Key intelligent response method, user terminal, storage medium and program product

Abstract

A key intelligent response method, a user terminal, a storage medium and a program product, relate to the field of digital data processing, and the method comprises the following steps: based on the first time sequence touch data collected during the period when the mouse device does not generate displacement signals, obtained from sample users, a two-hand mistaken touch recognition model is trained; based on the second time sequence touch data collected during the period when the mouse device generates displacement signals, obtained from sample users, a single-hand mistaken touch recognition model is trained; during the running of the user terminal, the target displacement signal of the mouse device is monitored; according to the target displacement signal, one of the two-hand mistaken touch recognition model and the single-hand mistaken touch recognition model is determined as the current effective recognition model; the real-time touch data is input into the current effective recognition model for calculation to obtain a key position effective score; when the key position effective score is lower than a preset effective threshold, the corresponding error key signal is prevented. By implementing the present application, the mistaken detection accuracy on a high-sensitivity keyboard can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric digital data processing, and relates to a key intelligent response method, a user terminal, a storage medium and a program product. BACKGROUND

[0002] With the development of human-computer interaction technology, as a core input device, the response speed and input efficiency of the keyboard become the key indicators for measuring user experience. In order to meet the pursuit of high-speed and smooth input experience by users, the design trend of the keyboard gradually develops towards high sensitivity, aiming to shorten the key response time and reduce the user's operation fatigue. However, the high sensitivity keyboard, while improving the response speed, also increases the probability of accidental touch.

[0003] In the related technology, in order to reduce the probability of accidental touch, a misjudgment detection method based on multi-dimensional physical parameter threshold is provided. When a key touch event is monitored, the method will evaluate multiple physical parameters of the event, such as the duration of the key signal and the contact pressure. When the physical parameters of a touch event (such as the duration exceeding the preset duration threshold T1 and the peak pressure being greater than the preset pressure threshold P1) meet the preset combination condition, the system will determine it as a valid key input. This method can filter out invalid signals caused by mechanical shaking or extremely slight and short transient contact to a certain extent.

[0004] However, in complex use scenarios, such as during a game, the long-time compression signal caused by the palm accidentally leaning on the keyboard, or the high-pressure unintended signal caused by an object falling, the related technology will not detect the accidental touch because the signal meets the corresponding threshold, and the misjudgment detection on the high sensitivity keyboard is not accurate enough. SUMMARY

[0005] The present application provides a key intelligent response method, a user terminal, a storage medium and a program product, for improving the accuracy of misjudgment detection on a high sensitivity keyboard.

[0006] In a first aspect, the present application provides a key intelligent response method applied to a user terminal, the method comprising: training a double-hand accidental touch recognition model based on first time-series touch data collected from sample users during a period when a mouse device does not generate displacement signals; training a single-hand accidental touch recognition model based on second time-series touch data collected from sample users during a period when the mouse device generates displacement signals; monitoring a target displacement signal of a connected mouse device during the running of the user terminal; determining one of the double-hand accidental touch recognition model and the single-hand accidental touch recognition model as a current effective recognition model according to the target displacement signal; inputting real-time touch data generated by a current user into the current effective recognition model for calculation to obtain a key position validity score; and preventing a corresponding false key signal when the key position validity score is lower than a preset validity threshold.

[0007] In the above embodiment, the user terminal distinguishes between single-hand and double-hand operation scenarios according to the mouse displacement signal, calls the recognition model for targeted training, and makes judgments on the effectiveness of the keys in both double-hand high-speed typing and single-hand mouse operation scenarios, thereby improving the accuracy of the false touch detection.

[0008] In combination with some embodiments of the first aspect, in some embodiments, the step of training the single-hand false touch recognition model based on the second time-series touch data collected from the sample user during the generation of the displacement signal by the mouse device specifically includes: performing clustering calculation on the second time-series touch data collected from the sample user during the generation of the displacement signal by the mouse device, determining false touch key points with preset large-area characteristics and preset low-pressure characteristics, and generating negative sample labels; determining effective key points within the single-hand stretching range based on the trigger position and time-series correlation of the key events, in combination with the single-hand effective operation area, and generating positive sample labels; extracting the position parameters, pressure values, duration, and spatial distance from the previous key event of each key point in the key event, and generating feature vectors; inputting the feature vectors and corresponding positive sample labels and negative sample labels into a neural network for training to obtain the single-hand false touch recognition model.

[0009] In the above embodiment, the user terminal identifies data with typical false touch characteristics as negative samples, and determines positive samples in combination with the ergonomics characteristics (single-hand stretching range), thereby ensuring the high quality of the training data, enabling the model to learn the single-hand operation mode from multiple perspectives such as position, pressure, and time, and more accurately identifying complex false touches caused by palm movement.

[0010] In combination with some embodiments of the first aspect, in some embodiments, before the step of monitoring the target displacement signal of the connected mouse device during the running of the user terminal, the method further includes: displaying a single-hand operation calibration interface containing a preset key sequence to collect calibration touch data generated during the user pressing the preset key sequence; calculating the maximum spatial distance between the key points in the calibration touch data to obtain the target single-hand stretching range of the user; taking the key position with the highest key frequency in the calibration touch data as the center and taking the target single-hand stretching range as the radius to determine the target single-hand operation area of the user; and correcting the single-hand false touch recognition model based on the target single-hand stretching range and the target single-hand operation area.

[0011] In the above embodiment, the user terminal corrects the pre-trained model based on the key data of a specific user, so that the model is more accurate in judging the effective key area of the user, reduces the false judgments caused by the size and operation posture differences of the user's palm, realizes the personalized adaptation of the model, and improves the detection accuracy.

[0012] In some embodiments of the first aspect, in some embodiments, the step of determining one of the double-hand mistouch recognition model and the single-hand mistouch recognition model as the current effective recognition model according to the target displacement signal specifically comprises: taking the double-hand mistouch recognition model as the current effective recognition model; collecting displacement signals and key timestamp of the keyboard within a preset time period, and calculating cumulative displacement of the displacement signals and key frequency; extracting position coordinates of each key point within the preset time period, and calculating spatial distribution variance of the position coordinates; when the cumulative displacement exceeds a preset mouse activity threshold, the key frequency is lower than a preset double-hand input threshold, or the spatial distribution variance is lower than a preset concentration threshold, changing the double-hand mistouch recognition model in the current effective recognition model to the single-hand mistouch recognition model.

[0013] In the above embodiments, the user terminal comprehensively considers mouse displacement, key frequency and key position distribution variance as the basis for judgment, which can more reliably identify the switching of scenarios from double-hand typing to single-hand gaming, for example, even if the user temporarily moves the mouse during double-hand typing, the system will not be switched incorrectly, ensuring the stability and accuracy of model switching, and avoiding the decline of recognition performance caused by incorrect judgment of scenarios.

[0014] In some embodiments of the first aspect, before the step of changing the double-hand mistouch recognition model in the current effective recognition model to the single-hand mistouch recognition model, the method further comprises: reading a foreground process identifier currently running on the user terminal, and determining a single-hand scenario bias coefficient corresponding to the foreground process identifier according to a process scenario mapping table; and adjusting the preset mouse activity threshold and the preset concentration threshold according to the single-hand scenario bias coefficient.

[0015] In the above embodiments, the user terminal can predict that the user is most likely in single-hand or double-hand operation mode by identifying the currently running program (such as a game or a document editor), and dynamically adjust the switching threshold using the single-hand scenario bias coefficient, for example, lower the threshold for switching to single-hand mode in a game, so that the scenario switching decision is more in line with the actual intention of the user, improving the sensitivity and accuracy of switching.

[0016] In some embodiments of the first aspect, after the step of preventing the corresponding error key signal when the key position effective score is lower than the preset effective threshold, the method further comprises: recording the key identifier corresponding to the blocked key signal and the blocking time; when a repeated key signal with the same key identifier is received within a preset time window after the blocking time, marking the error key signal as a misjudgment sample and storing it in a correction sample library; when the number of samples in the correction sample library reaches a predetermined value, extracting a misjudgment feature vector of the misjudgment sample; and based on the misjudgment feature vector and the corresponding positive sample label, calculating a weight update amount of the misjudgment feature vector to the current effective recognition model by a gradient descent algorithm, and loading the weight update amount to the current effective recognition model for incremental update.

[0017] In the above embodiment, when the user terminal misjudges (i.e. mistakenly blocks the valid key of the user), the user terminal identifies the misjudgment by monitoring the repeated key behaviors of the user in a short time, collects the misjudgment samples, and uses the misjudgment samples to incrementally update the current model to continuously improve the recognition accuracy of the specific user operation mode.

[0018] In combination with some embodiments of the first aspect, in some embodiments, when the repeated key signal identical to the key identifier is received within the preset time window after the blocking moment, the step of marking the error key signal as a misjudgment sample and storing the misjudgment sample in the correction sample library specifically includes: when the repeated key signal identical to the key identifier is received within the preset time window after the blocking moment, inputting the touch data corresponding to the error key signal and the touch data corresponding to the repeated key signal into the current effective recognition model respectively to obtain a first score and a second score; calculating the absolute value of the difference between the first score and the second score; when the absolute value of the difference is less than a preset fluctuation threshold, marking the touch data corresponding to the error key signal as a misjudgment sample and storing the misjudgment sample in the correction sample library; when the absolute value of the difference is greater than or equal to the fluctuation threshold and the second score is higher than a preset effective threshold, extracting the time points corresponding to the pressure peaks in the error key signal and the repeated key signal and calculating the time difference; and when the time difference is less than a preset time length threshold, marking the touch data corresponding to the error key signal as a misjudgment sample and storing the misjudgment sample in the correction sample library.

[0019] In the above embodiment, the user terminal compares the model scores of the two times of key pressing, the time difference of the pressure peaks, and other deep features to determine whether the blocked key is the real intention of the user, which can effectively filter out the pseudo-repeated key caused by the hesitation or operation habit of the user, avoid the interference of noise data on the model optimization process, and ensure the effectiveness of the incremental update.

[0020] In the second aspect, the embodiments of the present application provide a user terminal, which comprises one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program codes, the computer program codes comprising computer instructions, and the one or more processors invoke the computer instructions to enable the user terminal to perform the method described in the first aspect and any possible implementation manner of the first aspect.

[0021] In the third aspect, the embodiments of the present application provide a computer program product comprising instructions, which, when the computer program product is executed on a user terminal, enable the user terminal to perform the method described in the first aspect and any possible implementation manner of the first aspect.

[0022] In a fourth aspect, the embodiments of the present application provide a computer readable storage medium comprising instructions which, when executed on a user terminal, cause the user terminal to perform the method as described in the first aspect and any possible implementation of the first aspect.

[0023] It can be understood that the user terminal provided in the second aspect, the computer program product provided in the third aspect and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiments of the present application. Therefore, the beneficial effects achieved thereby can refer to the beneficial effects in the corresponding method, which will not be described here again.

[0024] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0025] 1. Since the technical solution of dynamically determining and switching the double-hand and single-hand mis-touch recognition models based on the mouse displacement signal is adopted, it can be determined according to the objective index of whether the user is moving the mouse that the user is in the scene of double-hand typing or single-hand operation with the mouse, and the model optimized for the scene is called to perform mis-touch recognition. This solution effectively solves the problem that the single fixed strategy in the prior art cannot take into account the different mis-touch features in different use scenarios, and thus realizes high-precision and self-adaptive detection of keyboard mis-touch in complex and changeable human-computer interaction scenarios, and improves the overall input experience.

[0026] 2. Since the single-hand model training method combining clustering calculation and ergonomic features for sample labeling is adopted, the positive and negative samples required for training can be automatically and accurately generated. The typical mis-touch data with large area and low pressure characteristics are identified as negative samples by clustering, and the effective keys are determined as positive samples according to prior knowledge such as single-hand stretching range, so as to ensure the quality and discrimination of the training data. This solution effectively solves the problem of difficult sample labeling and low quality in the prior art, and thus realizes training of a more robust and accurate single-hand mis-touch recognition model.

[0027] 3. Since the multi-dimensional scene switching judgment mechanism combining the accumulated displacement amount of the mouse, the key frequency and the spatial distribution variance of the key position is adopted, the change of the user operation mode can be more stably and reliably recognized. This solution makes a judgment by comprehensively considering multiple indexes, avoiding frequent or incorrect switching caused by relying on a single index (such as instantaneous mouse movement). This solution effectively solves the problem of simple scene switching logic and easy interference in the prior art, and thus realizes smooth and accurate model switching, and guarantees the continuous effectiveness of the mis-touch recognition function during the scene transition. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a flowchart of the method for intelligent response of keys in the embodiments of the present application;

[0029] Figure 2 is another flowchart of the intelligent response method of the key in the embodiment of the application;

[0030] Figure 3 is an entity device structure diagram of the user terminal in the embodiment of the application. DETAILED DESCRIPTION

[0031] The terms used in the following embodiments of the application are only for the purpose of describing specific embodiments and are not intended to be limiting of the application. As used in the specification, the singular forms "a," "an" and "the" are intended to include both singular and plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "and / or" as used herein refer to any or all possible combinations of one or more of the associated listed items.

[0032] Hereinafter, the terms "first" and "second" are only for the purpose of description, and cannot be understood as implying or suggesting relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features, and in the description of the embodiments of the application, the meaning of "multiple" is two or more, unless otherwise specified.

[0033] The technical solution of the application is that, by perceiving the use scenario of the user (mainly distinguishing between single-hand keyboard operation and double-hand keyboard operation), and applying a mis-touch recognition model matched with the scenario, the accurate judgment of high-sensitivity keyboard input is realized. Under this technical framework, some key technical terms have specific connotations. For example, the time-series touch data refers to a series of original data streams collected by the keyboard sensor, including time stamp, key position, pressure value and other multi-dimensional information, which is the basis for model analysis. The double-hand mis-touch recognition model and the single-hand mis-touch recognition model are two independent machine learning models, the former is trained for regular typing scenarios, and the latter is optimized for scenarios such as single-hand operation keyboard and another hand operating mouse. Whether the mouse device generates a displacement signal is the core external signal to distinguish between the two scenarios. The key position effective score is the quantitative evaluation result of the model on whether each key event is the user's real intention, and the score directly determines whether the key signal is accepted by the system.

[0034] The method provided by the embodiment is described in the following flow. Please refer to Figure 1 is a flowchart of the intelligent response method of the key in the embodiment of the application.

[0035] S101, based on the first time-series touch data collected during the period when the mouse device does not generate a displacement signal, a double-hand mis-touch recognition model is trained.

[0036] wherein the sample users refer to a user group providing original operation data for model training, and the mouse device not generating displacement signals refers to a time period when the user's hands do not operate the mouse, and the mouse cursor is static on the screen, which usually corresponds to a scenario that both hands are focused on keyboard input. The first time-series touch data refers to time-series data collected from the keyboard sensor during the period, which contains information such as key position, pressure, duration, etc.

[0037] Specifically, this step is performed offline before the method is deployed. The data acquisition module of the user terminal collects touch data of a large number of sample users when they perform typical two-handed keyboard operations such as text input and programming. During the collection process, the system synchronously records the state of the mouse, and filters out the corresponding keyboard touch data in the mouse static period as the first time-series touch data. This data contains a large number of labeled valid keys (positive samples) and accidental touches (negative samples). Then, these labeled data are input into a preset neural network structure (such as a recurrent neural network RNN or a long short-term memory network LSTM) for training. Through back propagation and parameter optimization, a two-handed accidental touch recognition model capable of recognizing accidental touch features in a two-handed operation mode is finally obtained.

[0038] It should be noted that the two-handed accidental touch recognition model is a deep learning network for processing time-series data, and its core is to accurately distinguish valid keys and accidental touches of users in a two-handed typing scenario.

[0039] A large amount of keyboard time-series touch data with clear labels is required to train the model. The input data is a fixed-length time-series segment for each key event, for example, a continuous sensor data stream of 100 milliseconds forward and 150 milliseconds backward centered on the key pressure peak time, which contains multi-dimensional information such as coordinates, pressure, and area of all touch points at each time point. The labels of these data segments are binary (0 or 1), where "1" represents a valid key annotated by human or verified by other reliable means, and "0" represents an accidental touch. The training standard is to minimize the binary cross-entropy loss function (Binary Cross-Entropy Loss) between the predicted results of the model and the true labels. During the training process, optimization algorithms such as Adam are used to continuously adjust the network weights through back propagation until the accuracy, recall rate, and other indicators of the model on the independent validation set reach the preset convergence standard.

[0040] The model usually adopts a long short-term memory network (LSTM) or its variant gated recurrent unit (GRU) as the core architecture. The model receives the above-mentioned formatted time series data segment as input, captures the dynamic pattern of physical quantities such as pressure and area changing over time during the key pressing process, and the spatio-temporal correlation between multiple touch points through several LSTM layers. For example, LSTM can learn that an effective key pressing usually has a clear pattern of "rapid pressure rise-short maintenance-rapid decline", while palm scratching may be a pattern of "slow pressure change and large area". The final hidden state of the LSTM layer will be integrated through one or more dense layers (Dense Layer), and finally output a single floating-point number between 0 and 1 through a sigmoid activation function.

[0041] In the actual operation of the user terminal, when the keyboard sensor detects a key pressing event, the system will capture the original time series touch data around the event in real time and process it into the same format as during training. Subsequently, this data segment is input into the trained two-hand mis-touch recognition model. The model performs a forward propagation calculation, and the output floating-point number between 0 and 1 is the "key position effective score". This score represents the likelihood of the model considering this key pressing as the user's real intention, and the subsequent system will determine whether to release or block the key signal based on the comparison result of this score and the preset threshold.

[0042] S102, based on the second time series touch data collected during the generation of the displacement signal of the mouse device from the sample user, a single-hand mis-touch recognition model is trained.

[0043] Specifically, this step is also performed offline before method deployment. The user terminal collects touch data of the sample user when playing first-person shooter games (FPS), real-time strategy games (RTS), and other applications that require frequent simultaneous operation of the mouse and keyboard. The system selects the data generated by the keyboard during the period when the mouse is moving as the second time series touch data. Unlike the two-hand scenario, the mis-touch patterns in the single-hand scenario are more diverse, such as the palm side touching other keys in order to press the far keys. Therefore, the training of this model requires more targeted features and samples. After labeling the collected data, the user terminal uses it to train an independent single-hand mis-touch recognition model, which is specifically designed to learn and recognize the input patterns and mis-touch patterns specific to single-hand operation.

[0044] In some embodiments, the model training in this step can be implemented in various ways: optionally, the user terminal can construct a more feature-rich input vector, in addition to the basic pressure, duration, add features such as the distance of the key and the commonly used area of the single hand (such as WASD), the time and spatial interval with the previous key, etc., and then input it to a gradient boosting decision tree (GBDT) model for classification training; optionally, the user terminal can use a similar sequence model as in step S101, but use a data set collected and labeled in a single-hand scenario to train the model, so that the model learns the specific timing dependency of single-hand operation. It can be understood that other machine learning algorithms suitable for this specific scenario can also be used to implement the construction of the single-hand mis-touch recognition model, which is not limited here.

[0045] It should be noted that the single-hand mis-touch recognition model is a machine learning model focusing on single-hand keyboard, single-hand mouse scenarios such as games, and its feature engineering and model structure are more focused on processing the unique patterns of single-hand operation.

[0046] The data used to train this model comes from the keyboard input of the user in a single-hand operation scenario such as playing games. Unlike the two-hand model, its input data is usually not the original time series, but a carefully designed feature vector. As described in S204, this vector contains static features of the current key event (such as key position coordinates, pressure peak, duration) and context-related features (such as spatial distance and time interval with the previous valid key, distance to the commonly used game key area WASD, etc.). The labels (0 for mis-touch, 1 for valid) of these feature vectors are automatically generated through the combination of clustering and heuristic rules described in S202 and S203. The training standard is also to minimize the loss function between the prediction and the true label, but since the input is a structured feature vector rather than a sequence, in addition to neural networks, integrated learning models such as gradient boosting decision trees (GBDT) are also commonly used for training, with the goal of enabling the model to learn to distinguish between valid instructions and mis-touches caused by palm movement or finger extension in the context of single-hand operation.

[0047] The model can be a multi-layer perceptron (MLP) or a gradient boosting decision tree (GBDT) set. If MLP is used, the number of input layer nodes is consistent with the dimension of the feature vector, and after several hidden layers with ReLU activation function, the final score is output by a Sigmoid output layer. If GBDT is used, the model is composed of hundreds of decision trees, and each tree is fitted to the residual of all trees before fitting. An input feature vector will be judged by all trees, and the final prediction results of all trees will be weighted and summed, and then converted into a probability score through a logic function (such as Sigmoid). GBDT model usually performs well for tabular feature data and has good interpretability.

[0048] When the system determines that the user enters the single-hand operation scenario through mouse displacement signals and the like, the model is enabled. For each keyboard touch, the system extracts the same feature vector as in the training, for example, the R key is pressed, the system calculates its coordinates, pressure, and duration, and finds the previous key (which may be W) to calculate the spatial distance, and packs these values into a vector. This vector is sent to the single-hand mis-touch recognition model (whether MLP or GBDT), and the model outputs a "key position validity score" between 0 and 1. This score combines the physical characteristics of the key itself and its rationality in the single-hand game operation logic chain, so as to more accurately filter out the mis-touches specific to the single-hand scenario, such as a slight scratch of the T key while trying to press the G key.

[0049] In some embodiments, there may be cases where the effective key region during single-hand operation varies from person to person. For this, the user terminal will require the sample user to perform a short calibration task when collecting the second time sequence touch data, such as pressing a specific set of keys in order (such as Q-E-R-F-V-C). By analyzing the key data when the user completes the task, the user terminal can preliminarily estimate the single-hand comfortable operation region and the limit extension range of each sample user. This information can be used as prior knowledge to assist in subsequent automatic or semi-automatic labeling of the second time sequence touch data, for example, touch data that falls outside the user's comfortable operation region, has low pressure, and short duration, has a higher probability of being labeled as a negative sample, thereby improving the quality of the training data.

[0050] S103、In the user terminal running period, the target displacement signal of the connected mouse device is monitored.

[0051] Among them, the user terminal running period refers to the stage after the method is deployed, and the user normally uses the computer. The target displacement signal refers to the raw data or processed displacement information that can represent the physical movement of the mouse, such as the pixel movement amount per unit time, obtained from the mouse driver program or the operating system level.

[0052] Specifically, after the user logs in to the operating system, a monitoring service built-in in the user terminal starts running in the background. The service obtains the underlying event data of the mouse in real time through hooking system messages or directly interacting with the input device driver. The core it cares about is the mouse movement event. Whenever the mouse moves, the monitoring service will record the timestamp and displacement information (such as the change amount of X, Y coordinates) of the event. These information will be cached and accumulated within a very short time window to form a quantitative target displacement signal for subsequent scenario judgment. The monitoring process is continuous and low-power, and will not have a perceptible impact on the normal performance of the user terminal.

[0053] In some embodiments, there can be cases where users have different DPI settings for their mouse, resulting in very different displacement signal values for the same physical movement distance. To address this, the user terminal provides a calibration option in the first run or in the settings interface. In the calibration process, the user terminal asks the user to move the mouse from one end of the screen to the other and records the accumulated displacement signal value during this process. By correlating this value with the screen resolution, the user terminal can calculate a standardization coefficient. In subsequent monitoring, all original displacement signals will be multiplied by this coefficient for normalization, thus eliminating the effects of different DPI and operating system pointer speed settings, ensuring that the subsequent displacement-based judgment logic is consistent and universal.

[0054] S104、According to the target displacement signal, determine one of the two-hand mis-touch recognition model and the single-hand mis-touch recognition model as the current effective recognition model.

[0055] Among them, the current effective recognition model refers to the model selected by the system at this moment for processing real-time keyboard input.

[0056] Specifically, the user terminal maintains a state variable inside to identify whether the current effective model is the two-hand model or the single-hand model. By default, the initial state is that the effective model is the two-hand mis-touch recognition model. The decision module of the user terminal continuously analyzes the target displacement signal from step S103. That is, within a predetermined time window (for example, 500 milliseconds), if the accumulated mouse displacement is zero or below a very small noise threshold, the decision module determines that the user is in a two-hand keyboard operation state, and continues to maintain or switch to the two-hand mis-touch recognition model; otherwise, if the accumulated displacement exceeds a predetermined activity threshold, it is determined that the user is actively using the mouse, and the state variable is changed to switch the current effective recognition model to the single-hand mis-touch recognition model.

[0057] In some embodiments, the model determination in this step can be achieved in various ways: optionally, the user terminal can use a simple threshold judgment logic with hysteresis. For example, set two thresholds T_high and T_low (T_high>T_low), when the displacement from low to high exceeds T_high, switch to the single-hand model; when the displacement from high to low is below T_low, switch back to the two-hand model, this hysteresis mechanism can prevent frequent switching in the critical state; optionally, the user terminal can introduce more complex judgment logic, for example, not only looking at the displacement of the current window, but also combining the displacement history of the previous few windows to form a moving average, based on which to make a judgment to obtain a smoother and more robust switching effect. It can be understood that more complex state machine or simple rule engine based methods can also be used to achieve dynamic determination of the model, which is not limited here.

[0058] In some embodiments, there can be a case where the user only slightly and briefly adjusts the mouse position during the typing gap, in which case the model should not be switched to the single-hand model. For this purpose, the user terminal will adopt a composite judgment standard. In addition to monitoring the target displacement signal, the user terminal will also monitor the keyboard key frequency synchronously. Only when both the mouse displacement exceeds the threshold and the keyboard key frequency is lower than a frequency threshold that should be the case for two-hand typing, the model will be switched to the single-hand false touch recognition model. The addition of this logic can effectively filter out the false signals in the above scenario, ensuring that the model is only switched when the user has indeed turned his attention to the mouse and the keyboard input intensity has decreased, thereby improving the accuracy of the switching.

[0059] S105, input the real-time touch data generated by the current user into the currently effective recognition model for calculation to obtain a key position effective score.

[0060] Among them, the real-time touch data refers to the instantaneous, unprocessed raw sensor data stream generated by each key pressing action of the user. The key position effective score is a quantitative output result, which is a floating-point number between 0 and 1. The higher the value, the greater the likelihood that the model considers the key pressing to be an effective operation.

[0061] Specifically, when the user performs any touch operation on the keyboard, the firmware or driver of the keyboard will generate a raw data packet containing key coordinates, time stamp, real-time pressure value, etc., i.e. real-time touch data. The response module of the user terminal will capture the data and perform corresponding preprocessing according to the currently effective recognition model (two-hand or single-hand model) determined in step S104. For example, it is constructed into a feature vector or a time sequence segment. The processed data is sent to the effective model for forward propagation calculation. The model will evaluate the input touch pattern according to the parameters it has learned internally and output a probability value representing effectiveness, i.e. the key position effective score.

[0062] In some embodiments, the score calculation in this step can be achieved in various ways: optionally, if the effective model is a neural network, the input data will undergo nonlinear transformation through each layer of the network, and finally pass through a Sigmoid or Softmax activation function to output the activation value of one or more nodes, which is the key position effective score; optionally, if the effective model is a tree model such as GBDT, the feature vector of the input data will be judged by multiple decision trees, and finally the output results of all trees will be weighted and summed to obtain the final key position effective score. It can be understood that the specific calculation method of the score is closely related to the model architecture adopted, and any model that can output a quantitative evaluation value can be applicable, which is not limited here.

[0063] In some embodiments, there are cases where the context of the key behavior is crucial to the judgment, such as rapid consecutive presses of the same key. In this case, the user terminal not only inputs the data of the current key event when inputting real-time touch data to the model, but also attaches context information. These information can include: the time interval, spatial distance from the previous key event, and the key density in the past short period of time (such as 100 milliseconds) and so on. These context features and the features of the current key itself (pressure, duration) are combined into a richer input vector, which is then sent to the model for calculation. In this way, the model can make judgments based on more comprehensive information, for example, it can learn to distinguish between two slight touches caused by hand shaking and the player's rapid double-click skill key in the game, which are similar on the surface but completely different in intention.

[0064] S106, when the key position effective score is lower than the preset effective threshold, preventing the corresponding error key signal.

[0065] The preset effective threshold is a key parameter, for example, 0.5, which is used as the decision boundary for judging whether the key is effective. Prevention means intercepting this key event so that it is not passed to the operating system and upper application, so that the user level shows that this key does not occur. The error key signal refers to the key event that is judged by the model to be invalid (i.e. the score is low) and is eventually prevented.

[0066] Specifically, after calculating the key position effective score in step S105, the decision module of the user terminal will immediately compare the score with a pre-set effective threshold. If the score is greater than or equal to the threshold, the decision module recognizes that the key is the user's real intention operation, and then allows the corresponding key signal to be normally reported to the operating system, and the user will see character input or application response on the screen. On the contrary, if the score is lower than the threshold, the decision module recognizes that the key is a false touch. At this time, it will perform the prevention operation, that is, discard the data packet of this key event, so that it does not enter the subsequent system processing flow. The whole judgment and prevention process occurs in a very short time (usually milliseconds), and the user can hardly perceive the delay.

[0067] In some embodiments, the threshold setting can be too strict, resulting in some users' special but effective keying habits (such as very light tapping force) being frequently misjudged as false touch. In this regard, the preset effective threshold of the user terminal is not fixed, but adjustable. In the setting interface of the user terminal, an option of false touch prevention sensitivity is provided, allowing the user to choose between conservative, balanced, aggressive and other levels, or directly drag the slider to fine-tune. Behind these settings, the preset effective threshold is actually adjusted. For example, the conservative mode corresponds to a higher threshold, which will prevent more suspicious keystrokes; while the aggressive mode corresponds to a lower threshold, which is more inclined to release the keystrokes, thus meeting the use preferences of different users.

[0068] The method provided by the present embodiment is further described in more detail below. Please refer to Figure 2 , another flowchart of the intelligent response method of the key in the present embodiment.

[0069] S201, based on the first time sequence touch data collected during the period when the mouse device does not generate displacement signals, obtained from the sample users, a two-hand false touch recognition model is trained.

[0070] Referring to step S101, which will not be repeated here.

[0071] S202, performing clustering calculation on the second time sequence touch data collected during the period when the mouse device generates displacement signals, obtained from the sample users, to determine false touch keystroke points with preset large area characteristics and preset low pressure characteristics, and generate negative sample labels.

[0072] Among them, the preset large area characteristic refers to the physical area covered by a touch event on the sensor exceeding the range of a single fingertip pressing normally, which is usually the palm or wrist pressing on the keyboard. The preset low pressure characteristic refers to the maximum pressure value generated by the touch event being far below the pressure level of normal effective keystrokes, which is usually caused by accidental light scratching or hovering of the finger. The negative sample label refers to marking the data points or data segments that meet these false touch characteristics as false touch (for example, the label is 0).

[0073] Specifically, this step is the data preprocessing link before the single-hand mis-touch recognition model training. The user terminal first collects a large amount of keyboard touch data (i.e., the second time sequence touch data) in a single-hand operation scenario, at this time these data are unlabeled. Then, the user terminal extracts the key physical parameters of each touch event, such as touch area, peak pressure, duration, etc., to form a multi-dimensional feature space. Then, clustering algorithms such as DBSCAN or K-Means are applied to cluster these feature points. The algorithm will automatically cluster similar feature data points together. The user terminal analyzes the feature centers of each cluster, identifies clusters with large area and / or low pressure statistical characteristics, and automatically labels all data points in these clusters as negative samples (mis-touch).

[0074] It should be noted that when performing clustering calculation to label negative samples, the system will extract a feature set for each independent touch event (from finger contact to leaving the keyboard a complete process), for example, the maximum area of the sensor covered by this touch event, the peak pressure, the average pressure, and the touch duration. These features together form a data point in a multi-dimensional feature space. Then, the system will use a density clustering algorithm like DBSCAN, because it does not need to pre-set the number of clusters, and can effectively identify the "large area, low pressure" touch point set formed by accidental palm edge or wrist resting on the keyboard, which is irregular in shape but similar in density. After the algorithm is executed, a number of data clusters and some outliers labeled as noise are generated. At this time, the system will not find the "feature center", but will perform overall statistical analysis on each formed cluster, calculate the area average, pressure average and duration average of all data points in the cluster. Then, the system compares these statistical values with the preset "mis-touch feature threshold", for example, if the area average of a cluster exceeds 80 square millimeters (much larger than the area of a single fingertip), and its pressure average is less than 30% of the normal key pressure, then the system will determine this cluster as a "palm pressure mis-touch cluster" and automatically label all touch event data points contained in it as negative samples. Through this way of judging based on the macro statistical characteristics of the entire cluster, the data group with typical mis-touch mode can be accurately identified and labeled at one time.

[0075] In some embodiments, the negative sample labeling in this step can be achieved in various ways: optionally, the user terminal can employ the DBSCAN algorithm, which is capable of discovering clusters of arbitrary shape and does not require pre-specification of the number of clusters. By setting a suitable neighborhood radius (Eps) and minimum number of samples (MinPts), a large piece of continuous low-pressure touch point formed by palm pressing can be identified as an independent cluster, thereby being accurately labeled as a negative sample; optionally, the user terminal can first perform preliminary screening through a simple threshold method, for example, directly labeling touch events with touch area greater than A_threshold or peak pressure less than P_threshold as negative samples, and then performing clustering analysis on the remaining ambiguous data to improve the efficiency and accuracy of labeling. It can be understood that various unsupervised and semi-supervised learning methods can also be combined to optimize the automatic labeling process of negative samples, which is not limited here.

[0076] In some embodiments, there may be certain special game operations (such as pressing multiple keys with the thumb at the same time) that may be similar in characteristics to palm pressing false touch, leading to false labeling as negative samples. In this regard, the user terminal will introduce a context-based verification logic after performing clustering calculation. For a large-area touch cluster preliminarily labeled as a negative sample, the user terminal will check whether there are other high-pressure, short-duration valid key events in the adjacent area and time window when the cluster occurs. If there are, and these events are reasonable in game operation logic (such as jump + squat), the user terminal will cancel the negative sample labeling of the large-area touch cluster, or label it as pending manual review, to avoid misjudging valid complex operations as false touch and ensure the purity of the training data.

[0077] S203, according to the trigger position and time sequence correlation of the key event, in combination with the single-hand effective operation area, determine the valid key points within the single-hand stretching range, and generate positive sample labeling.

[0078] Wherein, the time sequence correlation refers to the continuity and logic of a key event with other key events before and after it in time. The single-hand effective operation area refers to the keyboard area that a single hand (usually the left hand) can comfortably and quickly reach while maintaining control of the mouse, for example, the nearby key positions centered on WASD. The single-hand stretching range refers to the farthest distance that the fingers can reach from the core area. The positive sample labeling refers to labeling data points that meet the valid operation logic as valid (for example, labeled as 1).

[0079] Specifically, after the negative samples are labeled in step S202, this step focuses on identifying and labeling positive samples from the remaining data. The user terminal first defines a standard single-hand valid operation area (for example, the area surrounded by from to T, from Tab to G, and from Caps to B on the left side of the keyboard). For key events in the second time-series touch data that are not labeled as negative samples, the user terminal makes a judgment: first, if the position of a key event falls within the core area and its pressure, duration, and other characteristics conform to the normal key mode, it is highly likely to be labeled as a positive sample. Second, for keys that fall outside the core area, the user terminal checks their spatio-temporal relationship with the previous key event. If the distance between the key and the previous positive sample key is within the single-hand stretch range, and the time interval is reasonable (not too long or too short), the key is also considered part of the continuous operation and is labeled as a positive sample.

[0080] It should be noted that labeling positive samples based on temporal correlation and ergonomic features is a state and rule-based decision-making process. First, the system will determine an "anchor area" based on the application scenario (such as learning from the process mapping table that it is currently a certain game), which is usually the most core movement key area in the game, such as "WASD". Any event occurring in this anchor area and having physical characteristics (such as pressure, duration) conforming to normal key presses will be unconditionally prioritized and labeled as a positive sample, and will serve as the starting point for subsequent correlation judgments. For key events occurring outside this area, the system will initiate a correlation checking logic. It will backtrack to find the most recent key event that has been labeled as a positive sample, and calculate the physical spatial distance and time interval between the current key and the previous positive sample key. The system maintains a "single-hand maximum stretch distance" (for example, 15 centimeters, which can be personalized set by the calibration step S206) and a "continuous hit effective time window" (for example, 300 milliseconds). Only when the calculated spatial distance is less than the maximum stretch distance and the time interval is less than the continuous hit effective time window, the system will consider this key press as a continuous action of the previous effective operation, and thus label it as a positive sample. For example, while playing a game, the player presses the "W" key (within the anchor area, labeled as a positive sample), and then presses the "G" key 150 milliseconds later. If the distance between "W" and "G" is less than the user's single-hand maximum stretch distance, then "G" will also be labeled as a positive sample. This logic chain will continue to pass down, accurately identifying a complete set of continuous single-hand operation sequences.

[0081] In some embodiments, the positive sample labeling in this step can be achieved in various ways: alternatively, the user terminal can build a rule-based expert system. The system contains a series of rules, such as if the key is one of W, A, S, D, and the pressure is greater than P_normal, then label it as positive sample, if the current key is R, and the previous key is W, and the time difference between the two is less than T_combo, then label R as positive sample, etc. The data is automatically labeled by these rules; alternatively, the user terminal can use a semi-supervised learning method, first manually label a small part of the positive sample, then use these samples to train a simple classifier, and then use the classifier to predict the label of the remaining data points, and the prediction result with high confidence is used as the final positive sample label. It can be understood that the labeling process of the positive sample can combine various heuristic rules and machine learning models to achieve the best labeling effect, which is not limited here.

[0082] In some embodiments, there may be different core operation areas of different games or applications, for example, some games may use ESDF as the movement key. In this case, the user terminal will associate the currently running application when determining the single-handed effective operation area. The user terminal maintains a process-keyboard hot area mapping table internally. Before labeling the positive sample, the user terminal will detect the process running in the foreground (such as csgo.exe) during data collection, and query the core key area corresponding to the game (such as WASD, Shift, Ctrl, Space, etc.) from the mapping table. Then based on this specific core area, the subsequent judgment logic is executed. This labeling strategy bound to the application scenario makes the identification of positive samples more accurate, thereby providing a high-quality data basis for training a single-handed model with stronger scene adaptability.

[0083] S204, extracting the position parameter, pressure value, duration, and spatial distance from the previous key event of each key point in the key event to generate a feature vector.

[0084] Among them, the position parameter can be a two-dimensional coordinate (x, y) representing the position of the key on the keyboard matrix. The pressure value can be the pressure peak value during the key pressing process, or the pressure average value within a period of time. The duration is the time elapsed from the key being pressed to being lifted. The spatial distance from the previous key event is the Euclidean distance between the center of the current key and the center of the last triggered key. The feature vector is a one-dimensional array composed of these extracted numerical features in a specific order, which converts a key event and its context information into a mathematical form that can be processed by a machine learning model.

[0085] Specifically, for each key press event marked in steps S202 and S203, the user terminal will perform a feature extraction procedure. The procedure first parses the basic physical quantities from the raw touch data:

[0086] 1) the (x, y) coordinates of the key press;

[0087] 2) the maximum pressure value during the whole pressing process;

[0088] 3) the total time from when the pressure exceeds the trigger threshold to when it falls below the release threshold.

[0089] Subsequently, the user terminal will query the history record to find the previous key press event immediately before the current event and calculate the straight-line distance between the two key press coordinates. In addition, more derived features can be extracted, such as the pressure rise rate, shape parameters of the pressure curve, etc. Finally, all these numerical values (e.g., x, y, pressure_max, duration, distance_to_prev) are concatenated together to form a fixed-dimension feature vector.

[0090] In some embodiments, the feature vector generation in this step can be implemented in various ways: optionally, the user terminal can generate a vector containing the basic features, such as [x, y, pressure_peak, duration, time_since_prev, dist_to_prev]; optionally, in order to capture more rich dynamic information, the user terminal can extract statistical features of the pressure sequence, such as performing Fourier transform on the pressure value sequence during the pressing process to extract the coefficients of its low-frequency components, or calculating the mean, variance, kurtosis, etc. of the pressure sequence, and adding these statistics to the feature vector, so that the model can perceive the subtle differences in key press techniques. It can be understood that the selection of feature engineering is crucial to the performance of the model, and features can be designed and added or deleted according to specific needs, which is not limited here.

[0091] In some embodiments, there can be slight differences in the physical characteristics (such as spring pressure) of keys at different positions on the keyboard, resulting in different pressure values under the same force. For this, the user terminal will perform normalization when extracting the pressure value feature. The user terminal will store a keyboard calibration mapping table internally, which records the reference pressure value of each key or each region. After extracting the pressure value of a key, it will be divided by the reference value corresponding to the key to obtain a relative pressure value. Using this normalized pressure feature, the impact of keyboard physical differences can be eliminated, making the rules learned by the model more universal and not producing judgment bias due to different key positions.

[0092] S205, input the feature vector and the corresponding positive sample label and negative sample label to a neural network for training to obtain the single-hand mis-touch recognition model.

[0093] Wherein, the corresponding positive sample label and negative sample label refer to the label determined for each feature vector in steps S202 and S203 (such as 1 representing validity and 0 representing mis-touch).

[0094] Specifically, after a large number of feature vectors and their corresponding positive and negative sample labels are prepared, the user terminal divides the data set into a training set, a validation set and a test set. Then, a neural network model is constructed, the number of nodes of the input layer of which is the same as the dimension of the feature vector, and the output layer usually has one node (using a Sigmoid activation function, outputting a probability value of 0 to 1). In the training phase, the user terminal sends the feature vectors of the training set into the network in batches (batch), calculates the loss (such as cross-entropy loss) between the network output and the real label. Then, according to the loss value, the gradient of the loss to the weights of each layer of the network is calculated using the back propagation algorithm, and the weights are updated in the direction of gradient descent using the Adam, SGD and other optimizers. This process is repeated for multiple rounds (epochs) until the performance of the model on the validation set reaches convergence or meets the preset conditions.

[0095] In some embodiments, the model training in this step can be implemented in various ways: optionally, the user terminal can use a standard multi-layer perceptron (MLP) as the neural network structure. The network contains several fully connected hidden layers, uses a nonlinear activation function such as ReLU, has a simple structure and fast training speed, and is suitable for processing the extracted feature vectors; optionally, if the model is expected to automatically learn the complex interaction between features, a more complex network structure can be used, for example, part of the original time series data (such as pressure sequence) and the extracted feature vectors are processed separately, the former is extracted by a small one-dimensional CNN or RNN, the latter is directly input, then the extracted features of the two parts are spliced and then classified through a fully connected layer. It can be understood that the structure and training strategy of the neural network can be diversified to balance the model performance and computational overhead, which is not limited here.

[0096] In some embodiments, there is a risk of overfitting in the model training process, i.e. the model performs perfectly on the training set but performs very poorly on new data that has not been seen before. In this regard, the user terminal will employ various regularization techniques to prevent overfitting during the training process. For example, a Dropout layer is added after the hidden layer of the network, a portion of neurons is randomly discarded in each training iteration, forcing the network to learn more robust features; or an L1 or L2 regularization term is added to the loss function, which punishes the size of the network weights and limits the complexity of the model. At the same time, the user terminal will closely monitor the performance of the model on the validation set, and once it is found that the validation set loss no longer decreases or even starts to rise, the EarlyStopping mechanism will be triggered to stop training and save the current optimal model to obtain the best generalization ability.

[0097] S206、In the running of the user terminal, the target displacement signal of the connected mouse device is monitored.

[0098] Reference step S103, which is not described here.

[0099] In some embodiments, the user terminal will further improve the personalization level of the single-hand mis-touch recognition model by executing a personalized calibration process before the model switching logic is judged, i.e. the user terminal will display a single-hand operation calibration interface containing a preset key sequence to collect calibration touch data generated during the user pressing the preset key sequence; calculate the maximum spatial distance between each key point in the calibration touch data to obtain the target single-hand stretching range of the user; take the key position with the highest key frequency in the calibration touch data as the center and the target single-hand stretching range as the radius to determine the target single-hand operation area of the user; and correct the single-hand mis-touch recognition model based on the target single-hand stretching range and the target single-hand operation area.

[0100] Wherein, the preset key sequence is a key combination that needs to be completed by the user's single hand, for example QAZWSXEDCRFV. The calibration touch data is the keyboard input data collected when the user performs the task. The target single-hand stretching range is a value calculated from the data, representing the limit activity radius of the fingers of the specific user. The target single-hand operation area is the personalized hot area of the user determined based on the stretching range and the most commonly used key position.

[0101] Specifically, when the user first uses the function or actively triggers in the settings, the user terminal will pop up a calibration wizard. The wizard interface will prompt the user to maintain a single-handed operation posture (such as holding the mouse with the right hand and placing the left hand on the keyboard), and press the key sequence displayed on the screen in turn. During this process, the user terminal records the coordinates of all the keys. After calibration, the user terminal traverses all the collected continuous key pairs, calculates the Euclidean distance between them, and takes the maximum value as the target single-handed stretching range of the user. At the same time, the frequency of all keys is counted, and the key with the highest frequency (usually the starting key or core key of the sequence) is found, and its position is taken as the anchor point or center of the user's single-handed operation. With the center and the calculated stretching range as the radius, a circular target single-handed operation area that meets the user's habits can be defined. Finally, the user terminal updates these two parameters (stretching range, operation area) to the configuration of the single-handed mis-touch recognition model, and the model will preferentially use these personalized parameters in subsequent judgments, such as in the judgment of the context association of an effective operation key.

[0102] It should be noted that the single-handed mis-touch recognition model is corrected based on the user's personalized calibration data, which can be implemented through two levels of technical paths. The first is parameterized correction, which is a lightweight adaptation method. In this way, the single-handed mis-touch recognition model will call some ergonomics-related external parameters, such as "single-handed maximum stretching distance" and "single-handed operation area center point coordinates", when performing feature extraction or internal judgment. When not calibrated, these parameters use a general default value based on a large number of user statistics. After the user completes the personalized calibration process, the system will replace these default parameters called by the model with the calculated "target single-handed stretching range" and the center point coordinates of the "target single-handed operation area". In this way, when the model subsequently judges whether a key belongs to an effective continuous operation, the basis for its spatial distance judgment becomes the user's actual data, thus achieving personalization. The second is model fine-tuning, which is a more in-depth adaptation. The system will collect user key data (such as data generated by pressing the QAZWSXEDCRFV sequence) during the calibration process as a batch of high-quality, user-specific positive samples. Then, the system will unfreeze some or all network layers of the pre-trained single-handed mis-touch recognition model and use these personal samples to perform additional training on the model for several rounds. A very small learning rate is used during training to fine-tune the model's weights to the user's unique key habits (such as specific key spatio-temporal association patterns) without damaging the model's general knowledge learned from massive data (such as the ability to judge pressure and duration), so that the model itself is more personalized to the user's individual characteristics.

[0103] In some embodiments, there can be cases where the user does not operate the calibration process properly, resulting in noisy data collected, and inaccurate calculated stretch range or operation area. In this case, the user terminal can clean the calibration touch data before calculation. For example, the user terminal can calculate the duration of each key press and the time interval between adjacent key presses, and exclude abnormal data points with too long duration (possibly a pause in the middle) or too large time interval (possibly distracted) from the calculation. In addition, the user terminal can ask the user to repeat the calibration process several times, and then take the average or median of the calculated stretch range and operation area each time to obtain a more stable and reliable personalized parameter, reducing the error caused by a single accidental operation.

[0104] S207, taking the two-hand mis-touch recognition model as the current effective recognition model.

[0105] Specifically, when the user terminal starts or the current key intelligent response function is activated, the system will be initialized. Considering that in most non-game applications, two-hand typing is a more common operation mode, therefore, setting the two-hand mis-touch recognition model as the initial, default current effective recognition model is a safe and reasonable choice. This means that before the system collects enough evidence that the user is performing single-hand operation, all keyboard inputs will be judged by the two-hand mis-touch recognition model. The establishment of this initial state provides a clear starting point for subsequent dynamic switching.

[0106] In some embodiments, the initialization of this step can be implemented in various ways: optionally, when the relevant service of the user terminal starts, a pointer or identifier pointing to the two-hand mis-touch recognition model instance is directly assigned to a global current model variable; optionally, the user terminal can determine the initial model according to the last state before the user last shut down or logged out. For example, if the user last turned off the user terminal while playing a game, the single-hand model can be loaded as the initial model when starting next time, but this requires an additional state persistence mechanism. It can be understood that starting with the two-hand model by default is a more general and simple implementation, which is not limited here.

[0107] In some embodiments, there can be cases where the user's main use scenario is single-hand operation (for example, a professional game player or designer). Defaulting to the two-hand model every time the user terminal starts and waiting for the system to switch may cause misjudgment in the initial stage. In this case, the user terminal can provide an option for the default start mode in the settings, allowing the user to set the default model to automatic judgment, always two-hand mode or always single-hand mode. If the user selects always single-hand mode, the user terminal will take the single-hand mis-touch recognition model as the current effective recognition model in step S207, thereby better adapting to the fixed use habits of specific users.

[0108] S208, collect displacement signals and keyboard key timestamp in a preset time period, calculate the cumulative displacement of the displacement signals and the key frequency.

[0109] Wherein, the preset time period is a time window for state judgment, for example, 200 milliseconds. The cumulative displacement is the sum of the absolute values of all X and Y direction displacement of the mouse in the time period. The key timestamp is the accurate time point of each key event. The key frequency is the number of key events in the time period divided by the period length.

[0110] Specifically, the monitoring module of the user terminal segments and aggregates the data of the input device in a fixed time interval (i.e. the preset time period). In a time period, it will: 1) accumulate the absolute values of all displacement signals received from the mouse to obtain the total displacement in the period; 2) count the number of all key events received from the keyboard. At the end of the period, the instantaneous key frequency (unit: times / sec) in the time window is calculated by dividing the total number of keys by the period length (such as 0.2 seconds). These two calculated values, cumulative displacement and key frequency, will be important input parameters for subsequent scene switching decisions.

[0111] In some embodiments, the calculation of this step can be implemented in various ways: alternatively, the user terminal can maintain two counters, one for accumulating displacement and one for recording the number of keys. Every other preset time period, the values of the two counters are read, calculated, and then the counters are cleared to start the statistics of the next period; alternatively, in order to obtain smoother results, the user terminal can use the sliding window method. The window slides forward at a fixed step, and the cumulative displacement and key frequency in the window are recalculated each time the window slides. This way can provide more continuous and timely response state evaluation. It can be understood that the specific calculation implementation can be adjusted according to the demand for real-time and computing resources, which is not limited here.

[0112] In some embodiments, there may be a case where the mouse generates extremely weak jitter displacement signals due to the slight vibration of the desktop or sensor noise, which may cause the cumulative displacement to be non-zero, thereby interfering with the judgment. For this purpose, the user terminal will perform a noise filtering on the original displacement signal before calculating the cumulative displacement. For example, a very small single movement threshold (such as 1 pixel) is set, and any single displacement event less than the threshold will be ignored and not included in the cumulative displacement. In this way, hardware noise can be effectively filtered out, ensuring that only user-conscious mouse movements are counted, improving the reliability of cumulative displacement as a basis for judgment.

[0113] S209. Extract the position coordinates of each key point in the preset time period, and calculate the spatial distribution variance of the position coordinates.

[0114] The spatial distribution variance is a statistical quantity for measuring the degree of dispersion of the set of coordinate points in space. The greater the variance, the more extensive and dispersed the key position distribution; the smaller the variance, the more concentrated the key position.

[0115] Specifically, synchronously with step S208, in the same preset time period, the user terminal not only records the number of keystrokes, but also records the precise keyboard coordinates of each keystroke. At the end of a period, if keystrokes have occurred in the period (the number of times is greater than 0), the user terminal will collect all the key coordinates in this period. Then, the centroid of the set of coordinates (i.e., the average of all x coordinates and the average of all y coordinates) is calculated. Next, the square of the distance of each coordinate point to the centroid is calculated, and the average of all these squared distances is calculated, which is the spatial distribution variance. This value reflects whether the user's keystroke behavior is concentrated in a small area (such as the WASD area when playing single-handed games) or dispersed throughout the keyboard (such as when typing with both hands).

[0116] In some embodiments, the calculation of this step can be implemented in various ways: alternatively, the user terminal can calculate the variance of the X direction coordinates and the variance of the Y direction coordinates respectively, and then add them together as the total spatial distribution variance; alternatively, to simplify the calculation, instead of calculating the strict variance, the area of the minimum bounding box formed by all the key points, or the distance between the two farthest key points, can be calculated as an approximate measure of the spatial distribution dispersion. It can be understood that any statistical indicator that can effectively quantify the concentration of key positions can be used in this step, which is not limited here.

[0117] In some embodiments, there may be a situation where the number of keystrokes in the period is too small (for example, only 1 or 2 times), resulting in a variance that is not statistically meaningful. For this, the user terminal sets a minimum number of keystrokes threshold (for example, 3 times) when calculating the spatial distribution variance. Only when the number of keystrokes in the period reaches or exceeds the threshold, the variance calculation is performed. If the number of keystrokes is insufficient, the spatial distribution variance feature is marked as invalid or assigned a default value. This can avoid misleading the final decision based on unreliable variance values derived from a small number of samples, ensuring the effectiveness of the feature.

[0118] S210. When the accumulated displacement amount exceeds the preset mouse activity threshold, the key frequency is lower than the preset two-handed input threshold, or the spatial distribution variance is lower than the preset concentration threshold, change the two-handed accidental touch recognition model in the currently effective recognition model to the single-handed accidental touch recognition model.

[0119] Among them, the preset mouse activity threshold, the preset two-hand input threshold, and the preset concentration threshold are three key decision boundary parameters. The logic here is an or relationship, that is, any one of the three conditions is met, which will trigger the model switching. Changing the model means switching the pointer or identifier of the currently effective recognition model in the system from the two-hand model to the single-hand model.

[0120] Specifically, at the end of each preset time period, the decision module of the user terminal obtains the three core indicators calculated in steps S208 and S209: the accumulated displacement amount, the key frequency, and the spatial distribution variance. The decision module performs a composite logical judgment:

[0121] Check whether the accumulated displacement amount > preset mouse activity threshold is true? (indicating that the mouse is actively used)

[0122] Check whether the key frequency < preset two-hand input threshold is true? (indicating that the input intensity is not like high-speed typing)

[0123] Check whether the spatial distribution variance < preset concentration threshold is true? (indicating that the keys are highly concentrated in a small range)

[0124] If any one of the above three conditions is true, the decision module considers that the current user is most likely in a single-hand operation scenario. Therefore, it performs a switching operation to change the currently effective recognition model from the two-hand model to the single-hand model.

[0125] In some embodiments, the switching logic of this step can be implemented in various ways: optionally, the user terminal can use a weighted voting mechanism. A weight is set for each condition, and then a single-hand tendency score is calculated according to whether each condition is met and the degree of satisfaction. When the score exceeds a certain threshold, switching is performed, which is more flexible; optionally, the user terminal can define a state machine including a two-hand state, a single-hand state, and a transition state, and the combination of the three conditions determines the conversion between states, making the switching process smoother and more controllable. It can be understood that the specific decision logic can be adjusted according to the actual test effect to achieve the best switching performance, which is not limited here.

[0126] In some embodiments, there can be undefined cases for switching back from single-hand mode to double-hand mode. In this regard, the user terminal can define a symmetric switching back roll logic. For example, the model is switched back to double-hand mode only when and if the accumulated displacement amount is below the mouse inactivity threshold for N consecutive periods and the key frequency is above the double-hand input threshold and the spatial distribution variance is above the concentration threshold. Using and logic and stricter conditions can ensure that the model is switched back only when the user explicitly goes back to double-hand typing state, avoiding false switching back during short inactivity periods such as game pause. This asymmetric switching logic (lenient in-switching condition and strict in-switching condition) can improve the stability of the state.

[0127] In some embodiments, the user terminal can introduce application context information to make the scene switching more intelligent, i.e., the user terminal reads the foreground process identifier of the current running foreground process of the user terminal, determines the single-hand scene biasing coefficient corresponding to the foreground process identifier according to a process scene mapping table, and adjusts the preset mouse activity threshold and the preset concentration threshold according to the single-hand scene biasing coefficient before performing the switching judgment.

[0128] The foreground process identifier is the unique identification of the application program currently directly interacted with the user, such as the process name explorer.exe or the window title. The process scene mapping table is a preset database or configuration file, which maps the known process identifier to a specific scene type (such as game, office, browsing) and a quantitative single-hand scene biasing coefficient. The coefficient (for example, ranging from -1 to 1) indicates how large the probability of the application scene being a single-hand operation mode (positive value indicates biasing to single hand, negative value biasing to double hand). Adjusting the threshold means dynamically raising or lowering the boundary value used for decision-making in step S210.

[0129] Specifically, the monitoring module of the user terminal will periodically (for example, once every second) obtain the foreground window and the corresponding process identifier of the current operating system in addition to monitoring the input device. Then, it will use the process identifier to query the internal process scene mapping table. For example, when the foreground process is detected to be League of Legends.exe, the corresponding single-hand scene biasing coefficient obtained by querying the mapping table can be +0.8, indicating that it is a strong single-hand operation scene. Conversely, if it is detected to be WINWORD.EXE (Microsoft Word), the biasing coefficient can be -0.7. After obtaining the coefficient, the user terminal will use it to adjust the two key thresholds for scene switching: the preset mouse activity threshold and the preset concentration threshold. The adjustment formula can be new threshold = original threshold * (1-biasing coefficient). In this way, in a strong single-hand biasing scene (coefficient is positive), the threshold will be lowered, making it easier for the system to switch to the single-hand model; conversely, in a strong double-hand biasing scene, the threshold will be raised, making it more difficult to switch.

[0130] In some embodiments, the threshold adjustment in this step can be implemented in various ways: optionally, the linear multiplication formula described above is used for adjustment, which is simple and intuitive; optionally, the user terminal can preset several different threshold configuration schemes, respectively corresponding to game mode, office mode, standard mode, etc., and the process scene mapping table directly maps the process to one of the modes, and the system loads the complete set of thresholds corresponding to the mode, which is easier to manage and debug. It can be understood that the specific strategy of threshold adjustment can be designed according to the experimental effect, as long as it can reflect the adaptability based on the application scenario, which is not limited here.

[0131] In some embodiments, there may be a case that the process scene mapping table is incomplete and cannot judge the bias of a new application or game. In this case, the user terminal will have the ability to learn and update the mapping table. When a new process not in the table is encountered, the user terminal will temporarily use a neutral bias coefficient (such as 0). But in the subsequent running, the user terminal will count the single-handed / double-handed mode residence time ratio judged by the pure physical signal of step S210 during the process as the foreground process. If it is found that the user is in single-handed mode 90% of the time under this application, the user terminal will automatically generate a higher single-handed scene bias coefficient (such as +0.7) for this new process and store it in the mapping table. In this way, the mapping table can be continuously expanded and improved in use, and the intelligent level continues to improve.

[0132] S211, input the real-time touch data generated by the current user into the currently effective recognition model for calculation to obtain a key effective score.

[0133] Reference step S105, which is not repeated here.

[0134] S212, when the key effective score is lower than a preset effective threshold, the corresponding error key signal is prevented.

[0135] Reference step S106, which is not repeated here.

[0136] In some embodiments, the user terminal will start an online updating process to enable the model to have the ability of continuous learning and self-optimization after blocking a key once, i.e. the user terminal will record the key identification corresponding to the blocked key signal and the blocking time; when receiving a repeated key signal with the same key identification within a preset time window after the blocking time, the false key signal is marked as a misjudgment sample and stored in a correction sample library; when the number of samples in the correction sample library reaches a predetermined value, the misjudgment feature vector of the misjudgment sample is extracted; based on the misjudgment feature vector and the corresponding positive sample label, the weight update amount of the misjudgment feature vector to the current effective recognition model is calculated through the gradient descent algorithm, and the weight update amount is loaded to the current effective recognition model for incremental updating.

[0137] wherein the misjudgment sample refers to a key that is wrongly judged as a false touch and blocked by the model, but is actually a key of the user's real intention. The correction sample library is a buffer for storing these misjudgment samples.

[0138] Specifically, whenever a key signal is blocked due to a too low score, the user terminal will not discard all its information immediately, but will temporarily store its key identification (such as A key) and the time stamp of occurrence. At the same time, a timer is started to monitor a short preset time window (e.g. 500 milliseconds). If within this window, the user terminal receives a key signal exactly the same as the blocked key (the user repeats the key because he did not react the first time), the user terminal highly suspects that the previous blocking is a misjudgment. Therefore, it will mark the complete touch data of the first blocked key event as a misjudgment sample (i.e. should be a positive sample but is judged as negative), and store it in the correction sample library. When the number of misjudgment samples accumulated in this library reaches a predetermined value (e.g. 50), the user terminal will trigger an incremental update. It will extract the feature vectors of these samples, and use them as input and the effective key (positive sample) as expected output to perform several steps of gradient descent on the current effective model to calculate the small adjustment amount of the weight, and then apply this adjustment amount to the model to complete an online learning.

[0139] It should be noted that when implementing incremental updating of the model, in order to ensure the effectiveness and stability of learning, when a sufficient number of misjudgment samples (e.g. 50) are accumulated in the correction sample library, the system does not simply use these new samples for regular training. In order to prevent catastrophic forgetting, i.e. the model forgets the general knowledge learned from large-scale offline data after learning new samples, the system will adopt a mechanism similar to "experience replay". Specifically, when triggering incremental updating, the system not only takes out the 50 misjudged new samples from the correction sample library, but also randomly extracts a small number (e.g. 200) of "old" samples from the original large-scale data set used for offline training. Then, these new and old samples are mixed together to form a temporary training batch. The system uses this mixed batch to update the current effective model for several steps of gradient descent. During the updating process, a learning rate much smaller than that during offline training is used, which is a key hyperparameter, and the purpose is to let the model "review" the old knowledge while "correcting" the wrong judgment on the new samples, so as to make only a small and targeted adjustment to the weights. For example, if a blocked light key is repeatedly pressed by the user to confirm that it is valid, then this incremental update will slightly adjust the model weights so that in the future it will slightly increase the score of similar light keys, but not enough to judge all light touches as valid, thereby achieving a balance between correcting errors and maintaining generalization ability.

[0140] In some embodiments, there may be cases where the user does not repeat the key because it is blocked, but his operation habit is fast double-clicking, or the first key press is indeed a false touch, and the second one is the valid key. Directly regarding the first blocked key as a misjudgment sample will pollute the sample library. For this, the user terminal will introduce more stringent screening conditions when marking misjudgment samples. This will be described in detail in the next embodiment, by comparing the physical characteristics of the two key presses to more accurately determine the user's true intention, and ensure that only high-quality misjudgment samples are used for model updating.

[0141] In some embodiments, the user terminal will refine the sample labeling step in the online updating process in order to improve the discrimination accuracy of the misjudgment samples, that is, the user terminal will input the touch data corresponding to the error button signal and the touch data corresponding to the repeated button signal into the current effective recognition model when the same repeated button signal as the button identification is received within the preset time window after the blocking moment, to obtain a first score and a second score; calculate the absolute value of the difference between the first score and the second score; when the absolute value of the difference is less than the preset fluctuation threshold, the touch data corresponding to the error button signal is marked as a misjudgment sample and stored in the correction sample library; when the absolute value of the difference is greater than or equal to the fluctuation threshold and the second score is higher than the preset effective threshold, the time point corresponding to the pressure peak in the error button signal and the repeated button signal is extracted and the time difference is calculated; when the time difference is less than the preset time length threshold, the touch data corresponding to the error button signal is marked as a misjudgment sample and stored in the correction sample library.

[0142] Wherein, the first score and the second score are respectively the effectiveness evaluation of the blocked button and the subsequent repeated button by the model. The absolute value of the difference reflects the similarity of the user's two button pressing methods. The time point corresponding to the pressure peak is the moment when the pressure reaches the maximum in each button pressing process. The time difference is the time interval between the pressure peaks of the two button presses, reflecting the urgency of the user's repeated button pressing.

[0143] Specifically, when a repeated button press is detected, the user terminal will not immediately mark the previous blocked button as a misjudgment. Instead, a more detailed analysis process is performed: first, the touch data of the two button presses is sent to the current model to calculate the scores, obtaining a first score (low score) and a second score. Then discuss in two cases:

[0144] If the absolute value of the difference between the two scores is very small (less than the preset fluctuation threshold), it indicates that the user's two button pressing methods are very similar, but the second time is recognized as valid (or even if it is blocked, but the user is still repeating), which strongly suggests that the user's intention is consistent, and the model's judgment of this method is unstable or wrong. At this time, the first button press is marked as a misjudgment sample.

[0145] If the difference between the two scores is large, and the second score is much higher than the effective threshold, it indicates that the user's second button pressing action is clearer and more standard. At this time, the user terminal will further check the time difference between the pressure peaks of the two button presses. If the time difference is very short (less than the preset time length threshold, such as 150 milliseconds), it indicates that the user immediately and decisively pressed the button again after the first button press failed. This behavior pattern is also a strong signal of misjudgment. At this time, the first button press is also marked as a misjudgment sample. Only when one of the above two complex conditions is met, it is finally confirmed as a misjudgment sample and stored in the correction library.

[0146] In some embodiments, the feature comparison in this step can be implemented in various ways: alternatively, in addition to comparing the score difference and the time difference, more physical features of the two key presses can be compared, such as the touch area, the duration, etc., to construct a feature difference vector, and then a special misjudgment discrimination classifier can be trained to determine whether it is a real misjudgment; alternatively, a dynamic time warping (DTW) algorithm can be introduced to calculate the similarity between the pressure sequences of the two key presses, and DTW can better handle the slight stretching and shifting of the two key presses on the time axis, thereby more robustly measuring the similarity of the key pressing manners. It can be understood that the logic of discriminating misjudgment samples can be designed to be more complex and accurate as needed to maximize the quality of the correction samples, which is not limited here.

[0147] In some embodiments, the preset fluctuation threshold and duration threshold may not be universally applicable and may be too strict or too loose for some users. In this case, the user terminal can adaptively adjust these thresholds. For example, the user terminal can count the distribution of the score difference and the time difference of the samples in which a user is confirmed to be a real misjudgment. Then, the fluctuation threshold and the duration threshold can be dynamically adjusted according to the statistical quantity (such as the 80th percentile) of this distribution. In this way, the thresholds can gradually adapt to the unique error correction behavior pattern of each user, so that the misjudgment sample discrimination logic is also personalized, further improving the efficiency and accuracy of online learning.

[0148] In the embodiments of the present application, since the intelligent response method based on the mouse activity state is adopted to distinguish single-handed and double-handed operation scenarios and dynamically call the mis-touch recognition model matched therewith, the key effectiveness can be determined according to the actual use context of the user. This method effectively solves the problem that the single fixed strategy in the prior art cannot take into account the huge difference in mis-touch features in different scenarios (such as typing and gaming), resulting in a high misjudgment rate, and further realizes the recognition and filtering of various mis-touch signals (such as palm pressure and scratching) on a high-sensitivity keyboard, improving the input accuracy and overall experience of the user in various application scenarios.

[0149] The user terminal in the embodiments of the present application will be described from the perspective of hardware processing. Please refer to Figure 3 , which is a schematic diagram of an entity device structure of the user terminal in the embodiments of the present application.

[0150] It should be noted that Figure 3 The structure of the user terminal shown is only an example and should not impose any limitation on the functions and use range of the embodiments of the present application.

[0151] As Figure 3As shown, the user terminal includes a CPU 301 which can perform various appropriate actions and processes in accordance with a program stored in a ROM 302 or a program loaded into a RAM 303 from the storage section 308, for example, the method described in the above-described embodiments. In the RAM 303, various programs and data required for system operation are also stored. The CPU 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. An I / O interface 305 is also connected to the bus 304.

[0152] Connected to the I / O interface 305 are an input section 306 including an audio input device, a push button switch, and the like; an output section 307 including a Liquid Crystal Display (LCD), an audio output device, an indicator, and the like; a storage section 308 including a hard disk and the like; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, a modem, and the like. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the I / O interface 305 as necessary. A removable media 311 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is attached to the drive 310 as necessary, so that a computer program read therefrom is installed into the storage section 308 as necessary.

[0153] In particular, the processes described above with reference to the flow charts can be implemented as computer software programs in accordance with embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing computer programs for executing the methods illustrated by the flow charts. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 309, and / or installed from the removable media 311. When the computer program is executed by the CPU 301, various functions defined in the present application are performed.

[0154] The flow charts and block diagrams in the drawings are schematic illustrations of possible architectures, functions, and operations of systems, methods, and computer program products in accordance with various embodiments of the present application. In this regard, each block in the flow charts or block diagrams can represent a module, a segment, or a portion of code which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures.

[0155] In particular, the user terminal of the present embodiment includes a processor and a memory, and the memory stores a computer program which, when executed by the processor, implements the intelligent key response method provided by the above-described embodiments.

[0156] As another aspect, the application also provides a computer readable storage medium, which can be included in the user terminal described in the above embodiments, or can exist separately without being assembled into the user terminal. The storage medium carries one or more computer programs, which, when executed by a processor of the user terminal, enable the user terminal to implement the intelligent response method of the keys provided in the above embodiments.

[0157] The above-described embodiments are only used to illustrate but not to limit the technical solutions of the present application; even though the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: they can still make modifications to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0158] In the above embodiments, according to the context, the term "when" can be interpreted as meaning "if" or "after" or "in response to determining" or "in response to detecting". Similarly, according to the context, the phrase "upon determining" or "if detecting (the stated condition or event)" can be interpreted as meaning "if determining" or "in response to determining" or "upon detecting (the stated condition or event)" or "in response to detecting (the stated condition or event)".

Claims

1. A method for intelligent button response, characterized in that, Applied to a user terminal, the method includes: Based on the first time-series touch data obtained from sample users during the period when the mouse device did not generate a displacement signal, a two-handed accidental touch recognition model was trained. A single-handed accidental touch recognition model is trained based on second-sequence touch data collected from sample users during the period when the mouse device generates displacement signals. The steps of training the single-handed accidental touch recognition model based on the second-sequence touch data collected from sample users during the period when the mouse device generates displacement signals specifically include: performing clustering calculations on the second-sequence touch data collected from sample users during the period when the mouse device generates displacement signals to determine accidental touch key points with preset large-area features and preset low-pressure features, generating negative sample labels; determining effective key points within the single-handed reach range based on the trigger position and temporal correlation of key events, combined with the effective single-handed operation area, generating positive sample labels; extracting the position parameters, pressure values, duration, and spatial distance from the previous key event for each key point in the key event to generate feature vectors; and inputting the feature vectors and the corresponding positive and negative sample labels into a neural network for training to obtain the single-handed accidental touch recognition model. During user terminal operation, monitor the target displacement signal of the connected mouse device; Based on the target displacement signal, one of the two-handed accidental touch recognition models and the one-handed accidental touch recognition model is determined as the currently effective recognition model; The real-time touch data generated by the current user is input into the currently effective recognition model for calculation to obtain the key position validity score; When the valid score of a key is lower than a preset valid threshold, the corresponding erroneous key press signal is blocked.

2. The method according to claim 1, characterized in that, Prior to the step of monitoring the target displacement signal of the connected mouse device during operation on the user terminal, the method further includes: Display a one-handed operation calibration interface containing a preset key sequence to collect calibration touch data generated when the user presses the preset key sequence; Calculate the maximum spatial distance between each key point in the calibration touch data to obtain the user's target single-hand extension range; Using the key position with the highest key frequency in the calibrated touch data as the center and the target single-hand extension range as the radius, the user's target single-hand operation area is determined; Based on the target single-hand extension range and the target single-hand operation area, the single-hand accidental touch recognition model is corrected.

3. The method according to claim 1, characterized in that, The step of determining one of the two-handed accidental touch recognition models and the one-handed accidental touch recognition model as the currently effective recognition model based on the target displacement signal specifically includes: The aforementioned two-hand accidental touch recognition model is used as the currently effective recognition model; Collect the displacement signal and the key press timestamp within a preset time period, and calculate the cumulative displacement and key press frequency of the displacement signal; Extract the position coordinates of each key point within the preset time period, and calculate the spatial distribution variance of the position coordinates; When the cumulative displacement exceeds a preset mouse activity threshold, the key press frequency is lower than a preset two-hand input threshold, or the spatial distribution variance is lower than a preset concentration threshold, the two-handed accidental touch recognition model in the currently effective recognition model is changed to the one-handed accidental touch recognition model.

4. The method according to claim 3, characterized in that, Before the step of changing the currently effective recognition model from the two-handed accidental touch recognition model to the one-handed accidental touch recognition model, the method further includes: Read the identifier of the foreground process currently running on the user terminal, and determine the single-handed scene bias coefficient corresponding to the foreground process identifier according to the process scene mapping table; Based on the single-handed scenario bias coefficient, adjust the preset mouse activity threshold and the preset concentration threshold.

5. The method according to claim 1, characterized in that, After the step of blocking the corresponding erroneous key signal when the valid key score is lower than a preset valid threshold, the method further includes: Record the key identifier and the time of blocking corresponding to the blocked key signal; When a repeated key signal with the same key identifier is received within a preset time window after the blocking time, the erroneous key signal is marked as a misjudged sample and stored in the correction sample library; When the number of samples in the correction sample library reaches a predetermined value, the misjudgment feature vector of the misjudged sample is extracted; Based on the misjudged feature vector and the corresponding positive sample label, the weight update amount of the misjudged feature vector to the currently effective recognition model is calculated by the gradient descent algorithm, and the weight update amount is loaded into the currently effective recognition model for incremental update.

6. The method according to claim 5, characterized in that, The step of marking the erroneous key signal as a misjudged sample and storing it in the correction sample library when a repeated key signal identical to the key identifier is received within a preset time window after the blocking time specifically includes: When a repeated key signal with the same key identifier is received within a preset time window after the blocking time, the touch data corresponding to the erroneous key signal and the touch data corresponding to the repeated key signal are respectively input into the currently effective recognition model to obtain a first score and a second score. Calculate the absolute value of the difference between the first score and the second score; When the absolute value of the difference is less than a preset fluctuation threshold, the touch data corresponding to the erroneous button signal is marked as a misjudged sample and stored in the correction sample library; When the absolute value of the difference is greater than or equal to the fluctuation threshold and the second score is higher than the preset effective threshold, the time point corresponding to the pressure peak in the erroneous key signal and the repeated key signal is extracted and the time difference is calculated. When the time difference is less than a preset duration threshold, the touch data corresponding to the erroneous key signal is marked as a misjudged sample and stored in the correction sample library.

7. A user terminal, characterized in that, The user terminal includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the user terminal to perform the method as described in any one of claims 1-6.

8. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on a user terminal, the user terminal performs the method as described in any one of claims 1-6.

9. A computer program product, characterized in that, When the computer program product is run on a user terminal, the user terminal performs the method as described in any one of claims 1-6.

Images