Button switch adaptive optimization method and storage medium

Abstract

The invention discloses a button switch adaptive optimization method and a storage medium. The method comprises the following steps: acquiring a tactile signal through a high-density piezoelectric sensor, and extracting pressure distribution, gradient change rate and energy concentration degree characteristics in combination with a space-time convolution kernel and a dynamic baseline calibration algorithm; integrating time-frequency domain analysis of the historical operation sequence and neural differential equation modeling, predicting an operation probability and marking an anomaly; and an anti-interference control instruction is generated by fusing environmental parameters, and multi-source data are dynamically weighted by adopting a gating attention mechanism. The system constructs a multi-stage reliability protection mechanism, and fault immunity is realized through feature drift detection, digital twin residual verification and hybrid control switching. And responding to a real-time demand, dynamically optimizing heterogeneous computing resource allocation, executing three-dimensional storage reconstruction and precision bit width switching, and forming a closed-loop cognitive evolution link.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a method for adaptively optimizing a push-button switch and a storage medium. Background Art

[0002] Traditional push-button switches face significant technical challenges in complex industrial environments. In the prior art, the response delay of the button control system is relatively high, and it is difficult to meet the real-time requirements especially when dealing with high-frequency operations; the logic complexity of the indicator lights is high, resulting in a large cognitive load for users and difficult maintenance; the environmental adaptability is insufficient, and temperature fluctuations, mechanical vibrations, and electromagnetic interference are likely to cause misoperations.

[0003] In addition, existing methods mostly rely on a single signal modality and lack the ability of multi-source data fusion, making it difficult to accurately capture the user's operation intention. In terms of fault handling, traditional systems usually adopt fixed-threshold detection and cannot dynamically adapt to feature drift, resulting in low fault recovery efficiency. Although some studies have tried to introduce machine learning algorithms, they have not effectively combined the optimization of heterogeneous computing resources with the dynamic reconstruction of 3D storage, making it difficult to balance real-time performance and energy consumption efficiency. Summary of the Invention

[0004] (I) Technical Problems to be Solved To solve the above problems, the present invention proposes a method for adaptively optimizing a push-button switch and a storage medium, aiming to solve the problem that the prior art relies on a single signal modality and lacks the ability of multi-source data fusion, making it difficult to accurately capture the user's operation intention.

[0005] (II) Technical Solutions A method for adaptively optimizing a push-button switch according to the present invention, comprises: acquiring a tactile signal of the push-button switch, and obtaining a spatio-temporal feature vector based on the tactile signal, where the spatio-temporal feature vector includes pressure distribution, pressure gradient change rate, and energy aggregation degree; acquiring a historical operation event sequence of the push-button switch, and obtaining a timing prediction result based on the spatio-temporal feature vector and the operation event sequence, where the timing prediction result includes an operation probability distribution and an anomaly mark; acquiring an environmental state parameter of the push-button switch, and generating a control instruction based on the environmental state parameter, the timing prediction result, and the spatio-temporal feature vector, for converting the abstract decision of the neural network into an executable control instruction; constructing a multi-level reliability protection mechanism based on the execution state of the control instruction and the physical device response delay, and realizing fault immunity through feature drift detection, physical-digital state residual analysis, and hybrid control loop switching; In response to the feature drift detection result and the fault immunity requirement, dynamically optimize the heterogeneous computing resource allocation strategy, perform three-dimensional storage space reconstruction and dynamic precision bit-width switching based on real-time task priorities and energy consumption constraints, and form a closed-loop cognitive evolution link.

[0006] In the present invention, the method for extracting the spatio-temporal feature vector of the tactile signal includes: Construct a three-dimensional pressure field for collecting a pressure distribution matrix, and apply a spatio-temporal convolution kernel to smooth the pressure distribution matrix. Adopt a dynamic baseline calibration algorithm to eliminate environmental temperature drift. The dynamic baseline algorithm includes a smoothing factor, and the smoothing factor is adaptively adjusted according to the operating frequency.

[0007] In the present invention, the spatio-temporal convolution kernel smoothing the pressure distribution matrix further includes: Normalize the pressure gradient change rate to generate a standardized gradient vector. Calculate the energy aggregation degree of the pressure distribution matrix.

[0008] In the present invention, the method for generating the timing prediction result includes: Perform multi-scale wavelet scattering transform on the historical operation event sequence, extract the time-frequency domain feature tensor, and use a neural differential equation to model the implicit dynamic process. Predict the future operation intention probability distribution through a bidirectional gated recurrent unit, and mark it as an abnormal operation when the prediction residual exceeds the statistical threshold.

[0009] In the present invention, the environmental state parameters include temperature, main mechanical vibration frequency, electromagnetic interference intensity, and power supply ripple coefficient. The method for fusing the environmental state parameters and the adaptive control instruction is: Construct a preset environmental condition template library containing legal state transition paths for various typical environmental parameter combinations, and dynamically weight the influence weight of environmental parameters on the control strategy through an attention mechanism to generate an anti-interference optimization instruction. When the electromagnetic interference intensity exceeds the degradation threshold, forcibly enable the degradation mode and limit high-frequency state switching operations.

[0010] In the present invention, the multi-level reliability protection mechanism includes: Online monitor feature drift, detect abnormal inputs through Mahalanobis distance, and trigger an alarm when the Mahalanobis distance value is greater than 5.99. Monitor the ReLU activation rate of the neural network hidden layer, and freeze the network parameters when the ReLU activation rate exceeds the freezing threshold. Compare the sensor data with the digital twin prediction value, and switch to the standby PID controller when the residual exceeds three times the standard deviation.

[0011] In the present invention, the heterogeneous computing resource allocation strategy includes: Construct a heterogeneous computing resource partitioning mechanism to decouple the heterogeneous task flow into four parallel execution units of sensing signal parsing, decision logic generation, security constraint verification, and system interaction coordination according to functional modalities; Design a dynamic numerical representation mode switching strategy to adaptively adjust the data precision topology according to the criticality level of real-time tasks, and realize the dynamic migration between the low-entropy computing state and the high-fidelity parsing state.

[0012] Another computer-readable storage medium of the present invention stores a computer program thereon, and when the program is executed by a processor, it implements the button switch adaptive optimization method described in any one of the above technical solutions.

[0013] (III) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: In the present invention, through the spatio-temporal convolution kernel and the dynamic baseline calibration algorithm, the delay of state switching is greatly reduced and the response speed is improved. By adopting the combined entropy compression and the adaptive optical language system, the logic complexity of the indicator light is reduced.

[0014] In the present invention, through the multi-level protection mechanism and the hybrid control loop switching, the fault recovery time is greatly shortened. The dynamic precision bit-width switching and the three-dimensional storage reconstruction technology can greatly improve the inference and meet the strict power consumption constraints of industrial scenarios. Description of the drawings

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0016] Figure 1 It is a schematic flow structure diagram of the adaptive optimization method; Figure 2 It is a pressure thermal diagram of the button switch; Figure 3 It is a schematic flow structure diagram of the closed-loop cognitive evolution link. Specific implementation manners

[0017] This disclosure constructs a tactile intelligent agent that realizes the autonomous evolution of an industrial interaction interface through the deep meshing of spatio-temporal perception fusion and dynamic decision-making topology. The system kernel forms the construction and analysis of a multi-dimensional tactile field, elevates the spatio-temporal gradient of the pressure matrix to a continuous manifold of operation semantics, and reconstructs the topological space of human-machine dialogue at the differential geometry level. The fractal decision-making architecture converts mechanical pulses into a control language with semantic density through micro-meso-macro cognitive transitions, forming an embodied cognitive closed loop.

[0018] Embodiment 1

[0019] Such as Figures 1 - 3 shown, a method for self-adaptive optimization of a button switch includes the following steps: S100. Obtain the tactile signal of the button switch, and obtain a spatio-temporal feature vector based on the tactile signal. The spatio-temporal feature vector includes pressure distribution, pressure gradient change rate, and energy aggregation degree.

[0020] In the tactile signal acquisition stage, a high-density piezoelectric sensing array is used to construct a three-dimensional dynamic pressure field. The original signal flows through a baseline drift compensation module to construct a sliding window energy entropy model: , where represents the energy entropy value within the window, which is used to measure the degree of chaos of the pressure distribution, is the normalized probability of the th pressure value within the window is the th pressure sampling value within the window, is the sum of all pressure values within the window, is the number of samples of the sliding window. The sliding window energy entropy model dynamically detects and compensates for baseline drift by quantifying the degree of chaos of the pressure distribution, and solves the distortion problem of tactile signals caused by environmental interference or hardware aging. It is the core preprocessing module to ensure the accuracy of the spatio-temporal feature vector.

[0021] When the window entropy value exceeds the threshold, dynamic compensation is triggered. The threshold is 2.5, that is, when , dynamic compensation is triggered, and the compensation amount is generated by the weighted fusion of the historical baseline sliding mean and the current window minimum value: , where is the historical baseline sliding mean, calculated by exponential moving average (EMA), , , reflecting the retention degree of historical data, is the current baseline estimate value. The role of the historical baseline sliding mean is to capture the long-term baseline trend of the sensor signal and suppress low-frequency drift, such as signal offset caused by temperature change. is within the time window , the tactile signal The minimum value is used to detect signal instantaneous undershoot or abnormal low voltage in real time, reflecting the current environmental interference, such as sudden mechanical vibration.

[0022] Noise suppression adopts a combined filtering strategy - applying the anisotropic diffusion equation in the spatial dimension: , where represents the spatio-temporal pressure field function, describing the pressing force on the button surface at position , time . In the expression, the spatio-temporal pressure field function is input as the original tactile signal to be processed, containing noise and effective pressure distribution information; represents the spatial gradient operator of the pressure field, is used to quantify the local change rate of the pressure field, for identifying the edges and texture features of the pressing area; represents the magnitude of the gradient, used to measure the severity of the local pressure change, and a high magnitude value corresponds to the edge or mutation area; is the diffusion coefficient, , is the gradient threshold parameter, and in this disclosure . The gradient threshold parameter is mainly used to dynamically adjust the diffusion intensity. When the magnitude of the gradient is much larger than the gradient threshold parameter, strong diffusion is allowed to suppress noise; when the magnitude of the gradient is much smaller than the gradient threshold parameter, diffusion is suppressed to retain details. is used to represent the divergence operator acting on the diffusion flux , used to describe the anisotropic diffusion process of the pressure field, and adaptively smooths along the gradient direction.

[0023] In the time dimension, an improved Kalman filter is adopted, and the process noise covariance matrix is adaptively adjusted according to the pressure change rate, , where represents the process noise covariance matrix at the th moment in the Kalman filter, used to characterize the uncertainty of the system model, and the larger the value, the lower the confidence in the predicted value. In tactile signal processing, dynamically adjusts the balance between the response speed and stability of the filter. In the above expression, 0.1 represents the basic noise covariance, used to ensure the minimum noise tolerance under static or slowly changing pressure, and avoid signal distortion caused by over-sensitive filtering; 0.05 represents the pressure change rate weight coefficient, used to adjust the contribution degree of the pressure change rate to the noise covariance, and control the sensitivity of the dynamic response; represents the amplitude of the pressure change rate over time.

[0024] Specifically, the pressure distribution field is extracted through multi-scale morphological analysis: constructing a structural element , representing pixel disks with diameters of 3, 5, and 6 respectively, perform opening operations and closing operations , calculate the morphological gradient: , where represents the position at the morphological gradient value, represents taking the supremum, i.e., the maximum value, within the neighborhood defined by the structuring element , represents taking the infimum, i.e., the minimum value, within the neighborhood defined by the structuring element . respectively represent the values of the original signal at the position .

[0025] The final feature vector consists of three parts: centroid localization, shape description, and complexity evaluation. Among them, centroid localization calculates the centroid coordinates of the pressure field through zero-order and first-order moments to identify the core pressing area; shape description quantifies the overall shape of the pressure distribution, such as symmetry and extensibility, by extracting the first three-order features of Hu invariant moments; while complexity evaluation characterizes the randomness degree of the tactile signal by calculating the normalized entropy value of the pressure distribution.

[0026] The pressure gradient change rate is obtained through spatio-temporal joint analysis. In the spatial dimension, the Sobel operator is used to detect the horizontal and vertical gradients of the pressure field, and in the time dimension, the five-point central difference method is applied to calculate the instantaneous change rate. The system sorts the gradient data by significant values, extracts the top 5% of the high-gradient regions to generate histogram features, and statistically analyzes the spatial distribution frequency in 20 intervals to effectively capture the spatio-temporal characteristics of sudden operation behaviors, such as rapid consecutive presses.

[0027] The energy aggregation degree analysis fuses time-frequency domain information. In the spatial dimension, calculate the energy integral of each sensing unit within a 50ms window, and extract the aggregation direction of the energy field through principal component analysis; in the time dimension, use the short-time Fourier transform to monitor the center frequency of the energy spectrum. When a high-frequency energy event is detected, the system automatically triggers the transient response mode, such as increasing the filtering intensity in a strong mechanical vibration environment. In this embodiment, high-frequency energy means when the energy frequency is greater than 150Hz.

[0028] The spatio-temporal feature fusion adopts an attention-guided mechanism. The original pressure, gradient, and energy data are processed by three-branch encoders respectively, and the cross-attention is used to dynamically allocate feature weights. For example, when local sensing fails due to oil stains, the system automatically reduces the attention weight of the abnormal area and instead enhances the contribution of spatio-temporal gradient features. The fused feature vector is optimized through contrastive learning, forcing the features of the same type of operations to closely aggregate in the vector space and significantly separating the features of different types of operations.

[0029] ​​​S200. Obtain the historical operation event sequence of the button switch, and obtain the time series prediction result based on the spatio-temporal feature vector and the operation event sequence. The time series prediction result includes the operation probability distribution and the anomaly flag.

[0030] In the processing of the historical operation event sequence, the operation features at different time granularities are captured through a multi-scale sliding window mechanism. The short-term window focuses on the transient pressure change, and calculates the maximum value and variance of the pressure gradient within the window; the medium-term window counts the operation frequency and the pressure distribution entropy to quantify the operation randomness; The long-term window extracts the first 5 order components of the Fourier coefficients to identify the periodic pattern. The three-level features are concatenated into a 17-dimensional time series tensor , where is the number of time steps, forming a hierarchical time series representation.

[0031] The fusion of the spatio-temporal features and the operation sequence in the above formula has too much content and too little text content. Check the necessity of the formula and add text content. The fusion of spatio-temporal features and operation sequences adopts a dynamic attention mechanism. The spatio-temporal feature vector generates a query vector through a linear transformation, and the temporal features of the operation sequence are formed into key-value pairs through convolutional coding. By calculating the attention weights, the historical operation segments strongly related to the current spatio-temporal state are screened out. For example, when a specific pressure distribution pattern is detected, the model will focus on the operation interval rules under past similar patterns. The fused features are input into a bidirectional gated recurrent network, and its hidden state predicts the future operation probability and the anomaly risk simultaneously.

[0032] The operation probability distribution is dynamically generated through a mixture of Gaussian models. The model outputs the parameters (mean, variance, weight) of three Gaussian distributions, corresponding to three typical modes of rapid response, normal operation, and delayed trigger respectively. For example, if the current pressure is concentrated in the edge area and changes smoothly, the model will increase the weight of the delayed trigger mode and predict that the next operation interval may be extended to more than 800 ms. In real-time detection, if the actual interval continuously deviates from the predicted value by more than three standard deviations, a primary anomaly flag is triggered. To distinguish between accidental fluctuations and real faults, a secondary verification is introduced: after three consecutive primary anomalies, calculate the Mahalanobis distance of the operation features within the window. If it exceeds the statistical threshold, it is confirmed as the final anomaly.

[0033] Anomaly detection combines gradient analysis and time localization. Calculate the contribution of the hidden layer of the neural network to the anomaly flag through backpropagation, and locate the key time period that causes the anomaly. For example, a certain anomaly flag may be jointly caused by three abnormal pressure peaks within the past 2 seconds. When the system marks an anomaly, it generates an event report containing the time stamp and the influencing factors to assist the maintenance personnel in quickly locating the root cause of the problem.

[0034] Progressive multi-task learning is adopted for model training. The main task minimizes the operation interval prediction error, and the auxiliary task optimizes the anomaly classification accuracy. In the initial stage of training, normal operation data is mainly used, and 5% to 30% of abnormal samples are gradually introduced to enhance the model's sensitivity to rare events. To prevent overfitting, a temporal smoothing constraint term is added to force the hidden state changes at adjacent times to be gentle. In the deployment stage, a dynamic parameter update strategy is adopted. After every 1000 operations are accumulated, the model is fine-tuned using the latest data: the core weights are kept unchanged, and only the parameters of the output layer are updated. At the same time, an anomaly pattern library is maintained, and when a new type of anomaly is detected, incremental learning of the model is automatically triggered.

[0035] S300. Obtain the environmental state parameters of the button switch, and generate a control instruction based on the environmental state parameters, the temporal prediction result, and the spatio-temporal feature vector, which is used to convert the abstract decision of the neural network into an executable control instruction.

[0036] In the environmental state parameter acquisition stage, a three-dimensional perception network integrating temperature, humidity, and electromagnetic interference is integrated: the temperature sensor uses an NTC thermistor, and through a second-order polynomial calibration equation Convert the voltage signal into an accurate temperature value, where is the calibrated temperature value, representing the actual temperature calculated from the sensor voltage, is the original voltage input of the sensor, which is the unprocessed electrical signal output by the thermistor, is the quadratic term coefficient, representing the curvature of the voltage change with temperature, is the linear term coefficient, representing the linear proportional relationship between voltage and temperature, is the constant term, representing the temperature offset when the voltage is zero, such as baseline compensation, etc.; humidity measurement uses a capacitive sensing unit, and its output is compensated by adaptive baseline to eliminate the influence of temperature drift; the electromagnetic interference intensity is captured by a broadband detection circuit, and the energy integral value in the frequency band of 10 kHz - 1 GHz is extracted , is the electromagnetic interference intensity, representing the total electromagnetic energy in the specified frequency band, respectively represent the lower and upper limits of the integration frequency, defining the monitoring frequency band of electromagnetic interference, is the frequency domain representation of the electromagnetic interference signal, obtained by performing a Fourier transform (FFT) on the time domain signal; the energy integral value is converted into a standardized signal with a dynamic range of 0 - 5V through logarithmic compression. The fusion of environmental parameters and spatio-temporal feature vectors adopts a gated attention mechanism: , represents the comprehensive index of environmental stress, which is generated by fusing temperature and humidity generated by fusion, while are the normalized temperature and humidity values; is the time series prediction state of the LSTM network, representing the time evolution pattern of button operations; denotes learnable weight matrices, used to generate queries and keys respectively; is the query vector, generated based on spatio-temporal features, is the key vector, generated by concatenating environmental indicators and the LSTM state; denotes the attention weight matrix, representing the correlation strength between spatio-temporal features and environmental / temporal features, where represents calculating the similarity between the query and the key, is the probability distribution after scaling and normalized by Softmax.

[0037] is the comprehensive index of environmental stress, the fused features are input to the control policy generator, which consists of two parallel fully connected layers: the main network outputs the original control quantity , where is the weight matrix of the main network, used to map the fused feature to the control quantity space, is the fused feature vector, containing spatio-temporal features , environmental parameters and the time series prediction state 's comprehensive information, and the fused feature vector encodes multi-dimensional information of the current system state, such as button pressure distribution, environmental temperature and humidity, historical operation mode, etc.; is the bias term, used to adjust the baseline output of the control quantity; is the activation function, which is the Sigmoid function in this embodiment, used to compress the linear transformation result to the range. The original control quantity performs a linear transformation on the fused feature through the weight matrix , superimposes the bias , and then generates the normalized original control quantity through the activation function . For example, when the environmental temperature rises and the pressure gradient changes violently, the weight of the corresponding dimension in increases, resulting in the output of tending to 1 and triggering a fast state switching instruction; when in a confidence scenario (such as severe electromagnetic interference). The safety constraint network reduces the weight of through and preferentially adopts

[0038] to ensure system safety. The safety constraint network generates the confidence , is the weight matrix of the safety constraint network, which is used to map the fused feature vector to the linear transformation parameters in the confidence space. It is obtained through training and learning, and represents the influence weights of different feature dimensions on safety; is the bias term; the confidence is the dynamic safety gate, which realizes the balance between intelligent decision-making and robustness. The final control instruction is obtained through dynamic weighted fusion: , is the finally generated control instruction, which is used to directly drive the button switch or indicator light. Its main function is to synthesize the output of the main network decision and safety constraints, and balance the real-time response and system safety; is the preset safety instruction, which is generated based on historical data or expert rules. This expression adjusts the weights of the main network and safety policy in real time. For example, when the environmental pressure rises (such as a sudden increase in temperature), decreases, the proportion of increases to avoid the risk of overheating; when the electromagnetic interference decreases, increases to make full use of the optimization ability of the neural network.

[0039] Preset safety instruction is generated according to the historical operation mode. For example, in a high-temperature environment, the switch trigger sensitivity is forced to be reduced. The instruction discretization adopts an adaptive threshold mechanism: define the state switching critical value , where is the mean value of the last 100 , is the standard deviation. When , a high-level instruction is generated, otherwise is maintained. For the electromagnetic interference scenario, an instruction verification link is added: each control instruction is appended with a 3-bit parity check code, and the receiving end corrects single-bit errors through a majority voting mechanism. When the verification fails, a timing backtracking is triggered and the valid instructions of the previous two cycles are re-executed. In extreme working conditions, such as , the degradation mode is started: the neural network inference engine is turned off and switched to a conservative strategy based on a moving average filter to ensure the minimum functional availability of the system.

[0040] S400. Based on the execution status of the control instruction and the response delay of the physical device, a multi-level reliability protection mechanism is constructed to achieve fault immunity through feature drift detection, physical-digital state residual analysis, and hybrid control loop switching.

[0041] In the feature drift detection stage, the system constructs a dynamic protection mechanism by monitoring the execution status of control instructions in real time, such as response latency, execution success rate, and resource occupancy rate. Based on the sliding window statistical method, the system continuously calculates the deviation between the current state and the historical normal mode. When the state indicators, such as a sudden increase in latency or a sharp drop in success rate, deviate from the long-term statistical mean by more than three standard deviations, a security response is immediately triggered: freezing the parameter update of the online learning module, rolling back to the stable weight snapshot of the previous 5 minutes, and initiating a deep diagnosis process. This mechanism can effectively address feature drift problems caused by environmental mutations or model degradation. For example, in the control of industrial robotic arms, when a sudden load change leads to abnormal latency, the system can complete state freezing within 2 milliseconds to prevent the issuance of incorrect instructions.

[0042] Physical-digital state residual analysis is achieved through the two-way verification between the digital twin and the real device. The system has a built-in high-precision device dynamics model that generates predicted values of key parameters such as motor speed and temperature in real time and compares them with the physical sensor data at the millisecond level. The residual evaluation adopts a dynamic threshold strategy: comprehensively considering the historical fluctuation pattern and the instantaneous change rate, when the residual amplitude continuously exceeds the limit and is accompanied by a sharp increase in high-frequency noise energy, such as when the energy in the 2-4 kHz frequency band exceeds the baseline by 10 dB, it is determined as a hardware anomaly.

[0043] The hybrid control loop switching adopts a progressive migration strategy to ensure control continuity in fault scenarios. Initially, the neural network controller and the preset PID controller output instructions in parallel. By comparing the differences, such as when the output torque deviation exceeds 10% of the rated value, it is determined whether a switch is needed. During the transition period, exponential decay weighted fusion is used to gradually reduce the weight ratio of the neural network output. At the same time, Lyapunov stability constraints are injected to prevent sudden changes in the rotational speed or position signal. In the stable period, it is completely switched to PID control, and offline retraining of the neural network is initiated. This design performs excellently in the injection molding machine pressure control test. When the material fluidity suddenly changes and causes control instability, the system quickly completes a seamless switch, and the control accuracy of the output fluctuation amplitude is greatly improved.

[0044] The multi-level protection mechanism enhances the system's resilience through hierarchical collaboration. Feature drift detection serves as the first line of defense to address soft faults; residual analysis focuses on hardware anomaly diagnosis; and hybrid control provides the final guarantee. The three are linked through an event bus and a priority scheduler to build a closed-loop protection system from anomaly perception to decision execution.

[0045] S500. In response to the feature drift detection result and the fault immunity requirement, dynamically optimize the heterogeneous computing resource allocation strategy, perform three-dimensional storage space reconstruction and dynamic precision bit-width switching based on real-time task priorities and energy consumption constraints, and form a closed-loop cognitive evolution link.

[0046] In the feature drift detection phase, the system adopts a real-time monitoring mechanism that fuses multiple metrics. By comparing the probability differences between the real-time data stream and the historical baseline distribution, such as using KL divergence to measure the change in distribution shape, and the statistical deviation in the multi-dimensional feature space, such as using Mahalanobis distance to evaluate the overall distribution shift, a dynamic drift evaluation model is constructed. When moderate drift is detected, for example, when the metric exceeds 2 standard deviations of the historical fluctuation range, an adaptive sampling strategy is automatically triggered - the data acquisition frequency is increased from the normal 100Hz to 200Hz, and the incremental learning mode is activated to quickly fine-tune the parameters of the last layer of the model with a high learning rate of 0.01. If the drift continues to intensify, for example, if the metric does not fall back for 3 consecutive cycles, an emergency response protocol is initiated: load the baseline model snapshot stored in the safe area, and at the same time aggregate the latest feature patterns from the edge nodes through the federated learning framework to achieve online reconstruction of the model structure.

[0047] The dynamic scheduling of heterogeneous computing resources adopts a "time-space dual-driven" optimization strategy. In the time dimension, a dynamic evaluation model of task priority is designed, comprehensively considering the task urgency, processing deadline, and resource occupancy rate. For example, fault diagnosis tasks are given the highest priority, the closer to the deadline, the higher the weight, and non-critical tasks are downgraded when the GPU video memory occupancy exceeds 80%. In the space dimension, based on the hardware topology awareness technology, the working states of computing units such as CPUs, GPUs, and FPGAs are analyzed in real time, such as temperature, load, and energy consumption, and the system benefits of different task mapping schemes are predicted through graph neural networks. When it is detected that the FPGA temperature exceeds the safety threshold, the convolution operation task is immediately migrated to the idle GPU, and the matrix sparsification compression technology is enabled to greatly reduce the power consumption while ensuring the calculation accuracy.

[0048] As Figure 2 shown, the reconstructed three-dimensional storage architecture obtains the pressure heat map of the button switch through a pressure sensor, and the reconstruction of the three-dimensional storage architecture focuses on intelligent layering and resilience design. First, a data heat map analysis module is constructed. By tracking the access frequency, data correlation, and life cycle of storage units, high-frequency access parameters, such as the weights of the last layer of the neural network, are migrated to the near-compute core area (HBM2e) of the 3D stacked storage, greatly reducing the data access latency. Secondly, an energy consumption-aware storage model is established. Under the strict constraints of a 10ns access latency and a 5W peak power consumption, the dynamic voltage and frequency scaling technology (DVFS) is adopted, and the L3 cache voltage is reduced from 1.2V to 0.9V according to the real-time load. Finally, a cross-layer redundancy protection mechanism is deployed to implement multi-copy erasure code storage for key parameters - every 8 data blocks are extended with 4 parity blocks, allowing 3 storage units to be damaged simultaneously and still be able to fully recover the data, and at the same time ensuring data integrity through hash tree verification.

[0049] As Figure 3As shown in the figure, the dynamic precision regulation system adopts a closed-loop control chain of "perception - decision - execution". The perception layer monitors task characteristics, hardware status, and environmental constraints in real time. For example, image recognition tasks require a high dynamic range, an energy-saving mode is activated when the remaining battery power is less than 20%, and the frequency is forced to decrease in a high-temperature environment. The decision layer constructs a reinforcement learning model, transforming precision selection into a multi-objective optimization problem: dynamically switching between the FP32 full-precision mode and the INT8 quantization mode, and balancing the Pareto optimality of precision loss and energy consumption savings through a deep Q network. When a hardware anomaly is detected, the execution layer immediately switches to the FP32 mode and starts an error recovery process, ensuring output reliability through a recomputation mechanism.

[0050] The closed-loop cognitive evolution system achieves continuous optimization through a five-level linked adaptive mechanism. The perception network collects chip-level fine-grained metrics through multiple embedded probes, such as from transistor-level leakage current to system-level heat dissipation efficiency, and constructs a digital twin to map the physical system state in real time. The analysis engine uses a temporal graph convolutional network (T-GCN) to mine the spatio-temporal correlations of multi-dimensional metrics and predict performance inflection points within the next 5 seconds, such as the critical moment when the cache hit rate drops to 80%. The decision center uses a mixed integer programming algorithm to generate candidate configuration plans and selects the optimal strategy based on a multi-dimensional utility function, such as preferentially ensuring the resource supply of the fault recovery link. The execution unit converts control instructions into micro-operation codes through a hardware abstraction layer, ensuring the reallocation of computing resources, the reconstruction of storage topology, and the precision mode switching within an extremely short time.

[0051] Embodiment 2

[0052] An embodiment of the present invention provides a computer-readable storage medium.

[0053] The computer program stored on the computer-readable storage medium provided by the embodiment of the present invention can implement the steps of any one of the above button switch adaptive optimization methods when executed by a processor.

[0054] The computer-readable storage medium may include: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc that can store program codes.

[0055] For the introduction of the computer-readable storage medium provided by the embodiment of the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

[0056] The various embodiments in the specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0057] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0058] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of both. The software modules can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0059] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. An adaptive optimization method for a push-button switch, characterized in that: It includes: Obtain the tactile signal of the push-button switch, and obtain a spatio-temporal feature vector based on the tactile signal. The spatio-temporal feature vector includes pressure distribution, pressure gradient change rate, and energy concentration; Obtain the historical operation event sequence of the push-button switch, and obtain a time-series prediction result based on the spatio-temporal feature vector and the operation event sequence. The time-series prediction result includes operation probability distribution and anomaly marking; Obtain the environmental state parameters of the push-button switch, and generate a control instruction based on the environmental state parameters, the time-series prediction result, and the spatio-temporal feature vector, which is used to convert the abstract decision of the neural network into an executable control instruction; Based on the execution status of the control instruction and the physical device response delay, construct a multi-level reliability protection mechanism, and achieve fault immunity through feature drift detection, physical-digital state residual analysis, and hybrid control loop switching; In response to the feature drift detection result and the fault immunity requirement, dynamically optimize the heterogeneous computing resource allocation strategy, and perform three-dimensional storage space reconstruction and dynamic precision bit-width switching based on real-time task priority and energy consumption constraint to form a closed-loop cognitive evolution link.

2. The button switch adaptive optimization method according to claim 1, wherein The method for extracting the spatio-temporal feature vector of the tactile signal includes: Construct a three-dimensional pressure field for collecting the pressure distribution matrix, and apply a spatio-temporal convolution kernel to smooth the pressure distribution matrix; Use a dynamic baseline calibration algorithm to eliminate environmental temperature drift. The dynamic baseline algorithm includes a smoothing factor, and the smoothing factor is adaptively adjusted according to the operation frequency.

3. The button switch adaptive optimization method according to claim 2, wherein The spatio-temporal convolution kernel further smooths the pressure distribution matrix, including: Normalize the pressure gradient change rate to generate a standardized gradient vector; Calculate the energy concentration of the pressure distribution matrix.

4. The button switch adaptive optimization method according to claim 3, wherein The method for generating the time-series prediction result includes: Perform multi-scale wavelet scattering transform on the historical operation event sequence, extract the time-frequency domain feature tensor, and use a neural differential equation to model the implicit dynamic process; Predict the future operation intention probability distribution through a bidirectional gated recurrent unit, and mark it as an abnormal operation when the prediction residual exceeds the statistical threshold.

5. The button switch adaptive optimization method according to claim 1 or 4, characterized in that, The environmental state parameters include temperature, mechanical vibration main frequency, electromagnetic interference intensity, and power supply ripple coefficient. The fusion method of the environmental state parameters and the adaptive control instruction is: Construct a preset environmental condition template library, which contains legal state transition paths of various typical environmental parameter combinations, and dynamically weight the influence weight of environmental parameters on the control strategy through an attention mechanism to generate an anti-interference optimization instruction; When the electromagnetic interference intensity exceeds the degradation threshold, forcibly enable the degradation mode and limit high-frequency state switching operations.

6. The button switch adaptive optimization method according to claim 5, wherein The multi-level reliability protection mechanism includes: Online monitor feature drift, detect abnormal inputs through Mahalanobis distance, and trigger an alarm when the Mahalanobis distance value is greater than 5.99; Monitor the ReLU activation rate of the hidden layer of the neural network, and freeze the network parameters when the ReLU activation rate exceeds the freezing threshold; Compare the sensor data with the predicted value of the digital twin, and switch to the standby PID controller when the residual exceeds three times the standard deviation.

7. The button switch adaptive optimization method according to claim 6, wherein, The heterogeneous computing resource allocation strategy includes: Construct a heterogeneous computing resource partitioning mechanism to decouple the heterogeneous task flow into four parallel execution units of sensing signal parsing, decision logic generation, security constraint verification, and system interaction coordination according to the functional modality; Design a dynamic numerical representation mode switching strategy to adaptively adjust the data precision topology according to the criticality level of real-time tasks, and realize the dynamic migration between the low-entropy computing state and the high-fidelity parsing state.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the adaptive optimization method of the push-button switch described in any one of claims 1-7.

Images