup table method instead of complex calculation (e.g., pre-set error-adjustment mapping table), response time controlled within 10ms.

[0253] 2) Long-term optimization: Model Predictive Control (MPC) sub-module;

[0254] Applicable scenarios: periodic, non-linear long-term goals (e.g., "monthly satisfaction ≥90%" "promotion period resource allocation").

[0255] Algorithm logic: use lightweight prediction model (e.g. linear regression + time series decomposition) to predict performance trend in next 7 days; aim at "minimize long-term error + minimize control cost", solve optimal control sequence (e.g. "future 3-day knowledge base update frequency: 2 times -> 3 times -> 2 times"); only execute current step control, next period re-optimized based on new data (rolling optimization).

[0256] Lightweight design: simplify prediction model (discard complex neural network, use analytical formula), reduce optimization problem dimension from 100+ to 10+, calculation time ≤1 second.

[0257] 3) Adaptive adjustment: self-tuning control (STC) mechanism;

[0258] Applicable scenarios: system characteristic mutation (e.g. user demand type change, performance drift after large model version update).

[0259] Algorithm logic: online estimate system parameters (e.g. "sensitivity of large model generation quality to temperature coefficient"); when parameter change exceeds threshold (e.g. sensitivity decreases by 30%), automatically correct core parameters of PID / MPC (e.g. increase PID proportional coefficient); use forgetting factor recursive least squares (RLS) to update parameter estimation in real time, considering the weight of historical data and new data.

[0260] 4) Rule engine: emergency handling and boundary constraints;

[0261] Function: as a "safety net", handle extreme scenarios that algorithms cannot cover, avoid improper control.

[0262] Core rule examples: when system load > 90%: forcibly trigger "minimal recovery mode" (ignore long-term optimization of MPC, prioritize availability); when user explicitly complains (e.g. input "bad review"): skip PID adjustment, directly switch to manual (highest priority); consecutive 3 times invalid (error not reduced): suspend automatic adjustment, enable historical optimal parameter set.

[0263] Collaborative mechanism with large model: brain does not directly participate in inference process of large model, but realizes lightweight interaction through "signal interface":

[0264] 1) Control signal issuance: control brain transmits parameters to large model through API (e.g. {"temperature": 0.3, "max_tokens": 50}), without modifying model structure;

[0265] 2) Feedback data acquisition: large model output results (reply content, confidence) and user feedback (score, behavior) are synchronized to control brain in real time through log interface;

[0266] 3) Resource isolation: Control brain is deployed in independent lightweight servers (such as 4-core 8G), which is isolated from GPU resources of large models to avoid competition for computing resources.

[0267] Evaluation results comparison;

[0268] AI large model deployment is based on FastAPI framework, and inference optimization uses TensorRT-LLM0.7.0, quantized to FP16 precision.

[0269] According to the following models, the effects of intelligent customer service self-learning are compared: 1) Rule-based, based on regular expression matching, containing 2000 hand-crafted rules; 2) LSTM-CS: 3-layer LSTM + Attention, hidden layer dimension 512; 3) Qwen2.5: directly fine-tune Qwen2.5-32B model, without using domain adapter; 4) The model of the present application: train Qwen2.5-32B model based on the above automatic control method of the present application.

[0270] The following table is the index and definition of evaluating the performance of intelligent customer service:

[0271] Table 2 Intelligent customer service evaluation index and definition:

[0272] Indicator Definition Intention accuracy (Acc) Predicting intent and attention Figure 1 Proportion of successful Human intervention rate (HIR) The proportion of intelligent customer service that cannot be handled and is transferred to human (the lower the ideal value is, the better) Load standard deviation (σ Load) The standard deviation of the current number of sessions of human customer service (reflecting the balance of distribution, the ideal value tends to 0) Response delay (ms) The time interval from the user's question to the first reply (including model reasoning and routing time consumption)

[0273] The data set is as follows:

[0274] Labeled data: 20,000 conversations, covering 15 core intents (such as product consultation, return and exchange, invoice application, etc.);

[0275] Unlabeled data: 80,000 conversations for unsupervised clustering and data generation.

[0276] The following table is a comparison of the intelligent customer service index performance of each model:

[0277] Table 3 Comparison of index performance of each model

[0278] Model Acc (%) HIR (%) σ Load Response delay (ms) Rule-based 68.2 35.5 4.2 280 LSTM-CS 79.1 28.7 3.8 320 Qwen2.5 85.3 22.1 3.5 450 The model of the present application 89.7 18.4 2.1 380

[0279] Evaluation result analysis:

[0280] 1) Intent understanding improvement: Domain adapter enables LLM to quickly capture e-commerce terminology (such as "registration fee", "pre-sale deposit"), combined with dynamic knowledge graph updates, the recognition accuracy of emerging issues (such as "collective procurement rules") is 4.4% higher than Qwen2.5;

[0281] 2) Artificial efficiency optimization: the algorithm reduces the load standard deviation from 3.5 to 2.1, the maximum number of simultaneous conversations for customer service from 12 to 7, and the average response time is shortened to 98 seconds;

[0282] 3) Delay control: Through few-shot learning and gradient update of the model, the response delay is reduced by 46% compared to the original LLM, meeting the real-time interaction requirements of e-commerce.

[0283] Self-learning effect analysis:

[0284] 1) Unsupervised path adds about 12,000 new effective training data per month, making the model's Acc percentage on long-tail intents (such as "gift card purchase") increase by 15% on average per year;

[0285] 2) The number of knowledge graph entities increased from 5,000 initially to 18,000, and the number of relationship types expanded from 12 to 35, covering 92% of user consultation scenarios.

[0286] 3) Intelligent customer service has excellent performance in HIR (human access rate), with only 18.4%.

[0287] There are several alternatives to achieve the self-learning function of intelligent customer service, such as:

[0288] (1) Incremental model update: Use feedback data as new samples and update the model using incremental learning algorithms (such as online gradient descent, incremental SVM).

[0289] (2) Automatically mine new knowledge points: Extract high-frequency new questions (such as new product consultation, policy change-related questions) from historical dialogues, and automatically import them into the knowledge base after manual review.

[0290] (3) Reinforcement learning guided strategy optimization: Reward mechanism design, converting "user conversation time reduction", "human access rate reduction", "problem solving rate improvement" and other indicators into reward values, through reinforcement learning (such as Q-Learning, deep reinforcement learning), let the system try and error in the interaction with the user, gradually optimize the answer strategy.

[0291] (4) Transfer learning and pre-training model fine-tuning: Pre-training model foundation, initialize intelligent customer service model based on general large model (such as GPT, BERT), use its existing language understanding ability; domain fine-tuning, fine-tune the model with specific industry data (such as financial, e-commerce professional terms and scenarios), make it adapt to vertical field demand; continuous fine-tuning, fine-tune the model with new industry data generated regularly, maintain the understanding of new trends (such as emerging products, policies).

[0292] Based on the large model developed intelligent customer service, the above-mentioned alternatives to achieve the self-learning function have some limitations and deficiencies. According to the analysis and result evaluation in the above chapters, this application uses the innovative method of automatic control negative feedback principle to realize the self-learning function of intelligent customer service system with good performance.

[0293] The application also provides an intelligent customer service self-learning system, comprising a processor and a memory storing a computer program, and the processor implements the steps of any of the above methods when executing the computer program.

[0294] It can be understood that the above embodiments only express the preferred embodiments of the present application, which are described in detail and specifically, but cannot be understood as a limitation on the patent scope of the present application; it should be pointed out that for ordinary skilled in the art, the above technical features can be freely combined without departing from the concept of the present application, and several modifications and improvements can be made, which all belong to the protection scope of the present application; therefore, any equivalent transformation and modification within the scope of the claims of the present application shall belong to the scope of the claims of the present application.

Claims

1. An intelligent customer service self-learning method, characterized in that, The method comprises the following steps: Step S1, collecting user interaction data with intelligent customer service to form a multidimensional data set; Step S2, based on the PID controller, processing the performance indicators in the multidimensional data set, calculating the current error signal, and generating the corresponding control instruction according to the current error signal to adjust the response behavior of the intelligent customer service system in real time; Step S3, evaluating the system performance data adjusted by the control instruction to obtain evaluation feedback; Step S4, according to the evaluation feedback, dynamically adjusting the parameters of the PID controller through an adaptive control strategy, and updating the parameters of the PID controller.

2. The intelligent customer service self-learning method of claim 1, wherein, The interaction data includes interaction process data, performance indicator data, and environment and state data, and the step S1 comprises: Obtain the interaction process data, performance indicator data and environment and state data, and perform cleaning and structured processing to form a standardized multidimensional data set.

3. The intelligent customer service self-learning method of claim 1, wherein, The performance indicators include: answer accuracy, response time and user satisfaction; in the step S2, the PID controller is used to process the performance indicators in the multidimensional data set to calculate the current error signal, which comprises: Obtain the preset target accuracy, preset target response time and preset target user satisfaction; Calculate the accuracy deviation according to the answer accuracy and the preset target accuracy; Calculate the response time deviation according to the response time and the preset target response time; Calculate the user satisfaction deviation according to the user satisfaction and the preset target user satisfaction; Weight and fuse each deviation value to generate a comprehensive current error signal as the input of the PID controller.

4. The intelligent customer service self-learning method of claim 1, wherein, In the step S2, the corresponding control instruction generated according to the current error signal comprises: Input the current error signal into the PID controller to perform proportional operation, integral operation and differential operation respectively to obtain proportional adjustment component, integral correction component and differential prediction component; Superimpose and fuse the proportional adjustment component, integral correction component and differential prediction component to generate a total control signal; Generate a system control instruction according to the total control signal; The system control instruction is sent to the execution unit of the intelligent customer service system.

5. The intelligent customer service self-learning method of claim 1, wherein, The step S3 comprises: Based on Lyapunov stability theory, construct a system energy function according to the system performance data, and calculate the rate of change of the energy function over time in real time; Determine the system stability according to the rate of change to generate a stability analysis result; Calculate the system performance indicators based on the system performance data; Integrate the system performance indicators and the stability analysis result to generate the evaluation feedback report.

6. The intelligent customer service self-learning method of claim 5, wherein, In the step S4, the dynamic parameter adjustment through the adaptive control strategy comprises: According to the stability analysis result in the evaluation feedback report, adjust the proportional coefficient, integral coefficient or differential coefficient of the PID controller, and feed back the adjusted parameters to the PID controller in step S2 to make the system energy function decrease over time.

7. The intelligent customer service self-learning method of claim 1, wherein, The dynamic parameter adjustment through the adaptive control strategy further comprises: The dynamic model parameters of the intelligent customer service system are estimated in real time and on line through a self-correcting control strategy. The dynamic model parameters are used to represent changes in the mapping relationship between system inputs and control outputs, the system inputs include knowledge base matching weights or model generation parameters, and the control outputs include response accuracy or response speed.

8. The intelligent customer service self-learning method of claim 7, wherein, The dynamic parameter adjustment through the adaptive control strategy further includes: Based on the estimated changes in the dynamic model parameters, the degree of drift of system characteristics is determined; According to the degree of drift, the gain parameters of the PID controller are dynamically adjusted to make the performance indicators of the intelligent customer service system track the preset optimal performance target; The gain parameters include one or more combinations of proportional coefficient, integral coefficient and differential coefficient. 9.The intelligent customer service self-learning method of claim 7, wherein, The dynamic parameter adjustment through the adaptive control strategy further includes: A model predictive control strategy is used to predict the system performance change trend in the future period of time based on the current system state and the dynamic model parameters; The long-term performance indicators are taken as the optimal target for rolling optimization of the control sequence in the future limited time domain, and the optimal control action at the current time is executed; The long-term performance indicators include user satisfaction cumulative value or monthly problem solving rate.

10. An intelligent customer service self-learning system comprising a processor and a memory having stored therein a computer program, characterized in that, The processor implements the steps of the method of any one of claims 1-9 when executing the computer program.

Images