Coffee machine control error suppression method and system based on intent recognition model

Abstract

The application discloses a coffee machine control error suppression method and system based on an intention recognition model, and belongs to the field of intelligent control. The method comprises the following steps: deploying an operation verification threshold at the instruction response end of a target coffee machine; the prior of the operation verification threshold comprises the following steps: according to a false touch verification layer, performing denoising fitting distribution verification on a device interaction signal to obtain a first verification result representing false touch or non-false touch; if it is non-false touch, performing intention probability distribution analysis according to an ambiguity verification layer to output a device control instruction or an active inquiry pop-up window; encoding an error compensation expression in a symbolic regression mode, converting the device control instruction into a parameter control sequence, and performing error correction to obtain a target parameter control sequence; and performing multi-thread driving management according to the target parameter control sequence. The application solves the technical problem that the existing coffee machine control input end is prone to control deviation at the instruction level due to false touch, environmental interference and ambiguous instructions, resulting in insufficient extraction precision and poor user experience.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and specifically to a method and system for suppressing control errors in coffee machines based on an intent recognition model. Background Technology

[0002] As fully automatic coffee machines evolve towards intelligence and multimodal interaction, diverse interaction methods such as touch and voice control have been widely applied in various home and commercial coffee machines. However, the control error management of existing coffee machines mostly focuses on simple correction of a single parameter at the execution end, and has not yet built a systematic error suppression mechanism covering the entire chain from user interaction command input to end execution output. In practical applications, the interaction input end is prone to invalid or skewed control commands due to accidental touches, environmental electromagnetic interference, and ambiguity in the intent of vague commands. The execution end of the equipment is affected by factors such as ambient temperature fluctuations, unstable power grid voltage, individual hardware differences, and long-term wear. In addition, the timing matching and parameter decoupling accuracy of the coordinated drive of multiple components such as heating, grinding, and extraction are insufficient, which easily leads to parameter execution deviations. The deviations at the command end and the execution end are superimposed, ultimately resulting in insufficient accuracy of coffee extraction parameters, poor stability and consistency of finished product quality, and also reducing the smoothness of human-computer interaction and user experience. Summary of the Invention

[0003] This application provides a coffee machine control error suppression method and system based on an intent recognition model. It aims to solve the technical problem that existing coffee machine control input terminals are prone to control deviations at the instruction level due to accidental touches, environmental interference, and ambiguous commands, resulting in insufficient extraction accuracy and poor user experience. The goal is to improve the execution accuracy of coffee machine control parameters and the stability and consistency of coffee product quality, thereby enhancing the user experience.

[0004] In view of the above problems, this application provides a method and system for suppressing control errors of coffee machines based on an intent recognition model.

[0005] The first aspect disclosed in this application provides a method for suppressing control errors in a coffee machine based on an intent recognition model, the method comprising: An operation verification threshold is deployed at the command response end of the target coffee machine. Device interaction signals are collected in real time and imported into the operation verification threshold for prior analysis, generating device control commands carrying valid interaction signals. The prior analysis based on the operation verification threshold includes: performing a denoising fitting distribution verification based on inverse denoising on the device interaction signals according to the mis-touch verification layer to obtain a first verification result representing mis-touch or non-mis-touch; if it is a non-mis-touch, performing an intent probability distribution analysis based on the maximum entropy principle according to the ambiguity verification layer, outputting a device control command or an active query pop-up; encoding an error compensation expression using symbolic regression, converting the device control command into a parameter control sequence and performing error correction to obtain a target parameter control sequence; and performing multi-threaded drive management of the end effector based on the target parameter control sequence.

[0006] Another aspect of this application discloses a coffee machine control error suppression system based on an intent recognition model, the system comprising: The control module is used to deploy operation verification thresholds at the instruction response end of the target coffee machine. It collects device interaction signals in real time and imports these operation verification thresholds for prior knowledge, generating device control instructions carrying valid interaction signals. The prior knowledge module, based on the operation verification thresholds, includes: performing a denoising fitting distribution verification based on inverse denoising on the device interaction signals according to the erroneous touch verification layer to obtain a first verification result representing erroneous or non-erroneous touch; the analysis module, if the touch is non-erroneous, performing intent probability distribution analysis based on the maximum entropy principle according to the ambiguity verification layer, outputting device control instructions or an active query pop-up; the correction module, encoding an error compensation expression using symbolic regression, converting the device control instructions into a parameter control sequence and performing error correction to obtain a target parameter control sequence; and the management module, performing multi-threaded drive management of the end effector according to the target parameter control sequence.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: By deploying operation verification thresholds including a mis-touch verification layer and an ambiguity verification layer at the command response end of the target coffee machine, invalid signals of mis-touch are filtered out based on the denoising fitting distribution verification of inverse denoising, and the intention probability distribution is solved based on the maximum entropy principle to resolve command ambiguity, generating precise device control commands carrying valid interaction signals. Then, the error compensation expression adapted to the device characteristics is encoded using symbolic regression, and the device control commands are converted into a parameter control sequence and the pre-error correction is completed to obtain the target parameter control sequence. Finally, the end effector is subjected to multi-threaded decoupled drive and real-time feedback closed-loop adjustment based on the target parameter control sequence. The entire link from command input verification to execution output control systematically suppresses control errors, effectively improving the execution accuracy of coffee machine control parameters and the stability and consistency of coffee product quality, thus enhancing the user experience.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a coffee machine control error suppression method based on an intent recognition model is provided for embodiments of this application. Figure 2 A schematic diagram of the structure of a coffee machine control error suppression system based on an intent recognition model is provided for embodiments of this application; Figure 3 A schematic diagram of control parameter comparison curves for a coffee machine control error suppression method based on an intent recognition model is provided for embodiments of this application.

[0010] Explanation of reference numerals in the attached diagram: Control module 11, Prior module 12, Analysis module 13, Correction module 14, Management module 15. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] The overall concept of the technical solution provided in this application is as follows: This application provides a coffee machine control error suppression method and system based on an intent recognition model. By deploying an operation verification threshold including a mis-touch verification layer and an ambiguity verification layer at the instruction response end, and performing two-level pre-screening and ambiguity resolution of interactive signals based on inverse denoising fitting distribution verification and maximum entropy principle intent analysis, the system avoids control deviations caused by invalid and ambiguous instructions from the source. At the same time, a nonlinear error compensation expression suitable for the hardware characteristics of the device is fitted using a symbolic regression method to perform pre-error correction on the generated parameter control sequence. This is supplemented by multi-threaded decoupled drive control and real-time feedback closed-loop adjustment at the end, achieving systematic suppression of control errors throughout the entire chain from instruction input to execution output. This improves the control accuracy of the coffee machine's extraction parameters and the stability of the coffee product quality, and optimizes the user experience.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0014] Example 1, as Figure 1As shown in the embodiments of this application, a method for suppressing control errors in a coffee machine based on an intent recognition model is provided. This method includes: S100: Deploy an operation verification threshold at the instruction response end of the target coffee machine, collect device interaction signals in real time and import the operation verification threshold for prior knowledge, and generate a device control instruction carrying a valid interaction signal.

[0015] Specifically, the operation verification threshold is first deployed at the command response end of the target coffee machine. The operation verification threshold is a pre-verification unit embedded between the user interaction input and the device control execution link, which undertakes the function of command admission screening. Its deployment process is based on normal operation data to complete model training and encapsulation. First, normal interaction data under multiple modalities such as touch screen, voice, and physical buttons are collected from the target coffee machine to build a positive sample set. Then, based on the training logic of the diffusion model, random noise is gradually superimposed on the positive samples to simulate interference signals through forward diffusion. Then, the original normal signal is restored through the reverse denoising process. The forward diffusion refers to the progressive addition of Gaussian noise to the standard interaction data to gradually blur the original signal features to simulate abnormal input patterns caused by accidental touch and environmental interference. The reverse denoising is to gradually remove noise from the noisy signal and restore a clean signal that conforms to the normal operation rules. The entire training process aims to fit the high-dimensional signal distribution of normal operation interaction. Finally, the trained verification logic is encapsulated into a two-level verification structure containing a mis-touch verification layer and an ambiguity verification layer, and then deployed on the device's interaction response port.

[0016] After deployment, the system enters a real-time priori operation process, continuously collecting user-triggered device interaction signals and importing them into the operation verification threshold in real time to conduct two-level verification. The signal first enters the mis-touch verification layer to perform denoising fitting distribution verification based on inverse denoising. This layer performs inverse denoising calculation on the input interaction signal and outputs the corresponding negative log-likelihood value. The negative log-likelihood value is the core indicator for quantifying the degree of matching between the current signal and the distribution of normal operation signals. When the input signal belongs to normal operation, the inverse denoising restoration degree is high, and the corresponding negative log-likelihood value is low. If the input is an abnormal signal such as mis-touch or random interference, the denoising restoration effect is poor because it cannot match the distribution pattern of normal operation, and the negative log-likelihood value will increase significantly. Subsequently, the system quantifies the negative log-likelihood value into a directly comparable energy score and judges it against the preset first energy threshold. If the energy score is greater than the first energy threshold, it is judged as a mis-touch, and the interaction response end directly intercepts the signal without triggering any control response. If the energy score is less than or equal to the first energy threshold, it is judged as a non-mis-touch, and the signal is marked as a valid interaction signal and transferred to the ambiguity verification layer for the second layer of verification.

[0017] After entering the ambiguity verification layer, the system retrieves the corresponding user's historical interaction data as constraints. Based on the maximum entropy principle, it calculates the conditional entropy by solving the intent probability distribution of the valid interaction signal. The maximum entropy principle is the core criterion followed in intent probability modeling. It refers to selecting the most unbiased probability model with the most reasonable uncertainty distribution to describe the current user's intent under the constraint of known user's historical operating habits, avoiding misjudgment of intent caused by subjective assumptions. Conditional entropy is a quantitative value of the uncertainty of the intent of the current interaction signal. The higher the entropy value, the higher the ambiguity of the system's judgment of the user's true intent. For example, the entropy value is high when a fuzzy voice command matches multiple coffee-making needs at the same time, while the entropy value of a command that conforms to the user's daily fixed operating habits is at a lower level. After completing the solution, the system compares the conditional entropy with a preset second entropy threshold. If the conditional entropy is less than or equal to the second entropy threshold, the system determines that the valid interaction signal has no intent ambiguity and directly outputs the corresponding device control command. If the conditional entropy is greater than the second entropy threshold, the system determines that the signal has intent ambiguity and generates an active inquiry pop-up window with maximum information gain to confirm the specific needs of the user. After the user provides clear feedback, the corresponding device control command is generated.

[0018] S200: The prior based on the operation verification threshold includes: performing a denoising fitting distribution verification based on inverse denoising on the device interaction signal according to the accidental touch verification layer, to obtain a first verification result characterizing accidental touch or non-accidental touch.

[0019] Specifically, this step serves as the first-level judgment process in the two-level prior verification of the operation verification threshold. The false touch verification layer is a filtering unit deployed at the very front of the interaction signal input link, responsible for intercepting invalid signals of false triggering without operational intent. When the real-time collected device interaction signal flows to this layer, a denoising fitting distribution verification based on inverse denoising will be performed. The denoising fitting distribution verification is the core logic of this false touch judgment, that is, it does not make judgments based on fixed signal feature rules, but rather judges whether it matches the high-dimensional distribution of normal operation signals fitted in the training phase by the degree of restoration of the signal after denoising processing. The entire verification process relies on inverse denoising calculation, which is the reverse restoration process corresponding to the forward diffusion and noise addition in the threshold deployment phase, referring to the process of gradually removing noise from the original input interaction signal. The process involves removing random interference components and attempting to restore the signal to a standard signal that conforms to normal operating characteristics. After the calculation is completed, the corresponding negative log-likelihood value is output. The negative log-likelihood value is the core indicator for quantifying the degree of matching between the current input signal and the normal operating distribution. Its value is negatively correlated with the signal denoising and restoration effect: if the current signal is a normal operating signal initiated by the user, its characteristics are highly consistent with the high-dimensional distribution of normal operation, the restoration accuracy of inverse denoising is high and the restoration loss is small, and the final calculated negative log-likelihood value is at a low level; if the current signal is an abnormal signal caused by accidental touch, environmental electromagnetic interference, or touch jitter, it cannot match the distribution pattern of normal operation, the inverse denoising process is difficult to complete the effective restoration, the signal loss is large, and the corresponding negative log-likelihood value will be significantly increased.

[0020] The system then standardizes and quantizes the negative log-likelihood value to obtain an energy score that can be directly used for threshold comparison. The energy score is a quantified index after converting the negative log-likelihood value to a unified judgment scale. The higher the value, the stronger the deviation of the current signal from normal operation characteristics and the higher the probability of false touch. After quantization, the system compares the energy score with a preset first energy threshold. The first energy threshold is a critical value for false touch judgment set based on the statistical distribution of a large number of normal operation samples. If the energy score is greater than the first energy threshold, it means that the current signal deviates from the normal operation distribution to an unacceptable extent, and is judged as a false touch. A first verification result representing a false touch is generated, and the interactive response end directly intercepts the signal without triggering any subsequent control response. If the energy score is less than or equal to the first energy threshold, it means that the signal characteristics conform to the distribution law of normal operation, and is judged as a non-false touch. A first verification result representing a non-false touch is generated, and the signal is allowed to enter the subsequent ambiguity verification layer for the second layer of verification.

[0021] S300: If it is not a mis-touch, according to the ambiguity verification layer, perform the intent probability distribution analysis based on the maximum entropy principle, and output device control commands or proactive query pop-ups.

[0022] Specifically, after the accidental touch verification layer outputs the first verification result representing a non-accidental touch, the corresponding device interaction signal will be marked as a valid interaction signal and transferred to the ambiguity verification layer. This layer is the secondary verification unit in the operation verification threshold responsible for resolving the ambiguity of the instruction intent. It completes the ambiguity judgment through the intent probability distribution analysis based on the maximum entropy principle. The maximum entropy principle is the criterion followed in this intent probability modeling. It refers to selecting the probability distribution with the largest entropy value and the most unbiasedness under known constraints to describe the user's unknown intent, without introducing additional subjective preset judgment preferences, thereby ensuring the objectivity and comprehensiveness of intent recognition and avoiding intention misjudgment caused by human preset operation weights.

[0023] During the analysis, the system first retrieves the target user's historical interaction data as constraints. These constraints cover statistical characteristics such as the user's past operation category preferences, usage time patterns, and operation frequency distribution, serving as prior information for calculating the intent probability. For example, if over 90% of the user's historical brewing instructions point to Americano coffee, this statistical result will be used as a boundary constraint for the probability distribution calculation. Subsequently, the system inputs the characteristics of the effective interaction signal along with the aforementioned constraints into the solution model. Based on the maximum entropy principle, iteratively calculates the intent probability distribution of all executable coffee-making operations corresponding to the current signal, and further derives the conditional entropy. Conditional entropy is a quantitative indicator that measures the uncertainty of the intent of the current interaction signal. Its value is positively correlated with the ambiguity of the system's judgment of the user's true intent. A higher entropy value indicates a more even probability distribution of each alternative operation and stronger instruction ambiguity, while a lower entropy value indicates a more concentrated intent and higher judgment confidence.

[0024] After solving the problem, the system compares the calculated conditional entropy with a preset second entropy threshold. The second entropy threshold is a critical value for ambiguity judgment that balances control precision and interactive experience. If the conditional entropy is less than or equal to the second entropy threshold, it means that the intent of the current effective interaction signal is clear, the confidence level of intent judgment meets the standard under the constraint of the user's historical habits, and there is no significant ambiguity. At this time, the system directly parses the signal and generates a device control command that matches the intent and outputs it downstream. If the conditional entropy is greater than the second entropy threshold, it means that the ambiguity of the current signal's intent exceeds the range that can be directly executed, and there are multiple possible operations with similar probabilities. For example, if an ambiguous voice command matches multiple coffee categories at the same time and the probability difference is small, the system will generate an active inquiry pop-up that maximizes information gain. Maximizing information gain means that the content and options of the pop-up will prioritize the direction that can quickly narrow the scope of intent and reduce uncertainty, rather than listing all possible operations. This reduces the user's operation cost while resolving ambiguity. After the user gives clear feedback, the corresponding device control command is generated based on the confirmed clear intent.

[0025] S400: The error compensation expression is encoded using a symbolic regression method. The target parameter control sequence is obtained by converting the equipment control command into a parameter control sequence and correcting the error.

[0026] Specifically, the semantic functional instructions are first decomposed and converted into a structured parameter control sequence. The parameter control sequence maps the user's coffee making needs into a set of multi-dimensional control parameters that the equipment can directly parse. These parameters include the set values ​​of core execution dimensions such as heating power, grinder motor duty cycle, water pump duty cycle, and extraction time. This serves as a standardized input carrier for error correction.

[0027] The system will then encode and construct the error compensation expression based on symbolic regression. Symbolic regression is an adaptive function fitting method that does not require manual pre-setting of function forms. It uses basic operators such as addition, subtraction, multiplication, division, exponentiation, and summation, as well as input variables and constants, as the smallest building blocks. It automatically searches for the optimal mathematical expression that matches the data pattern through iterative evolution.

[0028] The system first collects parameter setpoint data and actual execution value data from sensors corresponding to various control commands during historical operation phases to form a basic dataset for error fitting. Simultaneously, it determines the equipment parameter control matrix, including heating power, grinder motor duty cycle, water pump duty cycle, and extraction time, as the input variable set. The error between the parameter setpoint and the actual execution value is used as the fitting output target, and the convergence of the error between the setpoint and the execution value is used as a constraint. This involves continuously optimizing the structure and internal parameters of the expression to maximize the compensation amount calculated by the expression to offset the deviation between the actual execution result and the theoretical setpoint. In the initial stage, the system randomly generates several sets of candidate expressions. Through multiple rounds of iterative verification, screening, and structural variation, the difference between the calculated expression result and the actual error is gradually reduced until the error converges to a preset acceptable range. Finally, an error compensation expression adapted to the current equipment hardware characteristics and operating environment is fitted.

[0029] After the expression encoding is completed, the system uses each control parameter in the original parameter control sequence as input variables and substitutes them into the error compensation expression to calculate the pre-error compensation amount corresponding to each control parameter. This compensation amount can offset the expected execution deviation caused by internal and external factors such as ambient temperature fluctuations, power grid voltage fluctuations, and uneven soybean density. Then, the original parameter values ​​and the corresponding compensation amounts are superimposed and corrected to complete the pre-compensation adjustment of all-dimensional control parameters, and finally obtain the target parameter control sequence after pre-error correction.

[0030] S500: Perform multi-threaded drive management on the end effector according to the target parameter control sequence.

[0031] Specifically, for the multi-threaded drive management of end effectors, the multi-threaded drive management adopts a parallel and independent control thread to manage the operation scheduling mode of different end effector components. It can achieve precise timing coordination and independent parameter control of multiple components. The execution process first targets the three core end effectors: heating tube, grinding motor, and water pump. The target parameter control sequence, which integrates full-dimensional control parameters such as heating power, grinding duty cycle, water pump duty cycle, and extraction time, is decoupled. Decoupling refers to the process of separating and separating the control parameters and timing information in the target parameter control sequence according to the function of the actuator. The originally integrated parameter control set is decomposed into independent control parameter packages corresponding to individual actuators. This generates multi-threaded drive instructions for each end effector. Each thread independently carries the full-cycle control logic of an execution component. The threads do not occupy computing resources or interfere with each other's parameter output, ensuring the accuracy of the control parameters of each component.

[0032] After instruction decomposition is completed, the system will execute spatiotemporal code constraints on multi-threaded drive instructions. Spatiotemporal code constraints are a rule mechanism for coordinating and calibrating multi-threaded instructions from two dimensions: time and space. In the time dimension, the start node, runtime, stop node, and linkage trigger conditions of each executor are uniformly calibrated. For example, the connection sequence between the completion node of the grinding process and the start node of the water pump extraction is precisely calibrated to avoid production process errors caused by process misalignment. In the space dimension, the matching degree between the parameters of each thread and the rated operating range of the corresponding executor is verified to prevent execution deviations and hardware damage caused by parameter out-of-bounds.

[0033] Finally, the drive commands, after being calibrated by the time-space code, will be converted into pulse width modulation (PWM) signals. PWM signals are standard drive signals that precisely adjust the duty cycle of the high level within a unit cycle to achieve precise regulation of power, speed, and flow rate. They can accurately convert digital control parameters into the actual heating power of the heating element, the actual operating speed of the grinder motor, and the actual water flow rate of the water pump. Ultimately, the system sends the PWM signals of the corresponding channels to the heating element, grinder motor, and water pump, driving each component to operate in parallel according to preset parameters and timing, completing the entire coffee-making process.

[0034] Furthermore, in the method provided in this application embodiment, deploying an operation verification threshold on the instruction response end of the target coffee machine includes: collecting multimodal interaction data under normal operating conditions through the target coffee machine as a positive sample set; based on the positive sample set, using a forward diffusion process to gradually add noise and reverse denoising as a training process, and using the high-dimensional distribution of normal operation interaction as the fitting target, deploying the operation verification threshold on the interaction response end; wherein, the operation verification threshold includes a mis-touch verification layer and an ambiguity verification layer, and the ambiguity verification layer is deployed using a judgment rule based on the maximum entropy principle.

[0035] Specifically, the target coffee machine first collects multimodal interaction data under normal operating conditions through its integrated touch sensing unit, voice acquisition unit, and physical button detection unit. This multimodal interaction data includes touch point coordinate sequences, press duration, and sliding acceleration of touch screen operations; voiceprint features, semantic segmentation features, and speech rate features of voice commands; and original interaction signals from multiple channels and dimensions, such as the pressure, trigger duration, and press interval of physical buttons. This covers daily usage scenarios for user groups with different operating habits and different ambient lighting and electromagnetic interference baselines. All valid normal operating signals confirmed by manual annotation are integrated into a positive sample set, which serves as the sole benchmark data for the model to learn the normal operating feature patterns, eliminating the need to collect additional negative samples of accidental touches.

[0036] Next, model training is performed, using a diffusion model as the network structure. The diffusion model is a generative model that learns the true distribution of data by progressively adding noise and reverse denoising. Its training process is divided into two completely symmetrical stages: forward diffusion and reverse denoising. The entire training aims to fit the high-dimensional distribution of normal operation interaction. The high-dimensional distribution refers to the joint statistical distribution law formed by all feature dimensions of normal operation signal, such as touch point coordinates, pressing time, and spectral characteristics. It is the overall feature boundary formed by the linkage of multi-dimensional features. In the forward diffusion stage, hundreds to thousands of time steps are set according to the preset noise scheduling strategy. Random noise that conforms to the Gaussian distribution is gradually superimposed on each group of original clean interaction signals in the positive sample set. Only a small amount of noise of the corresponding proportion is injected in each time step.

[0037] After continuous superposition across the entire step length, the original clear normal operation signal will gradually degenerate into a featureless signal that approximates random noise. This process essentially simulates the signal distortion patterns under abnormal scenarios such as accidental touches, touch jitter, and environmental electromagnetic interference, allowing the model to fully learn the mathematical logic of the evolution from normal signals to interference signals. The reverse denoising stage is the core training process of the model. The noisy signals generated at each time step of forward diffusion are used as network inputs, and the real noise components at the corresponding time steps are used as training labels. The network is trained to predict the noise components contained in the current noisy signal, and then the network parameters are iteratively optimized through backpropagation. The loss function adopts negative log-likelihood loss to constrain the deviation between the network's predicted noise and the real noise. After multiple rounds of full-sample iterative training, the network will have the ability to remove noise from any noisy interaction signal and restore the normal operation signal. At the same time, it will accurately grasp the high-dimensional distribution characteristics of the normal operation signal and be able to judge the matching degree between the input signal and the normal distribution by the difficulty of denoising and restoration.

[0038] After model training, a hierarchical structure for operation verification thresholds is encapsulated and rules are deployed. The operation verification thresholds adopt a two-level serial structure with progressive filtering. The first layer is the false touch verification layer. During encapsulation, the trained inverse denoising inference module, negative log-likelihood calculation module, energy score quantization module, and first energy threshold determination module are integrated into an independent verification unit. At the same time, a first energy threshold is preset as the false touch judgment threshold based on the denoising statistics of the positive sample set. This layer can directly perform denoising fit calculation and binary judgment of false touch / non-false touch on the input signal, undertaking the coarse screening function of the first layer. The second layer is the ambiguity verification layer, which adopts a maximum entropy principle. The judgment rules are deployed, with the maximum entropy principle being the guiding principle for intent probability modeling. This principle refers to selecting the probability distribution with the maximum entropy value and the most unbiased probability distribution to describe the unknown intent within the boundaries of known constraints, without introducing any subjective pre-set judgment preferences. During deployment, a set of instruction labels for all executable operations of the coffee machine is first preset, and an intent probability distribution solution module and a conditional entropy calculation module are built. At the same time, a constraint access interface for user historical interaction data is reserved, which can retrieve data such as user's past operation preferences and usage time patterns in real time as constraints for probability solution. Then, a second entropy threshold is preset as the ambiguity judgment threshold, and an active query generation logic that maximizes information gain is built accordingly.

[0039] Finally, the entire set of calculation and judgment logic is encapsulated into an independent ambiguity verification unit. Then, the accidental touch verification layer and the ambiguity verification layer are connected in series and integrated into a complete operation verification threshold in the order of accidental touch screening first and ambiguity resolution. This threshold is then fixed and deployed on the interactive response end of the coffee machine, that is, on the necessary link node between the user interaction input module and the device control command generation module, to ensure that all external input interaction signals must pass through the two-level verification of this threshold before entering the subsequent control process.

[0040] Furthermore, in the method provided in this application embodiment, according to the accidental touch verification layer, a denoising fitting distribution verification based on inverse denoising is performed on the device interaction signal, including: by real-time acquisition of the device interaction signal and importing it into the operation verification threshold, the accidental touch verification layer calculates a negative log-likelihood value based on the inverse denoising process of the device interaction signal, wherein the negative log-likelihood value characterizes the denoising fitting distribution of the device interaction signal; the negative log-likelihood value is quantized to obtain an energy score; if the energy score is greater than a first energy threshold, it is determined to be an accidental touch, and the interaction response terminal does not perform a control response.

[0041] Specifically, the accidental touch verification layer performs a denoising fitting distribution verification based on real-time collected device interaction signals. This denoising fitting distribution verification is a verification method that does not rely on preset accidental touch feature rules and only calculates the degree of matching between the input signal and the distribution of normal operation signals to determine accidental touches. It is achieved based on the reverse inference capability of the trained diffusion model, proceeding in three consecutive stages: reverse denoising calculation, quantization mapping, and threshold determination. First, the reverse denoising process is performed. The system uses the real-time collected raw device interaction signals (touchscreen, voice, buttons, etc.) as input samples x and imports them into the pre-trained diffusion model of the accidental touch verification layer for reverse inference. The network performs a complete multi-step inverse denoising inference. This inverse denoising is a signal restoration process completely symmetrical to the forward diffusion during model training. Forward diffusion progressively adds Gaussian noise to normal samples to simulate interference signal patterns, while inverse denoising starts with the input sample and, through a converged network trained over time, predicts and gradually removes the noise components contained in the signal, attempting to restore a clean signal conforming to the normal operating distribution. During this progressive denoising process, the model calculates the negative log-likelihood value of the corresponding input signal based on variational inference. The negative log-likelihood value is a core quantitative indicator characterizing the denoising fit distribution of device interaction signals, and its mathematical definition is: ; In the formula For parameters The diffusion model fits the high-dimensional distribution of the normal interaction signal for the input sample. The probability of occurrence is calculated by decomposing the variational lower bound of the forward diffusion process, transforming the negative log-likelihood, which is difficult to solve directly, into the cumulative sum of denoised prediction errors over T time steps, i.e.: ; in Let be the mean square error between the model's predicted noise and the actual noise component at time step t. If the input signal is a normal operation signal initiated by the user, its characteristics are highly consistent with the normal distribution learned by the model. The prediction error of progressive denoising is small, and the final accumulated negative log-likelihood value is low, which means that the signal has a high degree of fit with the normal distribution.

[0042] If the input signal is an abnormal signal caused by accidental touch, environmental electromagnetic interference, or touch jitter, it cannot match the distribution pattern of normal operation. The prediction error of the denoising process remains high, and the accumulated negative log-likelihood value increases significantly, representing a poor fit between the signal and the normal distribution. This achieves a quantitative characterization of the signal's normality. After calculating the negative log-likelihood value, the system enters the quantization mapping stage. To eliminate the numerical scale differences of different modal interaction signals and establish a unified judgment standard, the system quantizes the calculated negative log-likelihood value into an energy score through linear normalization. The quantization process uses the negative log-likelihood statistics of all samples in the positive sample set as a benchmark, first statistically determining the minimum value of the negative log-likelihood in the positive sample set: With the maximum value Then, using the formula: ; The original negative log-likelihood value is mapped to a unified numerical range of 0-100 to obtain a standardized energy score. The energy score is an indicator that intuitively represents the degree to which the input signal deviates from the normal operating distribution. The higher the score, the stronger the signal abnormality and the higher the probability of false touch. Finally, in the threshold determination stage, the system compares the calculated energy score with a preset first energy threshold. The first energy threshold is a critical value for false touch determination based on the statistical distribution of the positive sample set. It is usually taken as the 95th percentile value of the positive sample energy score to cover most normal operating scenarios. If the energy score is greater than the first energy threshold, it means that the current signal deviates from the normal operating distribution beyond the acceptable range and is judged as a false touch. A first verification result representing the false touch is generated, and the interactive response end directly intercepts the signal without triggering any subsequent control response. If the energy score is less than or equal to the first energy threshold, it is judged as a non-false touch, and the signal is marked as a valid interactive signal and transferred to the ambiguity verification layer for the second layer of verification.

[0043] Furthermore, in the method provided in this application embodiment, if it is determined that it is not a mis-touch, the interaction response terminal triggers a control response and uses the device interaction signal as a valid interaction signal; retrieves the target user's historical interaction data as a constraint condition, and transfers the valid interaction signal and the constraint condition to the ambiguity verification layer to output a device control command.

[0044] Specifically, once the accidental touch verification layer completes the denoising and fitting distribution verification and outputs the first verification result representing non-accidental touches, the interactive response terminal, which is in standby interception state, immediately releases the signal interception mechanism and triggers the complete control response chain. At the same time, the device interaction signal that has passed the first layer verification is officially marked as a valid interaction signal. A valid interaction signal refers to an interactive input carrier that has a clear operation direction and can enter the subsequent intent parsing and instruction generation process after excluding invalid triggers such as accidental touches, environmental electromagnetic interference, and touch jitter without real operation intentions. It is the valid input basis for all control logic.

[0045] Next, the system will retrieve the historical interaction data of the target user from the locally stored user behavior database based on the user identity identifier associated with the current interaction signal. These constraints are personalized prior information used to limit the boundary of intent probability solution. They cover the user's long-term accumulated coffee category selection preferences, commonly used concentration and extraction amount parameter settings, operation patterns at different times, high-frequency operation paths, and other behavioral statistical features. Their purpose is to inject user-specific behavioral pattern references into the general intent recognition framework, narrow the range of options for intent judgment, and avoid recognition bias caused by indiscriminate general judgment.

[0046] After the constraint conditions are retrieved, the system extracts the feature vector from the valid interaction signal and transfers it together with the corresponding user's constraint conditions to the ambiguity verification layer. This layer performs intent probability distribution analysis based on the preset maximum entropy principle solution framework. That is, within the boundary of the user's historical behavior constraints, following the principle of not introducing additional subjective judgment preferences, it calculates the intent probability distribution of all executable coffee-making operations corresponding to the current valid interaction signal. Then, based on this probability distribution, it derives the conditional entropy, which represents the uncertainty of intent. Conditional entropy is the core indicator for quantifying the degree of ambiguity of intent of the current signal under user behavior constraints. The higher the entropy value, the more even the probability distribution of each alternative operation is, and the lower the system's confidence in judging the true intent.

[0047] The system then compares the calculated conditional entropy with a preset second entropy threshold. The second entropy threshold is a critical value for ambiguity determination that balances instruction execution efficiency and judgment accuracy. It is usually determined based on the statistical distribution of a large amount of user operation data. If the conditional entropy is less than or equal to the second entropy threshold, it means that the intent of the current signal is clear enough under the constraints of the user's personalized behavior, and it is determined that there is no intention ambiguity. The ambiguity verification layer directly parses the core requirements of the valid interaction signal and generates the corresponding device control instructions with the corresponding functions and parameters to output to the downstream control module. If the conditional entropy is greater than the second entropy threshold, it is determined that there is significant intention ambiguity in the current signal. The system will generate an active inquiry pop-up window that maximizes information gain to confirm the specific requirements with the user. After the user gives clear feedback, the corresponding device control instructions are generated based on the confirmed clear intent.

[0048] Furthermore, in the method provided in this application embodiment, according to the constraints, the conditional entropy based on the intent probability distribution is obtained by solving the effective interaction signal using the maximum entropy principle; if the conditional entropy is less than or equal to the second entropy threshold, the effective interaction signal is determined to be unambiguous, and the device control command is output; if the conditional entropy is greater than the second entropy threshold, the effective interaction signal is determined to be ambiguous, and an active query pop-up window with maximum information gain is generated.

[0049] Specifically, when constraints carrying user historical interaction data and valid interaction signals are input into the ambiguity verification layer, the system will perform intent probability distribution calculation and ambiguity determination based on the maximum entropy principle. The maximum entropy principle is the criterion followed in this intent probability modeling. Its core logic is to select the distribution with the largest entropy value among all probability distributions that meet the known constraints as the final judgment basis, without introducing any subjective assumptions or preferences into unknown information, thereby ensuring the objectivity and universality of intent determination and avoiding misjudgment of intent caused by human-preset operation weights.

[0050] Subsequently, the system first presets a set of candidate intentions covering all executable operations of the coffee machine. It defines operation options such as different types of coffee making, different strength levels, and different extraction volumes as all selectable values ​​of discrete intention random variables. At the same time, it transforms the retrieved historical interaction data of the target user into quantitative constraints that can be substituted into the model. That is, it extracts statistical features such as category selection preferences, time-based operation patterns, and common parameter setting habits from the user's historical operation logs and transforms them into several sets of expected constraints of feature functions. For example, the statistically verified behavioral patterns such as the probability that a user chooses Americano coffee on a weekday morning are 0.9 are used as boundary conditions that must be met to solve the probability distribution, ensuring that the final solution matches the user's personalized operation habits.

[0051] The system then enters the stage of solving for the optimal distribution of maximum entropy. The objective function of the system is to maximize the information entropy of the intention probability distribution, i.e., to maximize the expression: ; in Let be the probability of the i-th candidate intent occurring. Let the information entropy corresponding to this probability distribution be denoted by , and the expectation constraints of all feature functions be used as boundary conditions for the optimization problem. The constrained entropy maximization problem is transformed into an unconstrained optimization problem by constructing Lagrange multiplier functions. By iteratively updating the values ​​of the Lagrange multipliers, the optimal solution is gradually approximated. Finally, the intention probability distribution that satisfies all constraints and has the largest entropy value is derived. Its form is a normalized exponential distribution. The probability of occurrence of each candidate intention is determined by the ratio of the weighted exponential value of the corresponding feature to the sum of the weighted exponential values ​​of all candidate intentions.

[0052] After obtaining the optimal intent probability distribution, the system further calculates the conditional entropy corresponding to this distribution. Conditional entropy is a core quantitative indicator that measures the uncertainty of intent of the current effective interaction signal under the user's historical constraints. Its mathematical expression is: ; in The set of constraint features representing the user's historical interaction data. Representing the random variable of candidate intent, the value of conditional entropy is positively correlated with the ambiguity of the system's judgment of the user's true intent: the higher the entropy value, the more even the probability distribution of each candidate intent, and the stronger the ambiguity of the instruction. For example, when a vague voice instruction corresponds to a probability distribution of 40% for Americano, 30% for latte, and 30% for cappuccino, the conditional entropy is at a high level. The lower the entropy value, the more concentrated the intent, and the higher the confidence of the judgment. For example, when an instruction that conforms to the user's daily habits corresponds to a probability distribution of 95% for Americano and 5% for latte, the conditional entropy is at an extremely low level.

[0053] After calculating the conditional entropy, the system compares it with a preset second entropy threshold. The second entropy threshold is a critical value for ambiguity determination that balances control precision and interactive experience and is based on statistical calibration of massive user operation samples. If the conditional entropy is less than or equal to the second entropy threshold, it means that the intent of the current effective interaction signal is clear enough under the user's personalized constraints, and the signal is determined to be unambiguous. The system directly parses the corresponding intent parameter with the highest probability and generates device control instructions that match the user's needs to output to the downstream control module. If the conditional entropy is greater than the second entropy threshold, it means that the ambiguity of the current signal's intent exceeds the range that can be directly executed, and significant intent ambiguity is determined. The system will generate an active inquiry pop-up window that maximizes information gain. Maximizing information gain means that the system will calculate the entropy reduction brought by different question directions and option combinations one by one, select the question content that can reduce intent uncertainty the fastest and narrow the candidate range to the greatest extent, thereby reducing the user's operation cost while resolving ambiguity. After the user gives clear feedback, the corresponding device control instructions are generated based on the confirmed clear intent.

[0054] Furthermore, in the method provided in this application embodiment, error correction is performed to obtain the target parameter control sequence, including: according to the device control command, the control center receives the effective interaction signal and generates the parameter control sequence; using symbolic regression to encode the error compensation expression; and according to the error compensation expression, the parameter control sequence is corrected to obtain the target parameter control sequence.

[0055] Specifically, the first step is to generate the parameter control sequence. The control center analyzes the functional requirements, product selection, and parameter level information carried by the effective interaction signals, mapping the semantic coffee-making instructions into a set of structured parameters that can be directly parsed by the underlying execution components of the equipment—the parameter control sequence. The parameter control sequence uses the equipment parameter control matrix as its core framework, encompassing the initial settings for four core execution parameters: heating power, grinder motor duty cycle, water pump duty cycle, and extraction time. This serves as the benchmark input for subsequent error correction. For example, the initial parameter control sequence corresponding to a standard single-serve coffee extraction instruction can be denoted as: These correspond to the theoretical settings for heating power, grinder motor duty cycle, water pump duty cycle, and extraction time, respectively.

[0056] After generating the original parameter control sequence, the system will encode and construct the error compensation expression using symbolic regression. Symbolic regression is an adaptive function fitting method that does not require pre-defined function structures. It uses basic operators such as addition, subtraction, multiplication, division, exponentiation, and summation, along with input variables and constants, as the smallest building blocks. Through iterative evolution, it automatically searches for the optimal mathematical expression that matches the data patterns. It does not require pre-defining the structural framework of the error model and can accurately capture the nonlinear error characteristics of multi-parameter coupling in coffee machines. It adapts to complex error patterns caused by environmental fluctuations and individual hardware differences. The complete encoding process is divided into two stages: data preparation and iterative fitting. In the data preparation stage, the system first collects the parameter setpoints corresponding to various control commands during the historical operating cycle of the equipment, along with the actual execution values ​​returned by sensors such as temperature, flow rate, and speed, to construct paired training datasets. Simultaneously, it determines the equipment parameter control matrix, including heating power, grinder motor duty cycle, water pump duty cycle, and extraction time, as the set of input variables. ; The error value is the difference between the set value and the executed value. ; As the target for fitting output; in the iterative fitting stage, several sets of candidate error compensation expressions composed of freely combined primitives are first randomly generated as the initial population, with mean squared error: ; As a fitness evaluation index, in the formula The predicted error value (MSE) is calculated for the candidate expression. The smaller the MSE, the higher the accuracy of the expression in fitting historical error patterns. Subsequently, the structure and internal constant parameters of the expression are continuously optimized through an evolutionary iterative mechanism of selection, crossover, and mutation. In each round of iteration, several expressions with the best fitness are selected as parents. Crossover is performed by exchanging the subtree structure of the expressions, and mutation is performed by randomly replacing operators, variables, or constant values ​​to generate a new generation of candidate expressions. At the same time, the convergence of the error between the set value data and the executed value data is used as a constraint condition, that is, the deviation between the predicted error output of the expression and the actual error is required to continuously reduce until it is lower than the preset convergence threshold. After dozens to hundreds of rounds of iterative optimization, the final fit is an error compensation expression that is adapted to the current hardware characteristics and operating environment of the device. The system then stores the encoded data in the control center. After encoding the error compensation expression, the system performs error correction calculations on the control sequence, substituting each control parameter in the original control sequence as input variables into the corresponding dimension's error compensation expression to calculate the pre-compensation amount for each parameter. This compensation amount can offset the expected execution deviations caused by internal and external factors such as ambient temperature fluctuations, power grid voltage fluctuations, and uneven soybean density in the storage area, and then be corrected using the following formula: ; By superimposing the original parameter settings with the corresponding compensation values, pre-compensation adjustments are made to all dimensional parameters within the parameter control sequence one by one, ultimately yielding the target parameter control sequence after pre-error correction. .

[0057] Furthermore, the method provided in this application embodiment employs symbolic regression to encode the error compensation expression, including: collecting setpoint data and executed value data under historical control commands; determining the equipment parameter control matrix, using basic operators and input variables, constants, and input variables as primitives, converting the equipment parameter control matrix into input variables, outputting the error value, and using the setpoint data and executed value data for error convergence constraints to fit the error compensation expression; wherein, the equipment parameter control matrix includes heating power, grinder motor duty cycle, water pump duty cycle, and extraction time.

[0058] Specifically, the first step is to collect and construct the basic dataset. The system extracts setpoint and execution value data corresponding to all historical control commands from the equipment's historical operation logs. The setpoint data consists of theoretical control parameter values ​​issued by the control center to the execution components, while the execution value data consists of actual operating parameter values ​​transmitted back by end-user acquisition modules such as temperature sensors, flow meters, and speed sensors. These two types of data are paired one-to-one according to the control command sequence, covering operating samples under different ambient temperatures, grid voltages, and soybean storage densities, forming a training dataset containing thousands of paired samples. Simultaneously, the actual error value for each sample set is calculated, i.e., the difference between the setpoint and the corresponding execution value, denoted as: ; in Set the parameter values ​​for the i-th sample group. The corresponding actual execution value, this error value will be used as the target output of the symbolic regression fitting. Subsequently, the equipment parameter control matrix is ​​determined as the input system of the model. The equipment parameter control matrix is ​​a standardized parameter matrix formed by integrating the core control parameters of the coffee machine according to their dimensions, denoted as: ; in Represents heating power, Represents the duty cycle of the coffee grinder motor. Represents the pump duty cycle, Representing extraction time, these four parameters cover the control dimensions of the three core processes in coffee making: heating, grinding, and extraction. They are the core source of execution errors. After the matrix definition is completed, an input transformation operation is performed, which maps each parameter in the equipment parameter control matrix to the input independent variable of the symbolic regression model. This transforms the physical equipment control parameters into input variables that can be processed by the mathematical regression model, realizing the mapping transformation from the hardware control dimension to the mathematical fitting dimension. This ensures that the symbolic regression can directly solve for the corresponding error compensation based on the control parameters.

[0059] Next, expression fitting is initiated based on symbolic regression, an adaptive function fitting method that does not require pre-defined function structures. It uses basic operators, input variables, and constants as the smallest building blocks, and an evolutionary algorithm automatically searches for the optimal mathematical expression matching the data patterns. These building blocks specifically include basic operators such as addition, subtraction, multiplication, division, and exponentiation; four input variables derived from the device parameter control matrix; and dynamically adjustable constant terms. All building blocks can be freely combined in a tree structure to form candidate expressions of arbitrary structures, accurately capturing the nonlinear error characteristics caused by the multi-parameter coupling of the coffee machine. The fitting process uses error convergence constraints as the core termination condition. Error convergence constraints refer to setting an error accuracy threshold to constrain the deviation between the predicted error and the actual error of the candidate expression, ensuring that the final fitted expression accurately reproduces the error patterns of the device. The complete iterative fitting process proceeds sequentially: First, an initial population is generated by randomly generating hundreds of candidate error compensation expressions composed of building blocks as the initial population, with each expression corresponding to a symbolic tree structure; second, fitness evaluation is performed, using mean squared error as the fitness evaluation index, calculated using the following formula: ; in For the first The prediction error value output by the candidate expression after inputting the control matrix into a set of samples. The smaller the MSE value, the higher the fitting accuracy and the better the fitness of the candidate expression to the historical error pattern.

[0060] Then, evolutionary iterations are carried out. Based on the fitness evaluation results, selection, crossover, and mutation operations are performed. The selection operation adopts the tournament selection method to select candidate expressions with high fitness ranking as parents. The crossover operation generates new candidate expressions by swapping the subtree structures of two symbol trees. The mutation operation generates new individuals by randomly replacing operators, variables, or constant values ​​in the symbol tree, thus generating the next generation of candidate expression population. Finally, convergence is determined. After each round of iteration, the MSE of the current population's best expression is calculated to determine whether it meets the error convergence constraint, that is, whether the best MSE is lower than the preset convergence threshold, or whether the MSE decreases less than a set value in multiple consecutive rounds of iteration. If the constraint is not met, the fitness evaluation and evolutionary iteration steps are repeated. If the constraint is met, the iteration is terminated, and the candidate expression with the best fitness is used as the final fitted error compensation expression, thus completing the encoding and solidification of the expression.

[0061] Furthermore, in the method provided in this application embodiment, multi-threaded drive management of the end effector is performed according to the target parameter control sequence, including: decoupling the target parameter control sequence for each end effector to determine the multi-threaded drive instruction corresponding to each end effector; subjecting the multi-threaded drive instruction to spatiotemporal code constraints, converting it into a pulse width modulation signal, and sending it to the heating tube, the grinding motor, and the water pump to execute the equipment control drive.

[0062] Specifically, the target parameter control sequence is first decoupled. This sequence is an integrated structured parameter set that incorporates four core parameters: heating power, grinder motor duty cycle, water pump duty cycle, and extraction time. It is denoted as: ; in To compensate for the heating power, For the duty cycle of the coffee grinder motor, For the water pump duty cycle, Let \(t\) be the extraction duration. Decoupling refers to the process of splitting and separating the integrated parameter control set according to the functional boundaries of the end effector. The system disassembles the parameters and timing information in the target parameter control sequence according to the respective functional attributes of the heating tube, the bean grinding motor, and the water pump, and matches the heating power, preheating duration, and insulation duration parameters for the heating tube, the bean grinding duty cycle and bean grinding duration parameters for the bean grinding motor, and the water pump duty cycle and extraction duration parameters for the water pump, eliminating the coupling relationship between multiple parameters, generating independent control parameter packets corresponding to individual actuators. Each parameter packet only contains the control quantities and timing requirements recognizable by the corresponding component, thereby determining the multi-threaded drive instructions corresponding to each end effector. Each drive instruction corresponds to an independent control thread, and each thread independently bears the full-cycle control logic of the corresponding execution component. The threads only achieve process linkage through synchronous trigger signals, do not occupy computing resources from each other, do not interfere with parameter output from each other, and ensure the independence and accuracy of the control parameters of each component.

[0063] Subsequently, the system performs space-time code constraints on all thread instructions. Space-time code constraints are a rule mechanism for collaborative calibration of multi-threaded instructions from two dimensions of time and space. It is to bind a unique timestamp and device address code to each drive instruction to achieve one-to-one correspondence of timing, device, and parameters. Among them, the constraint in the time dimension is based on the standard process flow of coffee making, and uniformly calibrates the start node, operation duration, stop node of each actuator, and the delay threshold for linkage trigger. For example, it is calibrated that after the bean grinding process ends, it delays to start the water pump for extraction, synchronizes the timing reference of each thread through the global clock, and avoids process errors caused by process misalignment. The constraint in the space dimension is bounded by the rated operating parameters of each actuator, and verifies whether the control parameters in the corresponding thread are within the legal range. For example, the heating power shall not exceed the rated maximum power of the heating tube, and the duty cycle value needs to be constrained within the range of 0 to 100%, preventing execution deviation and hardware loss caused by parameter overrun. After the space-time code constraint, the drive instruction will have accurate timing positioning and device attribution, effectively avoiding misfiring of instructions and timing disorder under multi-thread parallelism.

[0064] Subsequently, the system converts the constrained multi-threaded drive instructions into pulse width modulation signals. The pulse width modulation signal is the PWM signal, which is a digital drive signal that realizes precise adjustment of power, speed, and flow by adjusting the continuous ratio of the high level within a unit cycle. The calculation formula for the core parameter duty cycle is: ; where is the duration of the high level within a single cycle, During the low-level duration, the average output power, motor speed, and water pump flow rate are all approximately linearly positively correlated with the duty cycle. During the conversion process, the system maps the power and duty cycle parameters within each thread to the corresponding PWM channel's duty cycle value. For example, the heating tube channel divides the target heating power P by the heating tube's rated power. Obtain the heating PWM duty cycle: ; The grinder motor and water pump channels directly convert the target duty cycle parameters in the control sequence into the corresponding PWM duty cycle of the channel. Simultaneously, the start and end times of the PWM signal output are set according to the timing requirements of the time-space code constraints, ultimately generating independent PWM drive waveforms for each channel. After conversion, the system sends the PWM signals of each thread in parallel to the heating element, grinder motor, and water pump through the corresponding hardware drive channels. Each actuator receives the signal and operates collaboratively according to preset parameters and timing, completing the entire equipment control and drive process for coffee making. Furthermore, in the method provided in this application embodiment, after multi-threaded drive management of the end effector, the method includes: real-time acquisition of actual feedback values ​​of the end effector based on temperature sensors, flow meters, and speed sensors; comparison of the actual feedback values ​​with multi-threaded drive commands to calculate control deviation; conversion of the control deviation into feedback compensation values ​​using a linear adjustment method to perform real-time error correction of the end effector; storage of the control deviation in a temporary database; periodic extraction of error features that satisfy a preset generalization degree; and correction of the error compensation expression.

[0065] Specifically, the system first collects the actual operating physical quantities, i.e., the actual feedback values, of each end effector at a fixed sampling frequency through temperature sensors, flow meters, and speed sensors deployed in the corresponding links of each execution component. Specifically, the temperature sensor collects the real-time water temperature of the heating chamber of the heating element, the flow meter collects the real-time water flow rate of the water pump pipeline, and the speed sensor collects the real-time operating speed of the grinder motor. All collected actual feedback values ​​are bound to the corresponding actuator's drive thread according to the sampling timestamp, serving as the true reference for deviation calculation. Next, the system calculates the control deviation thread by thread. The control deviation is the difference between the target setpoint in the multi-threaded drive instruction and the actual feedback value collected by the corresponding sensor; it is the core quantitative indicator for measuring the degree of deviation from expectations during execution. For the k-th actuator thread, the formula for calculating the control deviation at time t is: ; In the formula This refers to the target setpoint for the thread-driven instruction at time t, such as the target parameters for heating power, water pump flow rate, and grinder speed. The calculated deviation is the actual feedback value collected by the corresponding sensor at time t. It can be positive or negative. A positive value means that the actual execution volume is lower than expected and needs to be compensated positively, while a negative value means that the actual execution volume exceeds expectations and needs to be reduced negatively.

[0066] After completing the deviation calculation, the system uses a linear adjustment method to convert the control deviation into the corresponding channel's feedback compensation. The linear adjustment method refers to an adjustment algorithm where the feedback compensation is linearly proportional to the control deviation. This method features low computational latency and fast response, adapting to the rapid correction requirements of real-time execution scenarios. Its core calculation formula is: ; In the formula The proportional adjustment coefficient corresponding to the k-th executor thread needs to be pre-calibrated based on the response characteristics of each execution unit. This refers to the feedback compensation amount output at time t. The system adds this compensation amount to the driving parameters of the current thread in real time, and adjusts the duty cycle value of the pulse width modulation signal accordingly to complete the real-time error correction of the end effector. For example, when the heating water temperature is lower than the target value, a positive deviation will generate a positive compensation amount, increasing the duty cycle of the heating tube pulse width modulation signal to increase the heating power. Conversely, the duty cycle will be reduced to achieve dynamic closed-loop correction throughout the execution cycle, ensuring that the actual operating parameters are always close to the target value.

[0067] Simultaneously, the system synchronously stores each set of control deviations with timestamps and operating condition labels to a temporary database. The temporary database is a local storage unit used to cache deviation data during operation, retaining only the full deviation samples within a set period to provide a data foundation for model iteration. Subsequently, the system performs error feature extraction and model correction at fixed operating cycle intervals. During the extraction process, all deviation data within the cycle is first cleaned and filtered to remove peak deviations caused by accidental factors such as single electromagnetic interference and material abnormalities. Then, error features that meet the preset generalization degree are extracted through statistical analysis. The preset generalization degree is the criterion for judging the universality of error features. Only systematic deviation features that repeatedly appear under different operating conditions and different production batches and have statistical significance meet the generalization degree requirements. For example, the overall heating deviation caused by seasonal changes in ambient temperature and the shift in grinding speed caused by long-term wear and tear of hardware. This eliminates error data without universal value caused by occasional interference and avoids the model being interfered with by accidental noise.

[0068] Finally, the system supplements the training dataset of the symbolic regression with the setpoint and executed values ​​corresponding to the deviation samples that meet the generalization requirements, restarts the iterative fitting process of symbolic regression, optimizes the parameters or fine-tunes the structure of the original error compensation expression, so that the updated error compensation expression can adapt to the current hardware status and operating environment of the device. The corrected expression will replace the original expression and be used in subsequent feedforward parameter control corrections, such as... Figure 3 As shown, with the four core control parameters of the coffee machine—heating power, grinding duty cycle, water pump duty cycle, and extraction time—as the horizontal axis and the parameter values ​​as the vertical axis, three sets of comparison curves were plotted. PID control is a commonly used temperature control scheme in existing coffee machine models, which uses a proportional-integral-derivative algorithm to achieve closed-loop temperature regulation. This application, by introducing intent recognition threshold, symbolic regression error compensation, and multi-threaded decoupling drive, significantly reduces the deviation from the set value in all four core control parameter dimensions, with an error suppression rate of 82.4%, directly verifying the technical advantages of this application over existing PID control schemes.

[0069] In summary, the coffee machine control error suppression method based on an intent recognition model provided in this application has the following technical effects: By deploying operation verification thresholds including a mis-touch verification layer and an ambiguity verification layer at the command response end of the target coffee machine, invalid signals of mis-touch are filtered out based on the denoising fitting distribution verification of inverse denoising, and the intention probability distribution is solved based on the maximum entropy principle to resolve command ambiguity, generating precise device control commands carrying valid interaction signals. Then, the error compensation expression adapted to the device characteristics is encoded using symbolic regression, and the device control commands are converted into a parameter control sequence and the pre-error correction is completed to obtain the target parameter control sequence. Finally, the end effector is subjected to multi-threaded decoupled drive and real-time feedback closed-loop adjustment based on the target parameter control sequence. The entire link from command input verification to execution output control systematically suppresses control errors, effectively improving the execution accuracy of coffee machine control parameters and the stability and consistency of coffee product quality, thus enhancing the user experience.

[0070] Example 2, based on the same inventive concept as the coffee machine control error suppression method based on the intent recognition model in the previous examples, such as... Figure 2 As shown in the embodiment of this application, a coffee machine control error suppression system based on an intent recognition model is provided. The system includes: Control module 11 is used to deploy an operation verification threshold at the instruction response end of the target coffee machine, and generate a device control instruction carrying a valid interaction signal by real-time acquisition of device interaction signals and importing the operation verification threshold for prior knowledge; Prior knowledge module 12 is used whereby the prior knowledge based on the operation verification threshold includes: performing a denoising fitting distribution verification based on inverse denoising on the device interaction signal according to the mis-touch verification layer to obtain a first verification result representing mis-touch or non-mis-touch; Analysis module 13 is used to perform intent probability distribution analysis based on the maximum entropy principle according to the ambiguity verification layer if it is non-mis-touch, and output device control instruction or active query pop-up window; Correction module 14 is used to encode the error compensation expression using symbolic regression, and obtain the target parameter control sequence by converting the device control instruction into a parameter control sequence and performing error correction; Management module 15 is used to perform multi-threaded drive management of the end effector according to the target parameter control sequence.

[0071] Furthermore, the control module 11 is also used to perform the following steps: collecting multimodal interaction data under normal operating conditions through the target coffee machine as a positive sample set; based on the positive sample set, using a forward diffusion process to gradually add noise and reverse denoising as a training process, and using the high-dimensional distribution of normal operating interactions as the fitting target, deploying an operation verification threshold and deploying it on the interaction response end; wherein, the operation verification threshold includes a mis-touch verification layer and an ambiguity verification layer, and the ambiguity verification layer is deployed using a judgment rule based on the maximum entropy principle.

[0072] Furthermore, the prior module 12 is also used to perform the following steps: by real-time acquisition of device interaction signals and importing the operation verification threshold, the mis-touch verification layer calculates the negative log-likelihood value based on the reverse denoising process of the device interaction signals, wherein the negative log-likelihood value characterizes the denoising fitting distribution of the device interaction signals; the negative log-likelihood value is quantized to obtain an energy score; if the energy score is greater than the first energy threshold, it is determined to be a mis-touch, and the interaction response terminal does not perform a control response.

[0073] Furthermore, the analysis module 13 is also used to perform the following steps: if it is determined that it is not a mis-touch, the interaction response terminal triggers a control response and uses the device interaction signal as a valid interaction signal; retrieves the target user's historical interaction data as a constraint, and transfers the valid interaction signal and the constraint to the ambiguity verification layer, and outputs the device control command.

[0074] Furthermore, the analysis module 13 is also used to perform the following steps: according to the constraints, by solving the effective interaction signal using the maximum entropy principle, a conditional entropy based on the intent probability distribution is obtained; if the conditional entropy is less than or equal to the second entropy threshold, the effective interaction signal is determined to be unambiguous, and the device control command is output; if the conditional entropy is greater than the second entropy threshold, the effective interaction signal is determined to be ambiguous, and an active query pop-up window with maximum information gain is generated.

[0075] Furthermore, the correction module 14 is also used to perform the following steps: according to the device control command, the control center receives the valid interaction signal and generates a parameter control sequence; using a symbolic regression method, it encodes an error compensation expression; according to the error compensation expression, it corrects the parameter control sequence to obtain the target parameter control sequence.

[0076] Furthermore, the correction module 14 is also used to perform the following steps: collecting setpoint data and execution value data under historical control commands; determining the equipment parameter control matrix, using basic operators and input variables, constants, and input variables as basic units, converting the equipment parameter control matrix into input variables, outputting the error value, and using the setpoint data and execution value data to perform error convergence constraints, and fitting the error compensation expression; wherein, the equipment parameter control matrix includes heating power, grinder motor duty cycle, water pump duty cycle, and extraction time.

[0077] Furthermore, the management module 15 is also used to perform the following steps: for each end effector, decouple the target parameter control sequence to determine the multi-threaded drive instructions corresponding to each end effector; constrain the multi-threaded drive instructions with spatiotemporal codes, convert them into pulse width modulation signals, and send them to the heating tube, the grinding motor and the water pump to execute the equipment control drive.

[0078] Furthermore, the management module 15 is also used to perform the following steps: real-time acquisition of the actual feedback value of the end effector based on the temperature sensor, flow meter, and speed sensor; comparison of the actual feedback value with the multi-threaded drive command to calculate the control deviation; conversion of the control deviation into a feedback compensation amount using a linear adjustment method to perform real-time error correction of the end effector; storage of the control deviation in a temporary database, periodic extraction of error features that satisfy a preset generalization degree, and correction of the error compensation expression.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A coffee machine control error suppression method based on an intent recognition model, characterized in that, The method includes: An operation verification threshold is deployed at the command response end of the target coffee machine. The device interaction signals are collected in real time and imported into the operation verification threshold for prior knowledge, and a device control command carrying a valid interaction signal is generated. The prior based on the operation verification threshold includes: according to the accidental touch verification layer, performing a denoising fitting distribution verification based on inverse denoising on the device interaction signal to obtain a first verification result characterizing accidental touch or non-accidental touch; If it is not a mis-touch, according to the ambiguity verification layer, perform an intent probability distribution analysis based on the maximum entropy principle, and output device control commands or an active query pop-up window; The error compensation expression is encoded using a symbolic regression method. The target parameter control sequence is obtained by converting the equipment control commands into a parameter control sequence and correcting the error. Based on the target parameter control sequence, the end effector is managed using multi-threaded drive.

2. The coffee machine control error suppression method based on an intent recognition model as described in claim 1, characterized in that, Deploy operation verification thresholds at the command response end of the target coffee machine, including: Multimodal interaction data was collected from the target coffee machine under normal operating conditions and used as a positive sample set. Based on the positive sample set, the training process involves progressively adding noise through forward diffusion and denoising in reverse. The high-dimensional distribution of normal operation interaction is used as the fitting target. An operation verification threshold is deployed and applied to the interaction response end. The operation verification threshold includes a false touch verification layer and an ambiguity verification layer. The ambiguity verification layer is deployed using a judgment rule based on the maximum entropy principle.

3. The coffee machine control error suppression method based on an intent recognition model as described in claim 2, characterized in that, Based on the accidental touch verification layer, a denoising fitting distribution verification based on inverse denoising is performed on the device interaction signals, including: By collecting device interaction signals in real time and importing them into the operation verification threshold, the accidental touch verification layer calculates the negative log-likelihood value based on the reverse denoising process of the device interaction signals. The negative log-likelihood value represents the denoising fitting distribution of the device interaction signals. The negative log-likelihood value is quantified to obtain the energy score; If the energy score is greater than the first energy threshold, it is determined to be a false touch, and the interactive response terminal will not perform a control response.

4. The coffee machine control error suppression method based on an intent recognition model as described in claim 3, characterized in that, If it is determined that the touch is not accidental, the interactive response terminal triggers a control response and uses the device interaction signal as a valid interaction signal. The target user's historical interaction data is retrieved as a constraint, and the valid interaction signal and the constraint are transferred to the ambiguity verification layer to output device control commands.

5. The coffee machine control error suppression method based on an intent recognition model as described in claim 4, characterized in that, Based on the constraints, the conditional entropy based on the intent probability distribution is obtained by solving the effective interaction signal using the maximum entropy principle. If the conditional entropy is less than or equal to the second entropy threshold, the valid interaction signal is determined to be unambiguous, and the device control command is output. If the conditional entropy is greater than the second entropy threshold, it is determined that the effective interaction signal is ambiguous, and an active inquiry pop-up window with maximum information gain is generated.

6. The coffee machine control error suppression method based on an intent recognition model as described in claim 1, characterized in that, Error correction is performed to obtain the target parameter control sequence, including: According to the device control command, the control center receives the valid interaction signal and generates a parameter control sequence; The error compensation expression is encoded using symbolic regression. The parameter control sequence is corrected according to the error compensation expression to obtain the target parameter control sequence.

7. The coffee machine control error suppression method based on an intent recognition model as described in claim 6, characterized in that, The error compensation expression is encoded using a symbolic regression approach, including: Collect setpoint and execution value data under historical control commands; The equipment parameter control matrix is ​​determined, using basic operators and input variables, constants, and input variables as primitives. The equipment parameter control matrix is ​​converted into input variables, with the error value as the output. The error convergence constraint is performed using the set value data and the executed value data, and the error compensation expression is fitted. The equipment parameter control matrix includes heating power, grinding motor duty cycle, water pump duty cycle, and extraction time.

8. The coffee machine control error suppression method based on an intent recognition model as described in claim 1, characterized in that, Based on the target parameter control sequence, the end effector is managed using a multi-threaded drive, including: For each end effector, the target parameter control sequence is decoupled to determine the multi-threaded driving instructions corresponding to each end effector; The multi-threaded drive instructions are constrained by a time-space code, converted into pulse width modulation signals, and sent to the heating tube, grinding motor, and water pump to execute equipment control drive.

9. The coffee machine control error suppression method based on an intent recognition model as described in claim 1, characterized in that, After implementing multi-threaded management of the end effector, the following is included: The actual feedback values ​​of the end effector are collected in real time based on temperature sensors, flow meters, and speed sensors. The control deviation is calculated by comparing the actual feedback value with the multi-threaded drive instructions. A linear adjustment method is used to convert the control deviation into a feedback compensation amount for real-time error correction of the end effector. The control deviation is stored in a temporary database, and error features that satisfy a preset generalization degree are periodically extracted to correct the error compensation expression.

10. A coffee machine control error suppression system based on an intent recognition model, characterized in that, For implementing the coffee machine control error suppression method based on the intent recognition model according to any one of claims 1 to 9, the system comprises: The control module is used to deploy operation verification thresholds at the command response end of the target coffee machine, and generate device control commands carrying valid interaction signals by real-time acquisition of device interaction signals and importing the operation verification thresholds for prior knowledge. The prior module, which is used to perform a priori verification based on the operation verification threshold, includes: performing a denoising fitting distribution verification based on inverse denoising on the device interaction signal according to the accidental touch verification layer, to obtain a first verification result characterizing accidental touch or non-accidental touch; The analysis module is used to perform an intent probability distribution analysis based on the maximum entropy principle, according to the ambiguity verification layer, if the touch is not accidental, and output device control commands or proactive query pop-ups. The correction module is used to encode the error compensation expression using a symbolic regression method, and to obtain the target parameter control sequence by converting the device control command into a parameter control sequence and performing error correction. The management module is used to perform multi-threaded drive management of the end effector according to the target parameter control sequence.

Images