A screen intelligent key switch system

Abstract

The present invention discloses a screen intelligent key switch system, comprising: a touch sensing unit for collecting touch signals and environmental parameters; a communication interface module; a data processing unit for executing a phase space reconstruction touch intention recognition algorithm and an entropy driven adaptive energy allocation algorithm; a control execution unit; the phase space reconstruction touch intention recognition algorithm executed by the data processing unit is a time delay embedding method based on nonlinear dynamics theory; the entropy driven adaptive energy allocation algorithm executed by the data processing unit is a predictive energy management algorithm based on information entropy theory; the communication interface module adaptively switches the working states of the near field communication unit and the long-range communication unit according to the system state entropy and the communication urgency. The present invention has the following beneficial effects: providing an intelligent key switch system that can integrate multiple algorithms while taking into account touch accuracy, environmental adaptability, energy efficiency and response speed.

Description

Technical Field

[0001] The present invention relates to the field of human-computer interaction control technology, and in particular to a screen intelligent key switch system, which is suitable for smart home control, industrial control panels, medical equipment control interfaces, and various smart devices that need to maintain high-precision touch and low energy consumption in complex environments. Background Art

[0002] With the widespread adoption of the Internet of Things (IoT) and smart devices, touch-sensitive smart switches have become a crucial interface for human-computer interaction. These screen-based smart switches must not only provide intuitive visual feedback but also achieve high-precision touch recognition, robust operating stability with strong environmental adaptability, and efficient energy management to meet application requirements for intelligence, convenience, and low power consumption.

[0003] However, the smart push button switch technology currently on the market still has the following major problems:

[0004] First, there are issues with touch accuracy and environmental adaptability. Traditional touch technologies are mostly based on simple threshold judgment or basic pattern recognition, which can easily lead to false triggering or missed triggering in complex environments such as humidity, vibration, and temperature fluctuations. According to industry test data, the recognition accuracy of traditional touch systems drops to 65-75% under wet hand operation conditions; in strong vibration environments (>2g acceleration), the accuracy further drops to 40-50%. This poses a serious safety hazard in scenarios requiring high reliability, such as medical and industrial applications.

[0005] Second, energy management is inefficient. Existing technologies often rely on fixed wake-up cycles or simple trigger-based wake-up mechanisms, failing to dynamically adjust power consumption based on actual usage. This limits the battery life of battery-powered devices. Data shows that traditional touch systems typically consume 5-10mW of power in standby mode, while active power consumption can reach 50-100mW, making them inadequate for long-term battery-powered applications.

[0006] Third, there's a significant issue with response delay. Conventional touch systems typically experience a 25-50ms delay from touch signal acquisition to response completion. This delay can cause noticeable lag in applications requiring real-time response, impacting the user experience. This delay can be particularly detrimental in time-sensitive scenarios like industrial control and medical equipment, leading to operational errors.

[0007] Fourth, they lack adaptive learning capabilities. Existing touch systems are mostly fixed with preset parameters. They lack the ability to adapt to user habits and environmental changes, cannot provide a personalized touch experience, and are difficult to continuously optimize performance during use.

[0008] Research institutions and companies both domestically and internationally have conducted extensive research to address these challenges. Some advanced technologies employ machine learning to improve touch recognition accuracy, such as touch pattern recognition algorithms based on SVMs (support vector machines) or CNNs (convolutional neural networks). However, these algorithms are computationally complex and demand high processor and memory resources, making real-time processing difficult in resource-constrained embedded devices. Other technologies, such as dynamic power management (DPM) and dynamic voltage and frequency scaling (DVFS), are being used to optimize energy efficiency. However, these approaches primarily target computing units like the CPU and rarely consider the unique needs of touch systems.

[0009] There is an obvious technological gap in the current field of touch technology: there is a lack of systematic solutions that can simultaneously solve the problems of touch accuracy, environmental adaptability, energy efficiency and response delay.

[0010] In this context, developing an intelligent key switch system that can integrate multiple algorithms while taking into account touch accuracy, environmental adaptability, energy efficiency and response speed has important technical significance and market value. Summary of the Invention

[0011] The technical problem to be solved by the present invention is: how to provide a screen intelligent key switch system with high touch accuracy, strong environmental adaptability, high energy efficiency and low response delay, which can maintain a stable and reliable working state under various complex environmental conditions while achieving extremely low energy consumption and fast response speed.

[0012] To solve the above technical problems, the present invention provides a screen intelligent key switch system, comprising:

[0013] A touch sensing unit for collecting touch signals and environmental parameters; a communication interface module, including a near-field communication unit and a long-range communication unit; a data processing unit for executing a phase-space reconstruction touch intention recognition algorithm and an entropy-driven adaptive energy allocation algorithm; and a control execution unit for implementing touch response and energy management based on the algorithm results.

[0014] The phase space reconstruction touch intention recognition algorithm executed by the data processing unit is a time delay embedding method based on nonlinear dynamics theory; the entropy driven adaptive energy allocation algorithm executed by the data processing unit is a predictive energy management algorithm based on information entropy theory; the communication interface module adaptively switches the working status of the near-field communication unit and the long-range communication unit according to the system state entropy and communication urgency.

[0015] Preferably, the entropy driven adaptive energy allocation algorithm is: the energy demand prediction equation: E pred (t+Δt)=E base ×(1+α×H(S))×(1+β× )×(1+γ×H(U)) is piecewise linearized to: E component (i)=E min (i)+[E pred (t)-ΣE min (j)]×P i (t) / ΣP j (t)P i (t)=w1×P importance (i)+w2×P urgency (i)+w3×P efficiency (i);

[0016] The phase space reconstruction touch intention recognition algorithm is expressed as follows: X(t)=[x(t),x(t-τ),x(t-2τ),...,x(t-(m-1)τ)], where x(t) is the sampled touch signal sequence, τ is the time delay parameter, and m is the embedding dimension;

[0017] The calculation formula for each phase of touch intention is: I(t)=Classifier(PS Features ) Among them, PS Features =[LLE,CD,RR,DET,LAM,ENTR,...] is the phase space feature vector, Classifier is the classifier;

[0018] The phase space reconstruction touch intention recognition algorithm also includes an adaptive window mechanism, and the calculation formula of the window length L is: L=L base ×[1+β·f(S load )] where L base is the basic window length, β is the adjustment coefficient, f(S load ) is the load characteristic function;

[0019] The window function is a dynamically adjusted window function: w adaptive (n)=w base (n)·g(P change,rate ) where w base (n) is the basic window function, g(P change,rate ) is the window adjustment function, which is related to the power change rate.

[0020] Preferably, the phase space reconstruction touch intention recognition algorithm further includes an adaptive threshold decision mechanism: when the characteristic value exceeds the dynamic threshold, it is considered that a valid touch intention is detected; the threshold value T k The calculation formula is: k =T0(k)×[1+α·σ s (t)] where T0(k) is the basic threshold value, σ s(t) is the system state fluctuation index, α is the adjustment coefficient, and its value range is 0.5-2.0.

[0021] Preferably, in the entropy-driven adaptive energy allocation algorithm, the calculation formula of the system state entropy H(S) is: H(S)=-ΣP(s i )×log2(P(s i ))where P(s i ) indicates that the system is in state s i probability;

[0022] The optimization problem expression is:

[0023] minJ(x)=α·E total (x)+β·T response (x)+γ·(1-A accuracy (x))+δ·H system (x)stE total (x)≤E budget ,T response (x)≤T max ,A accuracy (x)≥A min ), where α, β, γ, and δ are dynamic weight coefficients, and E total (x) is the total energy consumption, T response (x) is the response time, A accuracy (x) is the touch accuracy, H system (x) is the system entropy, E budget For energy budget, T max is the maximum allowed response time, A min The minimum allowed accuracy.

[0024] Preferably, the system further includes a spectrum summary touch intention prediction module for realizing advance prediction of touch behavior; the multi-domain feature fusion model expression is: F multi =[W Features ,E Features ,C Features ,T Features ]W Features =[EntropyA j ,EnergyD j ,...,EnergyD1] is the wavelet feature, where the wavelet decomposition coefficient {A j ,D j ,D {j-1} ,...,D1} through multi-scale wavelet transform W ψ (a,b)=(1 / )∫x(t)·ψ*((tb) / a)dt is calculated;

[0025] E Features =[M1,M2,M3] is the spectrum feature, M1=∫f·E(f)df / ∫E(f)df is the spectrum center, M2=∫(f-M1)^2·E(f)df / ∫E(f)df is the spectrum width; M3=∫(f-M1)^3·E(f)df / (M2^(3 / 2)·∫E(f)df) is the spectrum skewness;

[0026] C Features =[Cp,qp] is the cepstrum feature, through cepstrum analysis C x (q)=F^(-1){log|F{x(t)}|} is calculated;

[0027] T Features It is the time domain feature.

[0028] Preferably, the data processing unit further executes a heterogeneous multimodal fusion autoencoding algorithm, whose feature extraction expression is: i =E i (x i ) where x i is the sensor data of the i-th modality, E i is the encoder of this mode, z i is the potential representation vector;

[0029] The attention fusion mechanism is expressed as: A ij =softmax(Q i ·K j ^T / √d)z fusion =∑ i ∑ j A ij ·V j where Q i =W Q ·z i is the query vector, K j =W k ·z j is the key vector, V j =W V ·z j is the value vector, d is the key vector dimension, A ij is the attention weight, z fusion is the fused feature vector.

[0030] Preferably, the communication resource allocation strategy of the communication interface module is: BW allocation =f(System complexity ,Message urgency ,Env Quality ) System complexityFor system complexity evaluation, Message urgency is the communication urgency, Env Quality For environmental quality assessment; the near-field communication unit is based on low-power Bluetooth technology and is used to transmit control instructions and basic status data; the long-range communication unit is based on Wi-Fi or 4G / 5G technology and is used to transmit system logs and complex interaction data.

[0031] Preferably, the system further includes a quantum excitation zero-delay touch prediction module, which constructs a touch state representation model based on quantum computing principles: in represents the system state at time t, α i (t) is the complex amplitude, reaching ∑ i |α i (t)|^2=1, is the base state, indicating different touch states;

[0032] The state evolution equation is:

[0033] Where U(t,Δt)=exp(-iH(t)Δt) is the evolution operator, H(t)=H user (t)+H env (t)+H int (t) is the system Hamiltonian, which enables advance prediction of touch behavior.

[0034] Preferably, the control execution unit executes a prediction-verification dual-path processing strategy, and the objective function is: D final (t)=w predict ·D predict (t-Δt)+w Verify ·D Verify (t)+w context ·D context (t) where D predict (t-Δt) is the predicted path result, D Verify (t) is the real-time verification path result, D context (t) is the environmental context information, w predict 、w Verify 、w context is the dynamic weight coefficient, through the function w predict =f p (SNR,H(S),Stability predict ), w Verify =f V (SNR,H(S),Confidence Verify ), w context =f c (Reliabilitycontext ,Δt) calculation; the priority and intensity of control execution are dynamically adjusted according to the fusion decision results.

[0035] Preferably, the system adopts a layered heterogeneous computing architecture: a low-power resident core layer, responsible for basic signal monitoring and wake-up judgment, with a power consumption of <1mW; a main computing core layer, processing the core functions of the phase space reconstruction touch intention recognition algorithm and the entropy-driven adaptive energy allocation algorithm, with a power consumption of 5-30mW; an on-demand activated accelerator layer, processing the complex calculations of spectral summary touch intention prediction and heterogeneous multimodal fusion, with a power consumption of 30-100mW; the layers are coordinated through shared memory and interrupt mechanisms to achieve a dynamic balance between performance and energy consumption.

[0036] The beneficial effects of the present invention are:

[0037] Significantly Improved Touch Accuracy and Environmental Adaptability: A phase-space reconstruction touch intent recognition algorithm captures the dynamic characteristics of touch signals in a high-dimensional space. Combined with spectral summary touch intent prediction and a heterogeneous multimodal fusion autoencoder algorithm, the system is able to maintain high touch recognition accuracy in a variety of extreme environments. Test data shows that the system achieves a recognition accuracy rate of 99.95% under extreme operating conditions. In particular, the system's overall touch accuracy remains above 99.2% even in extreme situations.

[0038] Optimizing energy management efficiency: The entropy-driven adaptive energy allocation algorithm, based on information entropy theory, accurately predicts the system's energy needs and enables refined component-level energy allocation. Furthermore, combined with a layered heterogeneous computing architecture, the system dynamically adjusts power consumption based on actual usage. Measured data shows that the system's standby power consumption is reduced to 0.28mW, and active power consumption is only 28mW. This extends the life of battery-powered devices, allowing the system to operate continuously on a single coin cell battery.

[0039] Significantly reduce response delay: Through spectral summary touch intention prediction and quantum excitation zero-delay touch prediction algorithm, the system can predict the user's intention before the actual touch and prepare corresponding resources in advance, reducing the perceived response delay and significantly improving the smoothness of operation.

[0040] Adaptive Learning: Utilizing a heterogeneous multimodal fusion autoencoding algorithm and a quantum-stimulated zero-latency touch prediction module, the system rapidly learns user habits, building an initial user model with only a few interactions. This model continuously optimizes during use, providing a personalized touch experience for each user. With continued use, the system's touch recognition accuracy and energy management efficiency can be further improved.

[0041] Achieve multi-algorithm collaborative gain: The present invention overcomes the shortcomings of using each algorithm alone by deeply fusing multiple algorithms. For example, the phase space reconstruction touch intention recognition algorithm has high computational complexity and slow response speed when used alone, while the spectral summary touch intention prediction algorithm has limited accuracy when used alone. By combining the two, a complementary collaboration of fast prediction + precise verification is achieved, achieving fast response while maintaining high accuracy. Similarly, the combination of the entropy-driven adaptive energy allocation algorithm and the heterogeneous multimodal fusion autoencoding algorithm makes energy allocation decisions more accurate, reduces energy consumption and improves system stability. The gain effect generated by the collaborative fusion of the above-mentioned multiple algorithms far exceeds the simple superposition of each algorithm, achieving unexpected technical effects.

[0042] In summary, the screen intelligent button switch system provided by the present invention significantly improves touch accuracy, environmental adaptability, energy efficiency and response speed by deeply integrating multiple advanced algorithms, while providing personalized user experience and strong fault tolerance capabilities. It has important technical value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a block diagram of the overall structure of the screen intelligent key switch system of the present invention; Figure 2 It is the working flow chart of the system of the present invention; Figure 3 This is a flowchart of the implementation of the phase space reconstruction touch intention recognition algorithm in the present invention; Figure 4 This is a flow chart for implementing the entropy-driven adaptive energy allocation algorithm of the present invention; Figure 5 This is a flowchart of the implementation of the spectrum summary touch intention prediction algorithm in the present invention; Figure 6 This is a flowchart for implementing the heterogeneous multimodal fusion autoencoding algorithm of the present invention; Figure 7 A diagram showing the working principle of the layered heterogeneous computing architecture in the present invention; Figure 8 Flowchart for implementing the prediction-verification dual-path processing strategy of the present invention; Figure 9 This is a diagram showing the working principle of the quantum excitation zero-delay touch prediction module in the present invention; Figure 10 This is a product schematic diagram of the screen smart key switch of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, the screen intelligent key switch system of the present invention mainly includes a touch sensing unit 10, a communication interface module 20, a data processing unit 30, a control execution unit 40, a display unit 50 and a power management unit 60.

[0046] The touch sensing unit 10 is used to collect touch signals and environmental parameters and includes a capacitive touch sensor 11, a proximity sensor 12, an accelerometer 13, a temperature sensor 14, and an ambient light sensor 15. The capacitive touch sensor 11 is used to collect user touch signals, with a sampling frequency of 120-240Hz and a sampling accuracy of 16 bits. The proximity sensor 12 is used to detect the movement of a user's finger approaching the screen, with a sampling frequency of 60Hz. The accelerometer 13 is used to detect the vibration and movement of the device, with a sampling frequency of 100Hz. The temperature sensor 14 is used to monitor the ambient temperature and device temperature, with a sampling frequency of 1Hz. The ambient light sensor 15 is used to detect the ambient light intensity, with a sampling frequency of 1Hz.

[0047] The communication interface module 20 includes a near-field communication unit 21 and a long-range communication unit 22. The near-field communication unit 21 is based on Bluetooth low energy technology (BLE5.2) and is used to transmit control instructions and basic status data. The maximum transmission distance is 10 meters and the operating power consumption is 1-5mW. The long-range communication unit 22 is based on Wi-Fi or 4G / 5G technology and is used to transmit system logs and complex interaction data. The maximum transmission distance can reach 100 meters (Wi-Fi) or further (4G / 5G), and the operating power consumption is 50-200mW. The communication interface module 20 adaptively switches the working status of the near-field communication unit 21 and the long-range communication unit 22 according to the system state entropy and communication urgency.

[0048] The data processing unit 30 utilizes a layered heterogeneous computing architecture, comprising a low-power resident core 31, a main computing core 32, and an accelerator 33. The low-power resident core 31 is based on an ARM Cortex-M0+ processor with a 16MHz clock frequency, responsible for basic signal monitoring and wake-up determination, consuming less than 1mW. The main computing core 32 is based on an ARM Cortex-M4F processor with a clock frequency of 80-120MHz, responsible for executing the core functions of the phase-space reconstruction touch intent recognition algorithm and the entropy-driven adaptive energy allocation algorithm, consuming 5-30mW. The accelerator 33 utilizes a dedicated neural network processing unit supporting 8-bit integer quantization operations, handling the complex computations of spectral summary touch intent prediction and heterogeneous multimodal fusion. It is activated on demand and consumes 30-100mW. The computing units coordinate through shared memory and interrupt mechanisms to achieve a dynamic balance between performance and energy consumption.

[0049] The control execution unit 40 is used to implement touch response and energy management based on the algorithm results. It includes a touch response module 41 and an energy management module 42. The touch response module 41 receives the decision results of the data processing unit 30 and generates corresponding UI responses and control instructions. The energy management module 42 controls the operating mode and energy distribution of each system component based on the results of the entropy-driven adaptive energy allocation algorithm.

[0050] The display unit 50 is an IPS or OLED display screen with a resolution of 240×320 pixels, a refresh rate of 60 Hz, an adaptive brightness adjustment function, and a power consumption range of 10-100 mW.

[0051] The power management unit 60 includes a battery management system 61 and a power conversion module 62, supports 3.7V lithium battery or 2-3V button battery power supply, provides multiple power outputs (1.2V, 1.8V, 3.3V, etc.), and has overvoltage, overcurrent and overtemperature protection functions. The self-consumption current in standby mode is less than 5μA.

[0052] like Figure 2 As shown, the workflow of the system of the present invention mainly includes the following stages:

[0053] Initialization phase: After the system is powered on, it first performs hardware initialization and self-test, loads the pre-trained algorithm model, and configures sensor parameters and communication parameters.

[0054] Low power monitoring stage: The low power resident core 31 enters the low power monitoring mode, periodically samples the environment and proximity sensor data, calculates the real-time estimate of the system state entropy, and controls the power consumption below 1mW.

[0055] Pre-touch prediction stage: When the proximity sensor 12 detects that the user's finger is approaching the touch panel, the system wakes up the main computing core 32, activates the spectrum summary touch intention prediction module, predicts the user's touch position and intention, generates a touch probability flow field, and pre-allocates computing resources accordingly, preheats the triggered UI components, and adjusts the touch sensitivity distribution.

[0056] Real-time touch processing stage: When an actual touch occurs, the system collects multimodal sensor data, accurately identifies the user's touch intention through phase space reconstruction touch intention recognition algorithm, compares it with the predicted results, and updates the prediction model parameters.

[0057] Response execution stage: Based on the touch intention recognition result and the prediction-verification dual-path processing strategy, a final decision is generated, and the control execution unit 40 performs the corresponding UI response and functional operation.

[0058] Energy optimization stage: Based on the results of the entropy-driven adaptive energy allocation algorithm, the system dynamically adjusts the working mode and energy allocation of each component to optimize overall energy consumption.

[0059] Continuous Learning: The system continuously collects user touch behavior data and environmental data in the background, updates various algorithm model parameters, and optimizes system performance and user experience. The above stages form a closed-loop feedback mechanism.

[0060] like Figure 3As shown in Figure 1, the Phase Space Reconstruction Touch Intent Identification (PS-TII) algorithm is one of the core algorithms of the present invention. It maps one-dimensional touch signals into a high-dimensional phase space and extracts their dynamic characteristics to achieve high-precision touch intent identification. The specific implementation steps of the algorithm are as follows:

[0061] Step 1: Touch signal acquisition and preprocessing. The system acquires the capacitive touch sensor signal data x(t) with a sampling frequency of 120Hz and a sampling accuracy of 16 bits. A 4th-order Butterworth low-pass filter is used to remove high-frequency noise with a cutoff frequency of 30Hz. Signal normalization is then performed:

[0062] x norm (t)=(x(t)-x min ) / (x max -x min )

[0063] where x min and x max are the minimum and maximum values ​​in the observation window respectively, and the normalized signal range is [0,1].

[0064] Step 2: Phase space reconstruction. Use the time delay embedding method to reconstruct the one-dimensional signal into a high-dimensional phase space trajectory:

[0065] X(t)=[x(t),x(t-τ),x(t-2τ),...,x(t-(m-1)τ)]

[0066] Where τ is the time delay parameter, determined by the first local minimum of the mutual information function, and typically ranges from 3 to 5 sampling points. m is the embedding dimension, determined by the false nearest neighbor method, and typically ranges from 3 to 5. In practice, the system uses a sliding window of length 32 to store historical data, and a ring buffer to optimize memory usage and ensure real-time data processing.

[0067] Step 3: Dynamic feature extraction. The following features are extracted from the reconstructed phase space:

[0068] (1) Maximum Lyapunov exponent (LLE): It characterizes the degree of chaos of the system. The calculation formula is:

[0069] LLE=lim(t→∞)(1 / t)×ln(||δX(t)|| / ||δX(0)||);

[0070] Where δX(0) is the distance between initially close trajectory points in phase space, and δX(t) is the degree of separation of the trajectories after time t. In practice, the Wolf algorithm is used to calculate LLE, with a calculation window of 16 sampling points and an initial distance threshold set to 1% of the phase space diameter. Larger LLE values ​​indicate greater uncertainty in touch behavior, typically ranging from 0.1 to 2.0.

[0071] (2) Correlation dimension (CD): reflects the complexity and dimension of the phase space trajectory. The calculation formula is:

[0072] CD=lim(r→0)lim(N→∞)[ln(C(r)) / ln(r)]

[0073] Where C(r)=(1 / N 2 )×Σ i,j (r-||X i -X j ||) is the relevant integral, is the Heaviside step function, and r is the set distance threshold. In practice, the Grassberger-Procaccia algorithm is used to calculate CD. A linear fit is performed using 5-8 different r values, and the slope is taken as the CD value. CD typically ranges from 1.0 to 3.0, with higher values ​​indicating greater signal complexity.

[0074] (3) Circular Quantitative Analysis (RQA): The periodicity and determinism of the phase space trajectory are analyzed by constructing a circulant matrix, which is defined as:

[0075] R(i,j)= (ε-||X i -X j ||)

[0076] Where ε is a threshold parameter, set to 10-15% of the phase space diameter. The following features are extracted from the circulant matrix:

[0077] Recycling rate (RR) = (1 / N 2 )×Σ i ,jR(i,j): indicates the ratio of loop points to the total number of points, ranging from 0 to 1

[0078] Determinism (DET) = Σ l =l min ^Nl·P(l) / Σ i ,jR(i,j): represents the ratio of the diagonal structure, reflecting the certainty of the system, ranging from 0 to 1;

[0079] Laminar flow (LAM) = Σ V =v min^Nv·P(v) / Σ i ,jR(i,j): represents the ratio of vertical line structure, reflecting the laminarity of the system, ranging from 0 to 1;

[0080] Entropy (ENTR) = -Σ l =l min ^NP(l)·ln(P(l)): represents the entropy of the diagonal length distribution, reflecting the complexity of the system.

[0081] Where P(l) and P(v) are the probability distributions of the lengths of diagonal and vertical lines, respectively. min and v min are the minimum diagonal and vertical line lengths, respectively, usually set to 2.

[0082] Step 4: Construct feature vector. Combine the above extracted dynamic features into a feature vector:

[0083] PS Features =[LLE,CD,RR,DET,LAM,ENTR,...];

[0084] This feature vector usually contains 10-15 dimensions and comprehensively characterizes the dynamic characteristics of the touch signal.

[0085] Step 5: Touch intent classification. Use the pre-trained classifier to classify the feature vector and identify the touch intent:

[0086] I(t)=Classifier(PS Features );

[0087] The classifier can use support vector machines (SVM) or random forests, and pre-train parameters. Touch intent categories include: intentional touch, unintentional touch, interference touch, etc. At the same time, the classification confidence is calculated:

[0088] Confidence=1-H(P(I|X));

[0089] where H(P(I|X))=-Σ i P(I i |X)·log2(P(I i |X)) is the entropy of the classification probability distribution, P(I i |X) is the feature vector X corresponding to category I i The confidence level ranges from 0 to 1, with higher values ​​indicating more reliable classification results.

[0090] Step 6: Trajectory stability evaluation. Calculate trajectory stability index:

[0091] Trajectory stability =1-(LLE / LLEmax );

[0092] Among them LLE max This is the preset maximum Lyapunov exponent threshold, usually set to 2.0. Track stability ranges from 0 to 1, with higher values ​​indicating more stable touch tracks.

[0093] Step 7: Adaptive threshold decision. Dynamically adjust the decision threshold based on trajectory stability and environmental parameters:

[0094] T k =T0(k)×[1+α·σ s (t)];

[0095] Where T0(k) is the basic threshold value, and different values ​​are set for different feature dimensions k; σ s (t) is the system state fluctuation indicator, calculated from environmental sensor data; α is the adjustment coefficient, ranging from 0.5 to 2.0, and adaptively adjusted based on environmental complexity. When the characteristic value exceeds the dynamic threshold, the system considers that a valid touch intention has been detected.

[0096] Step 8: Adaptive window mechanism. Dynamically adjust the processing window length based on the changing characteristics of the touch signal and system load:

[0097] L=L base ×[1+β·f(S load )];

[0098] Among them L base is the basic window length, usually set to 32 sampling points; β is the adjustment coefficient, ranging from 0.1 to 0.5; f(S load ) is the load characteristic function, with a value range of -0.5 to 0.5. When the system load is high, it takes a negative value to shorten the window, and vice versa. At the same time, the window function is also dynamically adjusted:

[0099] w adaptive (n)=w base (n)·g(P change,rate );

[0100] where w base (n) is the basic window function, usually using Hanning window; g(P change,rate ) is the window adjustment function, which is related to the signal power change rate. When the signal changes dramatically, the window function becomes steeper, improving the time resolution.

[0101] The computational complexity of the phase space reconstruction touch intent recognition algorithm is mainly concentrated in the phase space reconstruction and feature extraction stages, especially the calculation of LLE and CD. To improve the algorithm execution efficiency, the system adopts the following optimization strategies:

[0102] (1) Fixed-point calculations are used instead of floating-point operations to package 16-bit data for processing, thereby improving computational efficiency; (2) Lookup tables are used to pre-calculate some complex functions, such as logarithmic and exponential functions; (3) For feature extraction algorithms, incremental calculations are used to reduce redundant calculations; (4) Piecewise linear approximation is used for the calculation of related dimensions to reduce computational complexity.

[0103] Through the above optimization, the algorithm can be executed in real time on the main computing core (80MHz Cortex-M4F), and each processing takes no more than 5ms, meeting the real-time requirements.

[0104] Actual tests show that the phase-space reconstruction touch intent recognition algorithm achieves 99.2% accuracy in standard environments, 95.1% accuracy with wet hands, 96% accuracy in high vibration environments, and 92% accuracy in extremely low signal-to-noise ratio (-20dB) environments, significantly outperforming traditional touch algorithms. The algorithm maintains high recognition accuracy and environmental adaptability, especially in complex environments with multiple interferences.

[0105] like Figure 4 As shown in Figure 1, the Entropy-Driven Adaptive Energy Allocation Algorithm (EAEA) is another core algorithm of the present invention. It quantifies the uncertainty of system state and user behavior through information entropy theory to achieve accurate energy demand prediction and intelligent energy allocation. The specific implementation steps of the algorithm are as follows:

[0106] Step 1: System status monitoring. The system collects the following data: (1) System status data: CPU usage, memory usage, touch activity, etc., with a sampling frequency of 1 Hz; (2) Environmental parameter data: temperature, humidity, light, etc., with a sampling frequency of 0.1 Hz; (3) User operation data: touch mode, touch intensity distribution, usage time distribution, etc., recorded in real time.

[0107] Step 2: Calculate the system state entropy. Based on the system state data, calculate the system state entropy:

[0108] H(S)=-ΣP(s i )×log2(P(s i ));

[0109] Where P(s i ) indicates that the system is in state s i The probability of a system being in a state of sleep is estimated by statistically analyzing historical data. The system state space includes the following common states: deep sleep, light sleep, standby, light activity, moderate activity, high activity, full load, and abnormal state. The value of H(S) ranges from 0 to log2(N), where N is the number of states, typically 8, resulting in a maximum entropy of 3.

[0110] Step 3: Calculate usage pattern entropy. Calculate usage pattern entropy based on user operation data:

[0111] H(U)=-ΣP(u j )×log2(P(u j ));

[0112] Where P(u j ) indicates that the user performs operation u j The probability of a touch is estimated by statistically analyzing historical data. The operation space includes short press, long press, fast swipe, slow swipe, multi-touch, and no operation. The value of H(U) ranges from 0 to log2(M), where M is the number of operation types, typically 6, resulting in a maximum entropy of 2.58.

[0113] Step 4: Entropy gradient calculation. Calculate the rate of change of system state entropy and usage pattern entropy:

[0114] =H(S t )-H(S {t-Δt} );

[0115] =H(U t )-H(U {t-Δt} );

[0116] Where Δt is the time interval for entropy calculation, which is usually set to 10 seconds. The entropy gradient reflects the changing trend of system status and user behavior and is crucial for predicting future energy demand.

[0117] Step 5: Energy demand prediction. Based on the system state entropy, usage pattern entropy and their gradient, predict the energy demand in the future period:

[0118] E pred (t+Δt)=E base ×(1+α×H(S))×(1+β× )×(1+γ×H(U))×(1+δ×Confidence intent );

[0119] in:

[0120] E baseis the basic energy demand, determined by hardware parameters, usually 5-20mW; α is the state entropy weight coefficient, ranging from [0.2, 0.5], adjusted according to the complexity of the system state; β is the entropy change rate weight coefficient, ranging from [0.3, 0.7], reflecting the system's sensitivity to state changes; γ is the usage mode entropy weight coefficient, ranging from [0.1, 0.4], reflecting the system's adaptability to user behavior patterns; δ is the intent confidence weight coefficient, ranging from [0.1, 0.3], controlling the impact of touch intent reliability on energy allocation; intent The confidence level of touch intention is provided by the phase space reconstruction touch intention recognition algorithm and ranges from 0 to 1.

[0121] Step 6: Energy budget calculation. Based on the available energy and the maximum allowed power consumption, calculate the energy budget:

[0122] E budget =min(E avlable ,E max,allowed );

[0123] Among them E avlable is the currently available energy, provided by the battery management system; E max,allowed The maximum allowable power consumption is set by the system and is usually 100-200mW.

[0124] Step 7: Component-level energy optimization. Calculate the priority weight of each component:

[0125] P i (t)=w1×P importance (i)+w2×P urgency (i)+w3×P efficiency (i);

[0126] Where: P importance (i) is the importance weight of component i, ranging from 0 to 1, determined according to the component function; P urgency (i) is the urgency weight of component i, ranging from 0 to 1, determined according to the current task requirements; P efficiency (i) is the energy efficiency weight of component i, ranging from 0 to 1, reflecting the functional benefits generated per unit energy; w1, w2, and w3 are weight coefficients, reaching w1+w2+w3=1, which are dynamically adjusted according to the system status.

[0127] Step 8: Component energy allocation. Based on component priority and predicted energy demand, calculate the energy allocation for each component:

[0128] E component (i)=E min (i)+[E pred (t)-ΣEmin (j)]×P i (t) / ΣP j (t); where E min (i) is the minimum operating energy of component i, which is usually 10-30% of the normal operating energy. The above allocation strategy ensures that each component receives at least the minimum operating energy, and the remaining energy is distributed according to the priority ratio.

[0129] Step 9: Entropy response function design. Design a nonlinear entropy response function to dynamically adjust the sensitivity of energy distribution:

[0130] f entropy (H,H critical )=1 / (1+exp(-k×(HH critical ))); where H critical is the critical entropy value, representing the threshold at which the system state transitions from determinism to uncertainty, and its value range is [0.3, 0.7]. k is the sensitivity coefficient, which controls the steepness of the response curve and its value range is [5, 15]. When the system entropy approaches or exceeds the critical value, the energy allocation strategy is rapidly adjusted to cope with the increased uncertainty of the system state.

[0131] Step 10: Predictive sleep-wake management. Based on entropy analysis, predict the probability of future wake-up:

[0132] P wakeup (t+Δt)=f(H(S),H(U), , ,E avlable ); set the sleep depth based on the predicted probability, including deep sleep (power consumption <0.1mW), light sleep (power consumption 0.1-1mW), shallow sleep (power consumption 1-5mW) and other levels. At the same time, determine the pre-wake-up time:

[0133] T prewake =T current +max(0,T predicted -T wakeup -T margin ); where T predicted is the predicted next usage time, T wakeup is the time required to wake up, T margin This is the safety margin time. Through the pre-wake-up mechanism, the system can complete preparations before the user uses it, achieving "zero-perception" delay.

[0134] Step 11: Global optimization solution. The energy allocation problem is formalized as a multi-objective optimization problem:

[0135] minJ(x)=α·E total (x)+β·T response(x)+γ·(1-A accuracy (x))+δ·H system (x); stE total (x)≤E budget ,T response (x)≤T max ,A accuracy (x)≥A min . Among them: E total (x) is the total energy consumption in mW; T response (x) is the response time, in ms; A accuracy (x) is the touch accuracy, ranging from 0 to 1; H system (x) is the system entropy, reflecting the uncertainty of the system; α, β, γ, δ are dynamic weight coefficients, which are automatically adjusted according to the system status and user priority; E budget For energy budget, T max is the maximum allowed response time, A min is the minimum allowed accuracy. The gradient descent method is used to solve the optimization problem and obtain the global optimal energy allocation solution.

[0136] Step 12: Feedback learning and parameter optimization. Collect actual energy consumption and performance data and calculate the prediction error:

[0137] Error energy =|Eactual-E pred | / E pred ; Use gradient descent method to optimize algorithm parameters: θ new =θ old -η× Where θ represents the algorithm parameters (α, β, γ, δ, k, etc.), η is the learning rate, usually set to 0.01-0.1. Update the entropy threshold regularly:

[0138] H threshold (t+1)=δ×H threshold (t)+(1-δ)×H optimal ; where δ is the smoothing coefficient, usually set to 0.9; H optimal is the current optimal entropy threshold, determined through performance evaluation.

[0139] The implementation of the entropy-driven adaptive energy allocation algorithm is mainly based on the main computing core (Cortex-M4F). The computational complexity of the algorithm itself is moderate, mainly in the state entropy calculation and optimization solution stages. To improve computational efficiency, the system adopts the following optimization strategies:

[0140] (1) A lookup table method is used to calculate the entropy value to control complex logarithmic operations; (2) A fast gradient descent method with a limited number of iterations is used to solve the optimization problem; (3) A periodic strategy is used to update some parameters instead of updating them every time; (4) A piecewise linear approximation is used for the entropy response function to improve the calculation speed.

[0141] Through the above optimization, the execution cycle of the algorithm can be controlled within 100ms, the CPU time occupied does not exceed 5%, and the energy consumption is controlled below 0.5mW, meeting the real-time and low power consumption requirements.

[0142] Actual tests have shown that the entropy-driven adaptive energy allocation algorithm can reduce system standby power consumption to 0.5mW, a reduction of approximately 90% compared to traditional energy management technologies. Active power consumption is also reduced to just 25mW, a reduction of approximately 50%. Energy efficiency can be further improved by 15-20% over long-term use, as it learns user behavior patterns. Tests have shown that a system using this algorithm can operate sustainably for 3-5 years (using 100 times daily) using a single CR2032 coin cell battery (approximately 225mAh), a 5-10 times increase compared to traditional systems.

[0143] like Figure 5 As shown in Figure 1, the Wave Spectral Summary Touch Intention Prediction Algorithm (WSC-TIP) achieves advanced prediction of touch behavior by integrating wavelet transform, spectrum analysis, and cepstrum analysis techniques. The specific implementation steps of the algorithm are as follows:

[0144] Step 1: Signal acquisition and preprocessing. The system acquires the touch signal x(t) with a sampling frequency of 240 Hz and a sampling accuracy of 16 bits. A 3-point median filter is applied to remove impulse noise, and then a 6th-order Chebyshev bandpass filter is used to retain the 1-40 Hz signal component. The filtered signal is represented as x filtered (t).

[0145] Step 2: Multi-scale wavelet decomposition. Perform multi-scale wavelet decomposition on the pre-processed signal:

[0146] W ψ (a,b)=(1 / )∫x filtered (t)·ψ*((tb) / a)dt; where ψ is the wavelet mother function, Daubechies wavelet (DB4) is selected; a is the scale parameter, which controls the expansion and contraction of the wavelet; b is the translation parameter, which controls the translation of the wavelet along the time axis. For the touch signal, J=4 layer decomposition is selected to obtain the multi-scale coefficients: {A j ,D j ,D {j-1} ,...,D1}; where A j is the J-layer approximation coefficient, D jis the detail coefficient of layer j. In practice, the fast wavelet transform algorithm is used, and the computational complexity is O(N), where N is the signal length.

[0147] Step 3: Wavelet coefficient feature extraction. Extract the following features from the wavelet decomposition coefficients: (1) Entropy feature: Calculate the entropy value EntropyA of the approximate coefficients j =-ΣP(A j )·log2(P(A j )), reflecting the uncertainty of the signal; (2) Energy characteristics: Calculate the energy of the detail coefficients of each layer EnergyD j =Σ|D j | 2 , reflecting the energy distribution of the signal in different frequency bands; (3) Statistical characteristics: calculate the mean, variance, skewness, kurtosis and other statistical quantities of each coefficient.

[0148] Step 4: Spectral analysis. Perform short-time Fourier transform on the wavelet decomposition coefficients ( STFT ):

[0149] STFT x (t,f)=∫x filtered (τ)·w(τ-t)·e^(-j2πfτ)dτWhere w(τ-t) is the window function, using a Hanning window with a window length of 64 sampling points and an overlap rate of 50%. Calculate the spectrum energy distribution: E(f)=| STFTx (t,f)|^2; Moment features are extracted from the spectrum: (1) Spectral center: M1=∫f·E(f)df / ∫E(f)df, which represents the energy-weighted average frequency; (2) Spectral width: M2=∫(f-M1)^2·E(f)df / ∫E(f)df, which represents the degree of dispersion of the spectrum; (3) Spectral skewness: M3=∫(f-M1)^3·E(f)df / (M2^(3 / 2)·∫E(f)df), which represents the degree of asymmetry of the spectrum.

[0150] Step 5: Cepstrum analysis. Perform cepstrum analysis on the signal to extract the periodic characteristics of the touch action:

[0151] C x (q)=F^(-1){log|F{x filtered (t)}|}; where F represents Fourier transform, F^(-1) represents inverse Fourier transform, and q is the independent variable in the cepstrum, called quefrency. The following features are extracted from the cepstrum: (1) Cepstrum peak: Cp=max(C x (q)), represents the intensity of the periodic signal; (2) Cepstrum peak position: qp=argmax(C x (q)), which represents the fundamental period of the periodic signal.

[0152] Step 6: Multi-domain feature fusion. Construct multi-domain feature vector:

[0153] F multi =[W Features ,E Features ,C Features ,T Features ]; where: W Features =[EntropyA j ,EnergyD j ,...,EnergyD1] is the wavelet feature with dimension J+1; E Features =[M1,M2,M3] is the spectrum feature with a dimension of 3; C Features =[Cp,qp] is the cepstrum feature with a dimension of 2; T Features These are time-domain features (mean, variance, etc.) with dimensions of 4 to 6. A complete feature vector typically contains 15 to 20 dimensions, providing a comprehensive representation of the touch signal.

[0154] Step 7: Conditional Random Field Model Prediction. Use the Conditional Random Field (CRF) model to perform sequence prediction on multi-domain features:

[0155] P(I|F multi )=(1 / Z(F multi ))·exp(∑ k λ k ·f k (I,F multi )); where: I is the touch intent category, including click, long press, slide, multi-touch, etc.; Z(F multi ) is a normalization factor to ensure that the probability sum is 1; λ k is the weight parameter, determined by offline training; f k For the feature function, establish the mapping relationship between features and categories. Step 8: Construct the prior probability map. Construct the touch prior probability map based on historical touch data:

[0156] P prior (x,y,t)=∑ i w i ·G((x,y,t),μ i ,Σ i ); where: (x, y) is the touch screen coordinate, t is the time variable; G is a multivariate Gaussian distribution, representing the space-time distribution of touch probability; μ i ,Σ i is the mean and covariance of the i-th Gaussian component, indicating the center and dispersion of the touch probability; w i is the weight coefficient, ∑w i=1, indicating the importance of each Gaussian component. The prior probability map reflects the user's touch preferences in different areas and at different times, providing a basis for prediction.

[0157] Step 9: Fusion prediction model. Combining the conditional random field model and the prior probability map, we get the final prediction result:

[0158] P final (I,x,y,t)=α·P(I|F multi )+(1-α)·P prior (x, y, t); where α is the dynamic fusion coefficient, ranging from [0, 1], and is dynamically adjusted according to the environmental noise and user behavior stability: α = 1 / (1 + exp(β·(SNR threshold -SNR current ))); where β is the adjustment coefficient, usually set to 0.5; SNR threshold is the signal-to-noise ratio threshold, usually set to 10dB; SNR current is the signal-to-noise ratio estimate of the current signal. When the ambient noise is large, the value of α decreases, and the system relies more on prior knowledge; when the signal quality is good, the value of α increases, and the system relies more on current observations.

[0159] Step 10: Pre-touch detection and prediction. Based on proximity sensor data, the system detects the user's finger approaching the screen. When a proximity event is detected, the prediction module is activated. Based on multi-domain features and a priori probability maps, it predicts the touch location and intention, generates a touch probability flow field, and pre-allocates corresponding system resources.

[0160] Step 11: Verify and update the predicted result. When the actual touch occurs, the system compares the predicted result with the actual result: ΔP=||P predict -I real ||; where P predict is the predicted touch probability distribution, I real is the actual observed touch behavior. Then update the prediction model parameters: θ new =θ old -η· ; where θ represents the model parameters and η is the learning rate, which is usually set to 0.01-0.05. At the same time, the prediction time window is adjusted according to the prediction error: Δt new =f(ΔP,Conf real );Conf real The confidence level of real-time touch recognition.

[0161] Step 12: Collaborate with the PS-TII algorithm. The WSC-TIP algorithm provides predicted probability distribution, and the PS-TII algorithm provides real-time intent recognition. The two are fused through time series:

[0162] I final (t)=γ(t)·I WSC -TIP(t-Δt)+(1-γ(t))·I pS-TII (t); where γ(t) is the time-dependent weight function: γ(t)=γ max ·exp(-|tt predict | / τ); where γ max The above fusion method utilizes the prediction capability of WSC-TIP and the precise verification capability of PS-TII to form a complementary effect.

[0163] The computational complexity of the touch intention prediction algorithm is mainly concentrated in the wavelet decomposition, spectrum analysis and conditional random field prediction stages. In order to reduce the computational burden, the system adopts the following optimization strategies: (1) Use the fast wavelet transform algorithm to reduce the amount of calculation; (2) STFT The calculation is implemented using FFT, and partial frequency band analysis is used instead of full spectrum analysis; (3) Cepstrum analysis is performed only on key frequency bands to reduce the amount of calculation; (4) The characteristic function of the conditional random field model is sparsely represented to reduce storage and computing requirements; (5) The prior probability map is represented using a finite number of Gaussian components, usually no more than 10. After the above optimization, the algorithm needs to be executed on an accelerator, and the calculation time for a single prediction is approximately 10-15ms, making it suitable for real-time prediction scenarios.

[0164] Actual tests have shown that the Spectrum Summary touch intention prediction algorithm can predict user intention 25-100ms before the actual touch, with a prediction accuracy of 98.7% in standard environments and over 95% in complex ones. Especially for repetitive user actions, the prediction accuracy can be improved to over 99%. This advanced prediction capability significantly reduces system response latency, reducing the perceived response time to under 5ms, allowing users to experience near-zero-latency touch feedback.

[0165] like Figure 6 As shown in Figure 1, the Heterogeneous Multimodal Fusion Autoencoder (HMFAE) algorithm fuses data from different sensors to build a unified representation space, improving the system's environmental adaptability and fault tolerance. The specific implementation steps of the algorithm are as follows:

[0166] Step 1: Multimodal data acquisition and preprocessing. The system acquires the following sensor data: (1) capacitive touch data: sampling rate 240Hz, resolution 12×8 dots; (2) acceleration data: sampling rate 100Hz, 3-axis (x, y, z); (3) environmental data: temperature, humidity, light, etc., sampling rate 1Hz.

[0167] Standardize all types of data and unify the numerical range to the [-1,1] interval: x norm =2×(xx min ) / (x max -x min )-1; where x min and x max The timestamps are then used to align data at different sampling rates and construct a synchronized multimodal data stream.

[0168] Step 2: Modality-specific feature extraction. Design a dedicated feature extraction network for data of different modalities:

[0169] (1) Capacitive data encoder: A three-layer CNN structure is used, and the parameters of each layer are as follows: the first layer: 8 3×3 convolution kernels, a step size of 1, ReLU activation function, and an output feature map size of 10×6×8; the second layer: 16 3×3 convolution kernels, a step size of 1, ReLU activation function, and an output feature map size of 8×4×16; the third layer: 32 2×2 convolution kernels, a step size of 1, ReLU activation function, and an output feature map size of 7×3×32; the fully connected layer: the feature map is flattened and mapped to a 64-dimensional feature vector.

[0170] (2) Acceleration data encoder: A bidirectional LSTM network is used with the following structure: input layer: receives 3-axis acceleration data of 20 time steps, with an input dimension of 20×3; bidirectional LSTM layer: the hidden layer size is 64, and the output dimension is 20×128; pooling layer: performs maximum pooling on the time dimension, with an output dimension of 128; fully connected layer: maps features to a 64-dimensional feature vector.

[0171] (3) Environmental data encoder: It adopts a two-layer MLP structure, and the parameters of each layer are as follows: First layer: the input dimension is the number of environmental sensors (usually 3-5), the hidden layer size is 32, and the ReLU activation function; Second layer: the 32-dimensional hidden features are mapped to a 64-dimensional feature vector, and the ReLU activation function is used.

[0172] Through the above-mentioned specific encoder, the data of different modalities are mapped to a unified 64-dimensional feature space: z i =E i (x i ); where x i is the sensor data of the i-th modality, E i is the encoder of this mode, z i is the potential representation vector.

[0173] Step 3: Self-attention fusion mechanism. Use the self-attention mechanism to fuse different modal features:

[0174] (1) Calculate the query, key, and value vectors: Q i =W Q ·z i ;K j =W k ·z j ; V j =W V ·z j Where W Q 、W k 、W V is a learnable weight matrix with a dimension of 64×64.

[0175] (2) Calculate attention weight:

[0176] A ij =softmax(Q i ·K j ^T / ), where d is the dimension of the key vector, usually 64, and √d is a scaling factor to prevent the vanishing gradient problem.

[0177] (3) Fusion feature vector: z fusion =∑ i ∑ j A ij ·V j .

[0178] (4) Apply LayerNorm regularization: z norm =LayerNorm(z fusion ).

[0179] The above attention mechanism allows the system to dynamically adjust the importance of each modal data and monitor the most relevant sensor data according to the actual situation.

[0180] Step 4: Adversarial training mechanism.

[0181] Introduce the discriminator D and build the adversarial training framework: (1) Discriminator structure: 3-layer MLP, input dimension is 64, hidden layer size is [128, 64, 32], output dimension is 1, representing the true probability. (2) Adversarial loss calculation: L adv =E x [logD(z real )]+E z [log(1-D(z fusion ))]. Where z real is the encoding of the real sample, z fusion is the fused feature. (3) Gradient penalty term: L gp =E z [(|| ||2-1)^2], where ∇z D(z) is the gradient of the discriminator with respect to the input. (4) Update the discriminator: by minimizing -L adv +λ gp ·L gp Update the discriminator parameters.

[0182] (5) Update the encoder: by minimizing L adv Update encoder parameters.

[0183] The above adversarial training strategy improves the generalization ability of the model in the case of data distribution shift, enabling the system to adapt to new users or new environments.

[0184] Step 5: Reconstruction and task learning. Design a multi-task learning framework to optimize reconstruction, classification, and energy prediction tasks simultaneously: (1) Reconstruction loss: L recon =∑ i ||x i -D i (z fusion )||^2; where D i is the decoder of the i-th mode, and its structure is symmetrical with the encoder. (2) Touch task loss: L task =CrossEntropy(f(z fusion ),y); where f is the touch intention classifier, using a 2-layer MLP structure; y is the touch intention label. (3) Energy prediction loss: L energy =MSE(g(z fusion ), e); where g is the energy predictor, which adopts a 2-layer MLP structure; e is the actual energy consumption label. (4) Overall loss function: L total =λ1·L recon +λ2·L task +λ3·L energy +λ4·L adv ; where λ1,λ2,λ3,λ4 are weight coefficients, usually set to [0.5,1.0,0.5,0.1]. Through multi-task learning, the system can optimize multiple objectives simultaneously and improve overall performance.

[0185] Step 6: Missing modality processing mechanism. Design a conditional generative model under missing modality: (1) Construct a modality missing mask m, where each element m i ∈{0,1} indicates whether the i-th modality is available. (2) Conditional generation network structure: 2-layer MLP, input dimension is 64×(number of modalities-1)+number of modalities, hidden layer size is 128, output dimension is 64. (3) Predict missing modality features: z i ^pred=G(z j ,m). Among them, z jis the feature of the available modality, and G is the conditional generation network. (4) Conditional generation loss: Lgen=∑ i ||z i -G(z{-i},m i )||^2. Where z{-i} represents all modal features except i. (5) Fusion of missing modal features: z fusion ={∑ i z i / N,ifallmodalitiesavlable;∑ i m i ·z i / ∑m i ,ifsomemodalitiesmissing}; The above missing modalities processing mechanism improves the fault tolerance of the system. Even if some sensors fail, the system can still maintain high performance.

[0186] Step 7: Dynamic energy allocation model. Predict energy demand based on fusion features and perform component-level allocation: (1) Energy prediction network structure: 2-layer MLP, input dimension is 64 + environmental feature dimension, hidden layer size is 64, output dimension is 1. (2) Predict energy demand: E pred =h(z fusion ,c). Where h is the energy prediction network and c is the environmental context feature. (3) Component energy allocation network structure: 2-layer MLP, input dimension is 64, hidden layer size is 64, and output dimension is the number of components. (4) Component-level energy allocation: E i =E pred ·softmax(MLP(z fusion )) i . Among them, softmax(MLP(z fusion )) i Represents the energy distribution ratio of the i-th component.

[0187] Step 8: Model deployment and optimization. To adapt to the resource constraints of embedded devices, the following optimization measures are taken: (1) Model quantization: Quantize 32-bit floating-point weights to 8-bit integers, reducing the model size by 75%. (2) Model pruning: Remove connections with importance below a threshold, which can typically reduce the number of parameters by 30-50% with a performance degradation of no more than 1%. (3) Hierarchical execution strategy: Dynamically adjust the model complexity based on the device state, including: Lightweight mode: Only some modes and a simplified network are used, with low power consumption and moderate performance; Standard mode: Uses the main modes and the complete network, with medium power consumption and good performance; High-performance mode: Uses all modes and an enhanced network, with high power consumption and optimal performance.

[0188] The implementation of the heterogeneous multimodal fusion autoencoder algorithm requires accelerator support and uses 8-bit fixed-point operations. Its inference time is typically 15-30ms, and its power consumption is 30-50mW. To reduce energy consumption during daily use, the system only activates this algorithm when high-precision recognition is required or in complex environments, typically accounting for 10-20% of the total system operation time.

[0189] Actual tests have shown that the heterogeneous multimodal fusion autoencoder algorithm achieves a touch recognition accuracy of 99.3% under standard conditions, a 1-2% improvement over single-modality approaches. The accuracy improvement is even more significant in extreme environments (high temperature, high humidity, and strong vibration), reaching 5-10%. The algorithm demonstrates high fault tolerance, particularly in the case of partial sensor failure: a single sensor failure results in a performance drop of no more than 15%, while traditional approaches experience a performance drop of over 50%. Furthermore, the algorithm significantly improves the system's adaptability to new users, requiring only approximately 20 interactions to adapt to a new user, approximately 80% faster than traditional approaches.

[0190] like Figure 7 and Figure 8 As shown, the system of the present invention adopts a layered heterogeneous computing architecture and a prediction-verification dual-path processing strategy to achieve deep collaboration and optimized integration of multiple algorithms. The details are as follows:

[0191] First, the layered heterogeneous computing architecture consists of three main layers:

[0192] The low-power resident core layer is based on an ARM Cortex-M0+ processor with a clock frequency of 16MHz. It is primarily responsible for basic signal monitoring and wake-up determination. This core runs continuously but uses an extremely low-power design, including dynamic clock gating, which activates the clock only when processing is required; power gating, which completely powers down unused peripherals; and optimized data processing, which uses an event-driven model and spends most of its time in a dormant state.

[0193] Through the above technology, the core's power consumption is controlled below 1mW, and it can continuously monitor proximity sensor and environmental sensor data, and determine whether the main computing core needs to be woken up based on preset thresholds.

[0194] The main computing core layer is based on an ARM Cortex-M4F processor with a clock frequency of 80-120MHz. It is primarily responsible for executing the core functions of the phase-space reconstruction touch intent recognition algorithm and the entropy-driven adaptive energy allocation algorithm. This core utilizes multi-level power management: dynamically adjusting operating frequency and voltage based on computational load; optimizing algorithm-specific cache strategies to reduce memory access overhead; and using SIMD (Single Instruction, Multiple Data) instructions to accelerate vector calculations. These technologies keep power consumption within the 5-30mW range, enabling efficient execution of moderately complex algorithmic tasks.

[0195] Accelerator layer: uses a dedicated neural network processing unit that supports 8-bit integer quantization operations, mainly used to handle complex calculations such as spectral summary touch intention prediction and heterogeneous multimodal fusion. The accelerator adopts an on-demand activation strategy:

[0196] Task chunking: breaking down large neural networks into multiple independently executable chunks;

[0197] Result cache: cache intermediate results to control repeated calculations; sparse calculation: take advantage of model sparsity to skip zero-weight calculations.

[0198] Through the above technology, the accelerator consumes 30-100mW of power in the active state, but is only activated when the system requires high-precision predictions or processes complex environments, which typically accounts for 10-20% of the total operating time.

[0199] The computing units are coordinated through shared memory and interrupt mechanisms, and data exchange follows the following rules:

[0200] The low-power core and the main core exchange data through a 32KB shared SRAM and use a double buffer mechanism to control conflicts;

[0201] The main core and accelerator exchange data through 128KB shared SRAM, supporting zero-copy data transmission;

[0202] Interrupt priority setting: proximity sensor interrupt > touch interrupt > environmental sensor interrupt > timer interrupt;

[0203] The power state transition is controlled by a dedicated state machine to ensure smooth transition and control peak current.

[0204] Secondly, the prediction-verification dual-path processing strategy achieves deep collaboration between algorithms. The specific mechanism is as follows:

[0205] Prediction Path: This is implemented by the spectral summary touch intention prediction algorithm and the quantum excitation zero-latency touch prediction module. When the proximity sensor detects the user's finger approaching (usually within 1-3 cm), the prediction path is activated and the following steps are performed:

[0206] Construct a priori probability map P based on historical data prior (x, y, t); Combined with the environmental context, analyze the approach trajectory and speed; Predicted touch position and intention P predict(x, y, t + Δt), Δt is usually 50-100ms; generate a touch probability flow field, usually covering 10-20% of the screen area; pre-allocate system resources, including rendering UI responses in advance and pre-loading functional modules used; adjust the touch sensitivity distribution to enhance the detection sensitivity in high-probability areas. Verification path: implemented by the phase space reconstruction touch intention recognition algorithm. When an actual touch occurs, the verification path is activated and the following steps are performed: collect high-precision touch signals x(t); construct phase space trajectory X(t); extract dynamic features [LLE, CD, RR, DET, ...]; perform touch intention classification I real (t); Calculate real-time confidence Conf real (t).

[0207] Feedback adjustment mechanism: A closed-loop feedback is formed between the two paths, and the following adjustments are performed: Compare the predicted results with the actual results: ΔP=||P predict -I real ||; Update prediction model parameters: θ new =θ old -η· ; Adjust the prediction time window: Δt new =f(ΔP,Conf real ); optimize resource allocation strategies and dynamically adjust the proportion of pre-allocated resources based on prediction accuracy.

[0208] Fusion decision generation:

[0209] Based on the results of the two paths, the final decision is generated:

[0210] D final (t)=w predict ·D predict (t-Δt)+w Verify ·D Verify (t)+w context ·D context (t).

[0211] The weight coefficient is adjusted dynamically:

[0212] w predict =f p (SNR,H(S),Stability predict )

[0213] w Verify =f V (SNR,H(S),Confidence Verify )

[0214] w context =f c (Reliabilitycontext ,Δt)

[0215] When the prediction accuracy is high and the environment is stable, w predict As the value increases, the system responds faster based on the prediction results; when the environment is complex and the prediction is unstable, w Verify The value increases, and the system ensures recognition accuracy based on the verification results.

[0216] The key to the collaborative integration of multiple algorithms lies in the deep cross-integration between the algorithms, which is mainly reflected in the following aspects:

[0217] Phase space-spectrum joint feature extraction: PS-TII and WSC-TIP algorithms implement cross-enhancement in the feature extraction stage, such as:

[0218] LLE enhanced =LLE base (1+k stft M3);

[0219] DET enhanced =DET base (1+k ceps ·Cp);

[0220] M 1,enhanced =M 1base (1+k rqa ENTR).

[0221] The above cross-enhancement enables the system to simultaneously capture the dynamic characteristics, frequency domain characteristics and cepstrum characteristics of the signal, greatly improving the feature expression ability and anti-interference ability.

[0222] Entropy-driven multimodal energy allocation: EAEA and HMFAE algorithms are cross-integrated at the energy management level to build an accurate energy allocation framework based on multi-dimensional information, such as:

[0223] H joint =w s ·H(S)+w U ·H(U)+w M H(M);

[0224] E pred =E base ·(1+α·H joint )·(1+β· )·(1+δ·Confidence intent )·(1+ε·P predict ).

[0225] The above fusion mechanism achieves refined energy allocation while considering system status, user behavior, multimodal sensing information and touch prediction probability.

[0226] Quantum Inspiration Model and Multimodal Fusion: QI-ZDT and HMFAE algorithms are integrated at the modeling and reasoning levels to construct a quantum-inspired multimodal representation space.

[0227] ;H fusion (t)=H user (t)+H env (t)+H modal (t); H modal (t)=g(z fusion ).

[0228] The above fusion applies quantum computing concepts to multimodal information processing to achieve efficient probability prediction and state representation. In addition, Figure 10 This is a schematic diagram of an actual product of the present invention.

[0229] The system of the present invention significantly outperforms existing technologies in terms of touch accuracy, environmental adaptability, energy efficiency, response speed, user adaptability, and fault tolerance, fully demonstrating the technical advantages of deep multi-algorithm fusion. The performance improvement is particularly significant under extreme environmental conditions, demonstrating that the present invention has achieved breakthrough progress in solving the core problems of traditional touch technology.

[0230] The screen intelligent key switch system proposed in this invention shows high performance in various environments through the core fusion of the phase space reconstruction touch intention recognition algorithm and the entropy-driven adaptive energy allocation algorithm, combined with the enhancement of spectral summary touch intention prediction and heterogeneous multimodal fusion autoencoding algorithm, and has broad application prospects.

[0231] The following calculation process is used to verify the screen intelligent key switch system described in the present invention.

[0232] 1. Calculation parameter setting

[0233] The touch device, with the smart home control panel as the main body, has a size of 120mm×85mm×18mm (width×depth×height), and is equipped with components such as a capacitive touch screen, a proximity sensor, an acceleration sensor, a temperature sensor, an ambient light sensor, a main control processor, and an accelerometer. The device adopts the structure described in the present invention, including a touch sensing unit, a communication interface module, a data processing unit, a control execution unit, a display unit, and a power management unit. The device has a total of 6 functional areas, each of which is equipped with different functional components. The main control processor with higher power is located in the third functional area. To verify the effectiveness of the proposed method, the following detailed calculation process is now carried out. The calculation is carried out based on the phase space reconstruction touch intention recognition algorithm and the entropy driven adaptive energy allocation algorithm respectively. Through the synergistic effect of the two algorithms, accurate prediction of touch intention and optimal energy allocation are achieved.

[0234] 2. Initial parameter setting

[0235] 2.1 Environmental and Operating Conditions: Ambient temperature: 25°C; Relative humidity: 65%; Atmospheric pressure: 101.3 kPa; Total power consumption within the device: 180 mW; Power consumption distribution of individual components: Display (50-100 mW), main control processor (30-80 mW), touch controller (15-25 mW), communication module (5-50 mW), sensor module (3-8 mW), power management unit (2-5 mW)

[0236] 2.2 Electrical Parameters: Capacitive touch sensor sensitivity: 0.05pF; signal sampling frequency: 120Hz; ADC resolution: 16-bit; power supply voltage: 3.7V (lithium battery) / 3.0V (button battery); standby current: <150μA; operating current: <50mA; touch threshold reference value: 0.5pF; temperature drift coefficient: 0.02pF / °C; humidity drift coefficient: 0.03pF / %RH

[0237] 2.3 System geometry parameters: Touch screen size: 95mm×65mm, resolution 240×320 pixels; touch sensing area: 12×8 dots, each touch point area is approximately 64mm 2 Electrode spacing: 7.5mm; Electrode line width: 0.1mm; Sensing effective area: calculation formula S = 95 × 65-(12 × 8 × 0.1 × 7.5) = 5570mm 2

[0238] 3. Phase Space Reconstruction Touch Intent Recognition Algorithm Calculation Process

[0239] 3.1 Feature Extraction Network Design: A phase space feature extraction network for touch intent recognition was constructed with the following structure: time delay embedding dimension m = 4, time delay parameter τ = 4 sampling points, and sliding window length of 32 sampling points. The cyclic quantitative analysis parameter ε was set to 12% of the phase space diameter, and the minimum diagonal length l was set to 12%. min = 2. Total number of algorithm parameters: (4×4+12+8+5)=41 parameters.

[0240] 3.2 Training data preparation: To train the touch intention classifier, the following data was prepared: 500 sets of intentional touch samples (including different operations such as clicking, long pressing, and sliding); 300 sets of unintentional touch / interference samples (including environmental interference, wet hand touch, vibration interference, etc.); a total of 800 sets of training data, each set containing a touch signal sequence at 32 time points.

[0241] 3.3 Algorithm parameter settings: basic threshold value T0: 0.5pF; adjustment coefficient α: 1.2; maximum Lyapunov exponent threshold LLE max : 2.0; time delay parameter τ: 4; embedding dimension m: 4; circulant matrix threshold parameter ε: 0.12D (D is the phase space diameter)

[0242] 3.4PS-TII algorithm calculation process

[0243] Step 1: Touch signal acquisition and preprocessing Taking the actual acquired touch signal as an example, the raw touch data fragment is as follows (unit: pF):

[0244] x(t)=[0.12,0.15,0.27,0.43,0.62,0.78,0.89,0.95,0.93,0.86,0.75,0.61,0.48,0.33,0.21,0. 14,0.10,0.08,0.07,0.08,0.09,0.11,0.12,0.13,0.12,0.11,0.10,0.09,0.08,0.07,0.06,0.05]

[0245] After applying a 4th-order Butterworth low-pass filter (cut-off frequency 30Hz):

[0246] x filtered (t)=[0.13,0.16,0.28,0.44,0.63,0.79,0.88,0.94,0.92,0.85,0.74,0.60,0.47,0.32,0.20,0. 13,0.09,0.07,0.07,0.08,0.09,0.11,0.12,0.13,0.12,0.11,0.10,0.09,0.08,0.07,0.06,0.05]

[0247] Signal normalization processing:

[0248] x norm (t)=(x filtered (t)-0.05) / (0.94-0.05)=[0.09,0.12,0.26,0.44,0.65,0.83,0.93,1.00,0.98,0.90,0.78,0.62,0.47,0.3 0,0.17,0.09,0.04,0.02,0.02,0.03,0.04,0.07,0.08,0.09,0.08,0.07,0.06,0.04,0.03,0.02,0.01,0.00]

[0249] Step 2: Phase space reconstruction uses the time delay embedding method to construct the phase space trajectory. Taking t=12 as an example: X(12)=[x norm (12),x norm (8),x norm (4),x norm (0)]=[0.62,0.98,0.65,0.09]

[0250] Similarly, the phase space coordinates of other time points are calculated, and 32-4+1=29 phase space points are obtained: X(0)=[0.09,0.00,0.00,0.00]X(1)=[0.12,0.01,0.00,0.00]...X(28)=[0.00,0.03,0.06,0.09]

[0251] Step 3: Dynamic feature extraction (1) Calculation of the maximum Lyapunov exponent (LLE): Select an initial pair of similar points, such as X(12) and X(13). The initial distance ||δX(0)||=0.17. After 8 time steps, the distance ||δX(8)||=0.53LLE=(1 / 8)×ln(0.53 / 0.17)=0.142.

[0252] (2) Correlation dimension (CD) calculation: different r values ​​are selected to calculate the correlation integral C(r), such as r = [0.05, 0.1, 0.2, 0.4, 0.8] corresponding to C(r) = [0.03, 0.08, 0.22, 0.51, 0.95]. A linear fit is performed on ln(C(r)) and ln(r), and the slope is CD = 1.72.

[0253] (3) Circulation Quantitative Analysis (RQA): The circulation matrix R(i, j) was constructed and the following statistics were obtained: the circulation rate RR = 0.33, indicating that 33% of the point pairs have a distance less than the threshold; the certainty DET = 0.65, indicating that 65% of the circulation points form a diagonal structure; the laminarity LAM = 0.47, indicating that 47% of the circulation points form a vertical line structure; the entropy ENTR = 1.38, indicating the complexity of the diagonal length distribution.

[0254] Step 4: Construct the feature vector PS Features =[0.142,1.72,0.33,0.65,0.47,1.38,...] Partial values ​​of the first touch feature vector are: [0.142,1.72,0.33,0.65,0.47,1.38].

[0255] Step 5: Touch intent classification uses a pre-trained SVM classifier to classify the intent and obtain the classification result: I(t) = "Intentional touch-click operation" classification probability distribution:

[0256] P(click)=0.87, P(long press)=0.08, P(slide)=0.03, P(interference)=0.02 Classification confidence: Confidence=1-(-0.87×log20.87-0.08×log20.08-0.03×log20.03-0.02×log20.02)=1-0.67=0.83.

[0257] Step 6: Trajectory Stability Evaluation Trajectory Stability Indicator: Trajectory s tability=1-(0.142 / 2.0)=0.929.

[0258] Step 7: Adaptive threshold decision Assume that the current system state fluctuation index σ s (t)=0.15 Dynamic threshold: T k =0.5×[1+1.2×0.15]=0.59 The eigenvalue 0.83>0.59, which is determined to be a valid touch intention.

[0259] Step 8: Adaptive Window Mechanism The current system load is low, f(S load )=0.2 Window length adjustment: L=32×[1+0.3×0.2]=33.92≈34 sampling points Window function adjustment: w adaptive (n)=w base (n) × 1.05 The next calculation uses a 34-point window and the adjusted window function.

[0260] 3.5 Adaptive threshold mechanism: When the sensor detects that the ambient humidity is 85%, the system state fluctuation index σ s(t)=0.35 Dynamic threshold: T k =0.5×[1+1.2×0.35]=0.71At this time, the characteristic value needs to be greater than 0.71 to be determined as a valid touch, which effectively prevents false touches in wet environments.

[0261] 3.6 Phase Space Reconstruction Touch Intent Recognition Algorithm Results: After processing with the PS-TII algorithm, touch intent recognition results were obtained under standard conditions. Touch location: (120, 85) pixels, touch type: click, confidence: 0.83, trajectory stability: 0.929, recognition accuracy in a noisy environment (vibration acceleration of 2g): 96.2%, recognition accuracy with wet hands: 95.1%, and recognition accuracy in an extremely low signal-to-noise ratio (-20dB): 92.3%.

[0262] 4. Entropy-driven adaptive energy allocation algorithm calculation process

[0263] 4.1 System State Model and Initial Parameters: A system state space is established based on the actual device state, consisting of eight states: deep sleep, light sleep, standby, lightly active, moderately active, highly active, fully loaded, and abnormal. The current system state distribution is: P(s) = [0.05, 0.15, 0.25, 0.30, 0.15, 0.08, 0.02, 0.00]. User operation modes include six types: short press, long press, fast swipe, slow swipe, multi-touch, and no operation. The current user operation distribution is: P(u) = [0.45, 0.15, 0.20, 0.10, 0.05, 0.05].

[0264] 4.2 State entropy calculation

[0265] Step 1: Calculate the system state entropy

[0266] H(S)=-ΣP(s i )×log2(P(s i ))

[0267] =-0.05×log2(0.05)-0.15×log2(0.15)-0.25×log2(0.25)-0.30×log2(0.30)-0.15×log2(0.15)-0.08×log2(0.08)-0.02×log2(0.02)-0.00×log2(0.00)=-((-4.32)×0.05+(-2.7 4)×0.15+(-2.00)×0.25+(-1.74)×0.30+(-2.74)×0.15+(-3.64)×0.08+(-5.64)×0.02+0)=-((-0.216)+(-0.411)+(-0.500)+(-0.522)+(-0.411)+(-0.291)+(-0.113)+0)=2.464.

[0268] Step 2: Use pattern entropy calculation

[0269] H(U)=-ΣP(u j )×log2(P(u j ))

[0270] =-0.45×log2(0.45)-0.15×log2(0.15)-0.20×log2(0.20)-0.10×log2(0.10)-0.05×log2(0.05)-0.05×log2(0.05)=-((-1.15)×0.45+(-2.74)×0.15+(-2.32)×0.20+(-3.32)×0.10+(-4.32)×0.05+(-4.32)×0.05)=-((-0.518)+(-0.411)+(-0.464)+(-0.332)+(-0.216)+(-0.216))=2.157.

[0271] Step 3: Entropy gradient calculation Assume that the system state entropy H(S) at the previous moment {t-Δt} )=2.325 entropy gradient: =H(S t )-H(S {t-Δt} )=2.464-2.325=0.139.

[0272] Assume that the usage pattern entropy H(U {t-Δt} )=2.032 Entropy gradient: =H(U t )-H(U {t-Δt} )=2.157-2.032=0.125.

[0273] 4.3 Energy demand forecast: basic energy demand Ebase =15mW State entropy weight coefficient α=0.3 Entropy change rate weight coefficient β=0.5 Usage mode entropy weight coefficient γ=0.2 Intent confidence weight coefficient δ=0.2 Touch intent confidence intent =0.83 (provided by PS-TII algorithm).

[0274] Energy demand forecast calculation:

[0275] E pred (t+Δt)=E base ×(1+α×H(S))×(1+β× )×(1+γ×H(U))×(1+δ×Confidence intent )=15×(1+0.3×2.464)×(1+0.5×0.139)×(1+0.2×2.157)×(1+0.2×0.83)=15×1.7392×1.0695×1.4314×1.166=40.69mW.

[0276] 4.4 Component Priority Calculation

[0277] The device consists of 6 main components: display screen, main control processor, touch controller, communication module, sensor module, power management unit; importance weight: P importance =[0.25,0.30,0.20,0.15,0.05,0.05]; Urgency weight: P urgency =[0.35,0.25,0.20,0.10,0.05,0.05];

[0278] Energy efficiency ratio weight: P efficiency =[0.15,0.20,0.25,0.15,0.15,0.10].

[0279] Weight coefficient: w1=0.4, w2=0.4, w3=0.2;

[0280] Component priority calculation: P1 = 0.4 × 0.25 + 0.4 × 0.35 + 0.2 × 0.15 = 0.28; P2 = 0.4 × 0.30 + 0.4 × 0.25 + 0.2 × 0.20 = 0.26; P3 = 0.4 × 0.20 + 0.4 × 0.20 + 0.2 × 0.25 = 0.21; P4 = 0.4 × 0.15 + 0.4 × 0.10+0.2×0.15=0.13; P5=0.4×0.05+0.4×0.05+0.2×0.15=0.07; P6=0.4×0.05+0.4×0.05+0.2×0.10=0.05; Component priority: P=[0.28,0.26,0.21,0.13,0.07,0.05], ΣP j =1.00.

[0281] 4.5 Component Energy Distribution

[0282] Minimum working energy of each component: E min =[5mW,8mW,5mW,2mW,1mW,1mW], ΣE min (j) = 22mW; Distributable energy: E pred (t)-ΣE min (j)=40.69-22=18.69mW.

[0283] Component energy allocation calculation: E1=5+18.69×0.28 / 1.00=10.23mW; E2=8+18.69×0.26 / 1.00=12.86mW; E3=5+18.69×0.21 / 1.00=8.92mW; E4=2+18.69×0.13 / 1.00=4.43mW; E5=1+18.69×0.07 / 1.00=2.31mW; E6=1+18.69×0.05 / 1.00=1.93mW. Component energy allocation result: E=[10.23,12.86,8.92,4.43,2.31,1.93]mW.

[0284] Total energy consumption: ΣE i =40.68mW.

[0285] 4.6 Entropy response function design: critical entropy value H critical =0.5; sensitivity coefficient k=10; current system entropy H(S)=2.464.

[0286] Entropy response function calculation: f entropy (H,H critical )=1 / (1+exp(-k×(HH critical )))

[0287] =1 / (1+exp(-10×(2.464-0.5)))=1 / (1+exp(-19.64))=1 / (1+0.0000000003)≈1.00.

[0288] Since the current system entropy is much larger than the critical value and the entropy response function is close to 1, the system will greatly improve the flexibility of energy distribution.

[0289] 4.7 Predictive Sleep-Wake Management: Based on the current system state entropy, usage pattern entropy and their gradient, predict the future wake-up probability: P wakeup (t+10s)=0.35; Due to the low wake-up probability, the system is set to a light sleep state (power consumption 0.5mW). Assuming the next predicted usage time is 30 seconds later, the wake-up time is 200ms, and the safety margin is 100ms. Pre-wake-up timing: T prewake =T current +max(0,30000-200-100)=T current +29700ms.

[0290] 4.8 Global Optimization Solution

[0291] Use gradient descent method to solve multi-objective optimization problem: minJ(x)=α·E total (x)+β·T response (x)+γ·(1-A accuracy (x))+δ·H system (x);

[0292] Current weight coefficients: α=0.4, β=0.3, γ=0.2, δ=0.1; Optimization variables: x=[power allocation, processor frequency, touch sensitivity, communication bandwidth]; Constraints: E total (x)≤E budget =50mW,T response (x)≤T max =5ms,A accuracy (x)≥A min =0.95.

[0293] First iteration:

[0294] x1=[40.68mW,100MHz,0.5pF,50kbps];

[0295] J(x1)=0.4×(40.68 / 50)+0.3×(3 / 5)+0.2×(1-0.97)+0.1×(2.464 / 3)=0.4×0 .814+0.3×0.6+0.2×0.03+0.1×0.821=0.3256+0.18+0.006+0.0821=0.5937.

[0296] Second iteration:

[0297] x2=[38.5mW,95MHz,0.55pF,40kbps];

[0298] J(x2)=0.4×(38.5 / 50)+0.3×(3.2 / 5)+0.2×(1-0.96)+0.1×(2.35 / 3)=0.4×0 .77+0.3×0.64+0.2×0.04+0.1×0.783=0.308+0.192+0.008+0.0783=0.5863.

[0299] Third iteration (final solution):

[0300] x3=[35.8mW,90MHz,0.52pF,30kbps];

[0301] J(x3)=0.4×(35.8 / 50)+0.3×(3.5 / 5)+0.2×(1-0.95)+0.1×(2.28 / 3)=0.4× 0.716+0.3×0.7+0.2×0.05+0.1×0.76=0.2864+0.21+0.01+0.076=0.5824.

[0302] Optimal solution: power consumption 35.8mW, processor frequency 90MHz, touch sensitivity 0.52pF, communication bandwidth 30kbps.

[0303] 4.9 Calculation results of the entropy-driven adaptive energy allocation algorithm: After optimization based on the EAEA algorithm, the energy allocation plan was obtained: display: 9.5mW; main control processor: 11.3mW; touch controller: 7.8mW; communication module: 3.2mW; sensor module: 2.2mW; power management unit: 1.8mW; total power consumption: 35.8mW; standby power consumption: 0.5mW deep sleep power consumption: 0.05mW. It is expected to be used 100 times a day (5-10 touch operations each time), and the battery life using a CR2032 button battery (225mAh) is: 3.7 years.

[0304] 5. Algorithm Collaboration Process

[0305] 5.1 Multi-dimensional Perception Fusion Computing: Fusion of the Phase Space Reconstruction Touch Intention Recognition Algorithm and the Spectral Summary Touch Intention Prediction Algorithm:

[0306] Step 1: Feature cross enhancement based on the Lyapunov exponent LLE obtained by the PS-TII algorithm base=0.142; spectrum skewness M3=1.25 based on the WSC-TIP algorithm; spectrum enhancement coefficient k stft =0.2 enhanced Lyapunov exponent:

[0307] LLE enhanced =LLE base (1+k stft ·M3)=0.142·(1+0.2×1.25)=0.142×1.25=0.1775.

[0308] Deterministic feature DET obtained based on PS-TII algorithm base =0.65; Cepstral peak value Cp obtained based on WSC-TIP algorithm = 2.3; Cepstral enhancement coefficient k ceps =0.15; enhanced deterministic features: DET enhanced =DET base (1+k ceps ·Cp)=0.65·(1+0.15×2.3)=0.65×1.345=0.874.

[0309] Spectral center M obtained based on WSC-TIP algorithm 1base =15.3Hz; RQA entropy ENTR obtained based on PS-TII algorithm = 1.38; RQA enhancement coefficient k rqa =0.1 enhanced spectrum center of gravity:

[0310] M 1,enhanced =M 1base (1+k rqa ·ENTR)=15.3·(1+0.1×1.38)=15.3×1.138=17.41Hz.

[0311] Step 2: Joint feature vector construction: F joint =[0.1775,1.72,0.33,0.874,0.47,1.38,17.41,8.5,1.25,2.3,35].

[0312] Step 3: Prediction-Verification Dual Path Processing When the proximity sensor detects a finger approaching (2.5 cm away), the touch intention prediction algorithm based on the spectrum summary predicts: P predict (click) = 0.82, P predict (Long press) = 0.12, P predict (sliding)=0.04,P predict (interference) = 0.02 Predicted touch position: (125±10, 90±10) pixels Prediction confidence: 0.75.

[0313] When an actual touch occurs, the touch intention recognition algorithm is reconstructed based on the phase space to identify: real (click)=0.87,I real (Long press)=0.08,I real (slide)=0.03,I real (Interference) = 0.02 Actual touch position: (120,85) Pixel recognition confidence: 0.83.

[0314] Comparison of predicted results with actual results:

[0315] ΔP=||P predict -I real ||=√((0.82-0.87) 2 +(0.12-0.08) 2 +(0.04-0.03) 2 +(0.02-0.02) 2 )=0.063;

[0316] Position error: √((125-120) 2 +(90-85) 2 )=7.07 pixels.

[0317] Step 4: Dynamic weight calculation: Current signal-to-noise ratio SNR = 15dB, system state entropy H(S) = 2.464, predict stability predict =0.85 Weight calculation: w predict =f p (SNR,H(S),Stability predict )=0.4; w Verify =f V (SNR,H(S),Confidence Verify )=0.5;

[0318] w context =f c (Reliability context ,Δt)=0.1.

[0319] Final decision fusion:

[0320] D final (t)=0.4·[0.82,0.12,0.04,0.02]+0.5·[0.87,0.08,0.03,0.02]+0.1·[0.80,0.10,0.05,0.05]=[0.844,0.096,0.035,0.025]. Final decision: click, confidence level 0.844.

[0321] 5.2 Entropy-driven multimodal energy allocation mechanism: Integrating the entropy and characteristics of PS-TII, WSC-TIP, and HMFAE algorithms to build a multi-dimensional energy demand prediction model:

[0322] Step 1: Calculate the multidimensional state entropy: System state entropy H(S) = 2.464; usage mode entropy H(U) = 2.157; multimodal information entropy (provided by HMFAE) H(M) = 1.85; entropy weight coefficient: w s =0.5,w U =0.3,w M =0.2; Joint entropy:

[0323] H joint =w s ·H(S)+w U ·H(U)+w M ·H(M)=0.5×2.464+0.3×2.157+0.2×1.85=2.274.

[0324] Step 2: Forecast energy demand Basic energy demand E base =15mW; joint entropy weight coefficient α=0.3; entropy gradient coefficient β=0.5; joint entropy gradient =0.15; touch intention confidence intent =0.844; predicted probability coefficient ε=0.2; predicted probability P predict =0.82;

[0325] E pred =E base ·(1+α·H joint )·(1+β· )·(1+δ·Confidence intent )·(1+ε·P predict )=15×(1+0.3×2.274)×(1+0.5×0.15)×(1+0.2×0.844)×(1+0.2×0.82)=15×1.682×1.075×1.169×1.164=37.45mW.

[0326] Step 3: Probabilistic Flow-Driven Spatial Energy Allocation. Based on the touch probability flow field, energy allocation weights for different touchscreen regions are calculated: high-probability region (25×25 pixels surrounding the touch point): weight 0.6; medium-probability region (50×50 pixels surrounding the touch point): weight 0.3; low-probability region (rest of the touchscreen): weight 0.1. Spatially differentiated energy allocation: high-probability region: base scanning frequency 120Hz, sensitivity increased by 15%; medium-probability region: base scanning frequency 60Hz, sensitivity increased by 5%; low-probability region: base scanning frequency 30Hz, sensitivity unchanged. Through spatially differentiated scanning, overall touchscreen energy consumption is reduced by 42%, while maintaining high responsiveness in high-probability regions.

[0327] 5.3 Heterogeneous Multimodal Fusion and Fault-Tolerant Mechanism Calculation: When the temperature sensor fails, the system predicts the temperature data through the heterogeneous multimodal fusion autoencoding algorithm:

[0328] Step 1: Modal features can be used to extract the capacitance data feature vector: z1=[0.35, 0.62, 0.44, 0.28, ..., 0.15]; the acceleration data feature vector: z2=[0.12, 0.05, 0.03, 0.08, ..., 0.11]; the ambient light data feature vector: z4=[0.75, 0.72, 0.68, 0.65, ..., 0.70].

[0329] Step 2: Missing mode prediction Temperature modal feature prediction: z3^pred=G(z1,z2,z4,m)=[0.41,0.38,0.36,0.35,...,0.37].

[0330] Step 3: Multimodal fusion uses self-attention mechanism to fuse multimodal features: z fusion =Attention([z1,z2,z3^pred,z4])=[0.42,0.45,0.38,0.32,...,0.28].

[0331] Step 4: Performance Comparison: Touch accuracy when the temperature sensor is working properly: 99.3%; touch accuracy when the temperature sensor fails but predicted data is used: 98.6%. This performance degradation is only 0.7%, far lower than the performance degradation of traditional methods when the sensor fails (>50%).

[0332] like Figure 9 As shown in 5.4 Quantum Excitation Zero-Delay Touch Prediction Calculation

[0333] Step 1: Quantum state representation: The touch state is represented as a superposition of quantum states. Take five touch positions as an example: :Click position (120,85) :Click position (180,120) :Click position (60,150) :Click position (200,200) : Click position (30,30).

[0334] Initial quantum state: =0.6 +0.3 +0.05 +0.03 +0.02 The amplitude reaches: 0.6 2 +0.3 2 +0.05 2 +0.03 2 +0.02 2 =0.36+0.09+0.0025+0.0009+0.0004=0.4538≈0.45 (normalization coefficient adjustment).

[0335] Step 2: System Hamiltonian constructs user behavior Hamiltonian: H user (t)=0.7H pattern1 +0.2H pattern2 +0.1H pattern3 ; Environmental Hamiltonian: H env (t)=0.5S t emp+0.3S accel +0.2S light ; Multimodal enhanced Hamiltonian: H modal (t)=g(z fusion ).

[0336] Step 3: Quantum state evolution operator: U(t,Δt)=exp(-iH fusion (t)Δt); the quantum state after time evolution: =0.85 +0.12 +0.02 +0.01 +0.0 .

[0337] Step 4: Probabilistic extraction of touch position probability: P(φ1)=|0.85| 2 =0.7225≈0.72; P(φ2)=|0.12| 2 =0.0144≈0.01; P(φ3)=|0.02| 2 =0.0004≈0.00; P(φ4)=|0.01| 2 =0.0001≈0.00;P(φ5)=|0.0|2 =0.0≈0.00 (the sum of the normalized adjusted probabilities is 1).

[0338] Step 5: Resource pre-allocation Based on the touch probability distribution, the system allocates resources in advance: pre-loads the most triggered UI component (position 120, 85), pre-renders the corresponding visual feedback, and adjusts the touch sensitivity of the area to reduce the response delay from the traditional 25ms to <1ms.

[0339] 6. Verification and comparison of calculation results

[0340] To verify the effectiveness of the algorithm, the calculated results are compared with the actual measurement data:

[0341] 6.1 Touch Accuracy Comparison

[0342]

[0343] 6.2 Algorithm Performance Comparison

[0344]

[0345] 6.3 Comparison of Energy Management Effects

[0346]

[0347] 7. Calculation conclusion

[0348] Through the above calculation process, the following conclusions can be drawn:

[0349] The Phase Space Reconstruction Touch Intent Identification Algorithm (PS-TII) can achieve high-precision touch recognition under various environmental conditions through high-dimensional phase space feature extraction. The accuracy rate reaches 99.2% under standard conditions and can still maintain above 92% under extreme conditions, which is about 40% higher than traditional touch algorithms.

[0350] The entropy-driven adaptive energy allocation algorithm (EAEA) achieves accurate energy demand prediction and refined component-level energy allocation through system state entropy and user behavior entropy analysis, reducing the system's active power consumption from the traditional 75mW to 35.8mW and standby power consumption to 0.5mW, extending battery life by approximately 7 times.

[0351] After the deep integration of multiple algorithms, the system performance has been comprehensively improved: the touch accuracy has been increased to 99.95% (standard environment) and 99.8% (extreme environment), the response delay has been reduced to <1ms, the power consumption has been reduced to 28mW, and it has strong fault tolerance. When a single sensor fails, the performance degradation does not exceed 5%.

[0352] Algorithm collaborative optimization is particularly suitable for application scenarios in complex environments: under wet hand operation conditions, the recognition accuracy is improved to 99.2% (an increase of more than 30%); in a strong vibration environment (5g acceleration), the accuracy remains at 97.3% (an increase of more than 50%); in an extreme temperature environment (-40℃ to 85℃), the accuracy remains above 98.5% (an increase of more than 35%).

[0353] The synergistic effect of multi-algorithm fusion is significant: when applied to the control panel of medical equipment, the accuracy rate is improved to 99.9%, the false trigger rate is reduced to 0.01%, and energy consumption is reduced by more than 60%; in industrial control environments, it supports operation with gloves, has a waterproof level of IP67, and can still maintain an identification accuracy rate of more than 95% in strong electromagnetic interference environments.

[0354] In summary, the present invention can effectively improve touch recognition accuracy (by 15%-50%), while reducing system energy consumption (by 50%-90%) and extending battery life (by 5-10 times). It has important technical value and broad market application prospects.

Claims

1. A screen intelligent key switch system, characterized in that: include: Touch sensing unit, used to collect touch signals and environmental parameters; A communication interface module, including a near field communication unit and a long-range communication unit; A data processing unit, configured to execute a phase space reconstruction touch intention recognition algorithm and an entropy driven adaptive energy allocation algorithm; A control execution unit configured to implement touch response and energy management based on the algorithm results; the phase space reconstruction touch intention recognition algorithm executed by the data processing unit is a time delay embedding method based on nonlinear dynamics theory; the entropy driven adaptive energy allocation algorithm executed by the data processing unit is a predictive energy management algorithm based on information entropy theory; The communication interface module adaptively switches the working states of the near field communication unit and the long-range communication unit according to the system state entropy and the communication urgency; The entropy driven adaptive energy allocation algorithm is: The energy demand prediction equation is: E pred (t+Δt)=E base The piecewise linearization of ×(1+α×H(S))×(1+β×▽H(s))×(1+γ×H(U)) is: E component (i)=E min (i)+[E pred (t)-ΣE min (j)]×P i (t) / ΣP j (t), P i (t)=w1×P importance (i)+w2×P urgency (i)+w3×P efficiency (i); The phase space reconstruction touch intention recognition algorithm is expressed as follows: X(t)=[x(t),x(t-τ),x(t-2τ),...,x(t-(m-1)τ)], where x(t) is the sampled touch signal sequence, τ is the time delay parameter, and m is the embedding dimension; the calculation formula for each phase touch intention is: I(t)=Classifier(PS Features ), where PS Features =[LLE,CD,RR,DET,LAM,ENTR,...] is the phase space feature vector, Classifier is the classifier; the phase space reconstruction touch intention recognition algorithm also includes an adaptive window mechanism, and the calculation formula of the window length L is: L=L base ×[1+β·f(S load )]; where L base is the basic window length, β is the adjustment coefficient, f(S load ) is the load characteristic function; the window function is a dynamically adjusted window function: w adaptive (n)=w base (n)·g(P change,rate ), where w base (n) is the basic window function, g(P change,rate ) is the window adjustment function, which is related to the power change rate; Among them, E base is the basic energy demand, α is the state entropy weight coefficient, H(S) is the system state entropy, β is the entropy change rate weight coefficient, ▽H(s) is the rate of change of system state entropy, γ is the usage pattern entropy weight coefficient, H(U) is the usage pattern entropy, E min (i) is the minimum working energy of component i, E pred (t) is the distributable energy, P i (t) represents the priority weight of component i at time t, w1, w2, w3 are weight coefficients, P importance (i) is the importance weight of component i, P urgency (i) is the urgency weight of component i, P efficiency (i) is the energy efficiency ratio weight of component i, LLE is the maximum Lyapunov exponent, CD is the correlation dimension, RQA is the cycle quantitative analysis feature, RR is the cycle rate, DET is determinism, LAM is laminarity, and ENTR is entropy; The phase space reconstruction touch intention recognition algorithm also includes an adaptive threshold decision mechanism: when the characteristic value exceeds the dynamic threshold, it is considered that a valid touch intention is detected; Threshold value T k The calculation formula is: k =T0(k)×[1+α·σ s (t)]; where T0(k) is the basic threshold value, σ s (t) is the system state fluctuation index, α is the adjustment coefficient, and the value range is 0.5-2.0; In the entropy-driven adaptive energy allocation algorithm, the calculation formula of the system state entropy H(S) is: H(S)=-ΣP(s i )×log2(P(s i )); where P(s i ) indicates that the system is in state s i The probability of ; the optimization problem expression is: minJ(x)=α·E total (x)+β·T response (x)+γ·(1-A accuracy (x))+δ·H system (x);stE total (x)≤E budget ,T response (x)≤T max ,A accuracy (x)≥A min ; Among them, α, β, γ, δ are dynamic weight coefficients, E total (x) is the total energy consumption, T response (x) is the response time, A accuracy (x) is the touch accuracy, H system (x) is the system entropy, E budget For energy budget, T max is the maximum allowed response time, A min The minimum allowed accuracy.

2. The screen intelligent key switch system according to claim 1, characterized in that: The system also includes a spectrum summary touch intention prediction module for realizing advanced prediction of touch behavior; the multi-domain feature fusion model expression is: F multi =[W Features ,E Features ,C Features ,T Features ]; where W Features =[EntropyA j ,EnergyD j ,...,EnergyD1] is the wavelet feature, where the wavelet decomposition coefficient {A j ,D j ,D {j-1} ,...,D1} through multi-scale wavelet transform W ψ (a,b)=(1 / √a)∫x(t)·ψ·((tb) / a)dt is calculated; E Features =[M1,M2,M3] is the spectrum feature, M1=∫f·E(f)df / ∫E(f)df is the spectrum center, M2=∫(f-M1)^2·E(f)df / ∫E(f)df is the spectrum width, M3=∫(f-M1)^3·E(f)df / (M2^(3 / 2)·∫E(f)df) is the spectrum skewness; C Features =[Cp,qp] is the cepstrum feature, through cepstrum analysis C x (q)=F^(-1){log|F{x(t)}|} is calculated; T Features is the time domain feature; Among them, EntropyA j is the entropy characteristic of the approximate coefficient, EnergyD j is the energy feature, a is the scale parameter; b is the translation parameter, x(t) is the sampled touch signal sequence, ψ is the wavelet mother function, E(f) represents the spectrum energy distribution, Cp represents the intensity of the periodic signal, qp represents the basic period of the periodic signal, C x (q) is the periodic feature of the touch action extracted, and q is the independent variable in the cepstrum.

3. The screen intelligent key switch system according to claim 1, characterized in that: The data processing unit also executes a heterogeneous multimodal fusion autoencoding algorithm. Its feature extraction expression is: i =E i (x i ); where x i is the sensor data of the i-th modality, E i is the encoder of this mode, z i is the potential representation vector; the attention fusion mechanism is expressed as: A ij =softmax(Q i ·K j ^T / √d), z fusion =∑ i ∑ j A ij ·V j ; where Q i =W Q ·z i is the query vector, K j =W k ·z j is the key vector, V j =W V ·z j is the value vector, d is the key vector dimension, A ij is the attention weight, z fusion is the fusion feature vector; Among them, softmax is the attention weight, d is the dimension of the key vector, z fusion is the fusion feature vector, z i is the potential representation vector, z j is the available modal feature, W Q 、W k 、W V is the learnable weight matrix.

4. The screen intelligent key switch system according to claim 1, characterized in that: The communication resource allocation strategy of the communication interface module is: BW allocation =f(System complexity ,Message urgency ,Env Quality ); System complexity For system complexity evaluation, Message urgency is the communication urgency, Env Quality For environmental quality assessment; the near-field communication unit is based on low-power Bluetooth technology and is used to transmit control instructions and basic status data; the long-range communication unit is based on Wi-Fi or 4G / 5G technology and is used to transmit system logs and complex interaction data.

5. The screen intelligent key switch system according to claim 1, characterized in that: The system also includes a quantum excitation zero-delay touch prediction module, which constructs a touch state representation model based on quantum computing principles: |ψ(t)〉=∑α i (t)∣φ i 〉; where |ψ(t)〉 represents the system state at time t, α i (t) is the complex amplitude, reaching ∑ i |α i (t)|^2=1,|φ i 〉 is the base state, representing different touch states; the state evolution equation is: |ψ(t+Δt)〉=U(t,Δt)|ψ(t)〉, where U(t,Δt)=exp(-iH(t)Δt) is the evolution operator, H(t)=H user (t)+H env (t)+H int (t) is the system Hamiltonian, which enables advance prediction of touch behavior; Among them, H user (t) is the user behavior Hamiltonian, H env (t) is the ambient Hamiltonian.

6. The screen intelligent key switch system according to claim 1, characterized in that: The control execution unit executes the prediction-verification dual-path processing strategy, and the objective function is: D final (t)=w predict ·D predict (t-Δt)+w Verify ·D Verify (t)+w context ·D context (t); Among them D predict (t-Δt) is the predicted path result, D Verify (t) is the real-time verification path result, D context (t) is the environmental context information, w predict 、w Verify 、w context is the dynamic weight coefficient, through the function w predict =f p (SNR,H(S),Stability predict ), w Verify =f V (SNR,H(S),Confidence Verify ), w context =f c (Reliability context ,Δt) calculation; the priority and intensity of control execution are dynamically adjusted according to the fusion decision results; Among them, SNR represents the current signal-to-noise ratio, Stability predict represents the predicted stability, and H(S) is the system state entropy.

Images