Multi-physical quantity collaborative perception brain synapse device and dynamic environment self-adaptive method thereof

Abstract

This invention discloses a multi-physical quantity collaborative sensing neuromorphic synaptic device and its dynamic environment adaptive method, relating to the fields of neuromorphic synaptic devices and multi-physical quantity sensing technology. It includes: a multi-physical quantity sensing layer, comprising a temperature-sensitive unit, a humidity-sensitive unit, and a gas-sensitive unit; a memristor synaptic modulation layer; a collaborative control unit; a read / write control module; and an interface and integration module. All of the above structures are integrated onto a wearable carrier, achieving hardware-level collaboration between multi-physical quantity sensing and memristor synaptic weight control. This invention achieves accurate identification of interference types and differentiated anti-interference processing through dynamically updated interference source features combined with a lightweight machine learning classifier, overcoming the shortcomings of existing technologies such as fixed filtering thresholds, lack of dynamic interference feature libraries, and difficulty in coping with various interferences in complex environments. Simultaneously, by dynamically adjusting the feature matching threshold and the minimum threshold, the probability of clutter misjudgment is reduced.

Description

Technical Field

[0001] This invention relates to the field of neuromorphic synaptic devices and multi-physical quantity sensing technology, and in particular to neuromorphic synaptic devices for multi-physical quantity collaborative sensing and their dynamic environment adaptive methods. Background Technology

[0002] With the rapid development of wearable electronic devices and neuromorphic computing technology, the demand for real-time sensing and precise control of multiple physical quantities (such as temperature, humidity, and gas concentration) is becoming increasingly urgent. Especially in wearable application scenarios such as health monitoring and environmental sensing, devices not only need to have high sensitivity, low power consumption, and dynamic environment adaptability, and be able to simulate the weighted control characteristics of neuromorphic synapses to achieve synergistic optimization of sensing signals and control logic, but also need to have efficient interference identification and anti-interference capabilities, as well as closed-loop iterative optimization capabilities of control logic, in order to cope with various interferences in complex dynamic environments and ensure the accuracy of sensing and control.

[0003] Currently, most existing multi-physical quantity sensing devices adopt a single physical quantity sensing unit design, and there is a lack of coordination and linkage between the sensing units. This makes it impossible to achieve synchronous acquisition and coupled analysis of multi-physical quantity signals, and the acquisition sensitivity is fixed, making it difficult to adapt to dynamically changing environmental scenarios.

[0004] In terms of dynamic environment adaptive control, the filtering of multi-physical quantity signals often adopts a fixed threshold method, which cannot dynamically adjust the threshold parameters according to the signal distribution characteristics and changes in environmental interference. This easily leads to problems such as misjudgment of clutter and loss of effective signals. Furthermore, there is no dynamically updated interference source feature library, and there is a lack of linkage with machine learning classifiers, making it impossible to achieve accurate identification of interference types and differentiated anti-interference processing, and making it difficult to cope with various interferences in complex environments. Moreover, the control of memristor synaptic weights often adopts fixed logic, which lacks adaptive response to changes in environmental physical quantities. It cannot dynamically adjust the control strategy according to the characteristics of the sensed signal, and there is no closed-loop optimization mechanism after control. It is impossible to optimize the control logic in reverse based on the control effect, making it difficult to ensure the stability of control accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-physical quantity collaborative sensing brain-like synaptic device and its dynamic environment adaptive method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-physical quantity collaborative sensing neuromorphic synaptic device, comprising:

[0007] The multi-physical quantity sensing layer includes a temperature-sensitive unit, a humidity-sensitive unit, and a gas-sensitive unit, which are used to collect temperature, humidity, and gas concentration signals in the environment;

[0008] The memristor synaptic modulation layer employs a reversible conductance-modulated memristor structure to simulate synaptic weights.

[0009] The collaborative control unit is electrically connected to the multi-physical quantity sensing layer and the memristor synapse modulation layer, respectively, and is used to generate weight adjustment instructions based on the signals collected by the multi-physical quantity sensing layer and the current weight state of the memristor synapse modulation layer.

[0010] The read / write control module is electrically connected to the collaborative control unit and the memristor synapse modulation layer, and is used to adjust the weight of the memristor synapse modulation layer according to the weight adjustment command;

[0011] An interface and integration module is electrically connected to the collaborative control unit and the read / write control module, and is used to communicate with external systems.

[0012] All of the above structures are integrated into a wearable carrier, enabling hardware-level collaboration between multi-physical quantity sensing and memristor synaptic weight control.

[0013] The dynamic environment adaptation method based on multi-physical quantity collaborative sensing specifically includes the following steps:

[0014] S1: Through the temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer, the temperature, humidity, and gas concentration signals in the environment are collected in real time.

[0015] S2: The multi-physical quantity sensing layer performs preliminary filtering on the collected multi-channel physical quantity signals, removes obvious noise, and then transmits them to the collaborative control unit;

[0016] S3: The read / write control module reads the current weight state of the memristor synaptic modulation layer in real time and transmits the weight state to the collaborative control unit;

[0017] S4: The collaborative control unit extracts and analyzes the features of the received physical quantity signals, combines them with the weight state of the memristor synaptic modulation layer, generates targeted weight adjustment instructions, and transmits them to the read / write control module.

[0018] S5: The read / write control module dynamically adjusts the pulse parameters according to the control instructions, and adaptively controls the weight of the memristor synapse modulation layer to achieve precise matching between physical quantity signals and synaptic weights;

[0019] S6: The collaborative control unit detects the control effect in real time, compares the memristor weight state and the accuracy of the sensing signal before and after control, dynamically optimizes the control logic, and forms a closed-loop adaptive control.

[0020] Preferably, the physical quantity signal in S1 specifically includes:

[0021] The temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer adopt a partitioned acquisition configuration. The inner unit is attached to the human skin to collect human-related physical quantities, while the outer unit is exposed to the environment to collect environmental physical quantities.

[0022] The acquisition sensitivity of each unit is dynamically adjusted based on the changing characteristics of environmental physical quantities.

[0023] After ensuring the integrity of the collected multiple physical quantity signals, they are transmitted to the subsequent processing stage.

[0024] Preferably, the preliminary filtering and transmission in S2 specifically includes:

[0025] Differentiated filtering methods are used to remove high-frequency interference for different physical quantity signals, while the original acquired signal and the filtered signal are transmitted to the collaborative control unit simultaneously.

[0026] The collaborative control unit acquires the original signal before filtering through a sliding window, extracts the signal statistical features and noise floor, dynamically calculates the clutter removal threshold based on the signal distribution characteristics, and the threshold parameter is adaptively adjusted.

[0027] A joint threshold judgment mechanism is established by combining the coupling characteristics of multiple physical quantities. A lightweight machine learning classifier is introduced to identify the signal mutation attributes and dynamically adjust the threshold parameters to distinguish between environmental changes and interference.

[0028] Record signal misjudgment cases, optimize threshold calculation parameters through closed-loop feedback mechanism, adapt to device aging and environmental changes, and support periodic threshold recalibration;

[0029] Threshold updates are implemented using an event-driven pattern.

[0030] Abnormal noise is removed based on dynamically adjusted thresholds, while valid physical quantity signals are retained.

[0031] The filtered multi-channel physical quantity signals are synchronized and aligned in time to ensure consistent transmission timing before being transmitted to the collaborative control unit.

[0032] Preferably, the current weight state control and transmission in S3 specifically includes:

[0033] The read / write control module sends probe pulses to the memristor synapse modulation layer via pulse detection to obtain the memristor conductance response signal;

[0034] Based on the memristor conductance response signal, the current memristor synaptic weight value is calculated using a preset algorithm, the weight stability range is defined, and the weight drift is recorded synchronously.

[0035] The current weight value, weight stability range, and drift amount are packaged and transmitted to the collaborative control unit via a preset bus.

[0036] Preferably, the generation of targeted weight adjustment instructions in S4 specifically includes:

[0037] The collaborative control unit performs feature extraction on the received multi-channel physical quantity signals, extracting three core features: the rate of change of the signal, the fluctuation amplitude, and the steady-state duration.

[0038] Based on the extracted signal characteristics, the influence of each physical quantity on the memristor synaptic weight is determined, and the control priority of each physical quantity is dynamically allocated in combination with the application requirements of wearable scenarios.

[0039] Based on the memristor weight status transmitted by the read / write control module, determine whether the current weight deviates from the adaptation range, and generate a targeted weight adjustment command including pulse amplitude, pulse width, and adjustment step size;

[0040] The control commands are encoded and transmitted to the read / write control module.

[0041] Preferably, the S5 weights are adaptively adjusted, specifically including:

[0042] The read / write control module parses the received control commands, extracts the core control parameters, and determines the control mode;

[0043] The output pulse parameters are dynamically adjusted according to the control mode. The degree of weight deviation is determined by two methods: quantization or dynamic threshold. Depending on the degree of deviation, small step fine-tuning or large step rapid control are used respectively.

[0044] The adjusted pulse signal is output to the memristor synaptic modulation layer, and the weight changes are monitored in real time to ensure the accuracy of regulation;

[0045] During the control process, the correspondence between pulse parameters and weight changes is recorded simultaneously to provide data support for subsequent control optimization.

[0046] Preferably, the control logic in S6 specifically includes:

[0047] The collaborative control unit receives the memristor weight status after adjustment from the read / write control module, compares it with the weight status before adjustment, calculates the weight adjustment deviation, and determines whether the adjustment effect meets the standard.

[0048] Synchronously collect the physical quantity sensing signals after regulation, compare the signal accuracy before regulation, and analyze the impact of interference factors on the regulation effect.

[0049] If the control effect is not satisfactory, adjust the control pulse parameters and control logic according to the weighted deviation and signal accuracy changes;

[0050] The optimized control logic is stored to form a closed-loop iteration mechanism, ensuring that the subsequent control effect continues to improve.

[0051] Preferably, the acquisition process in S1 further includes adaptive adjustment of the sampling timing, specifically including:

[0052] The collaborative control unit detects the rate of change of each physical quantity in real time and classifies it into slow rate and fast rate.

[0053] Based on the rate of change of physical quantities, the sampling timing of the sensing layer is dynamically adjusted, with the sampling interval extended for slow rates and shortened for fast rates.

[0054] The coordinated control unit synchronizes the timing of the read / write control module to ensure that the sampling and control of the sensing layer are synchronized, thus avoiding signal misalignment.

[0055] Preferably, the preliminary filtering process in S2 further includes:

[0056] Based on the characteristics of each unit in the multi-physical quantity sensing layer, a pre-defined interference source feature library is established and historical interference samples are stored in the collaborative control unit.

[0057] The collaborative control unit extracts the features of the signal before filtering, calculates its similarity with historical interference samples in the feature library, and dynamically adjusts the feature matching threshold.

[0058] If the similarity between the signal and all historical interference samples is lower than a preset minimum threshold, it is determined to be a suspected new type of interference, and its characteristics are automatically marked and recorded to complete the automatic discovery of new interference; the preset minimum threshold can be dynamically adjusted to reduce the probability of false positives.

[0059] The collaborative control unit shares signal features and similarity results with a lightweight machine learning classifier, reuses its classification results to verify the type of interference, and executes differentiated anti-interference strategies based on the type of interference source to eliminate abnormal signals.

[0060] After completing the anti-interference processing, subsequent filtering and signal transmission are performed to ensure signal purity. At the same time, newly discovered interference features are updated regularly to optimize the matching threshold. The update cycle can be adapted to the scenario as needed.

[0061] The technical effects and advantages of this invention are as follows:

[0062] (1) This invention achieves accurate identification of interference types and differentiated anti-interference processing by dynamically updating interference source features and combining them with a lightweight machine learning classifier. This solves the defects of existing technologies, such as fixed filtering thresholds, lack of dynamic interference feature library, and difficulty in dealing with various interferences in complex environments. At the same time, by dynamically adjusting the feature matching threshold and the minimum threshold, the probability of misjudgment of clutter is reduced, abnormal signals are effectively eliminated, and the integrity and purity of multi-physical quantity sensing signals are ensured. This provides reliable data support for subsequent weight control and is suitable for the complex and ever-changing interference environment of wearable scenarios.

[0063] (2) This invention constructs a closed-loop adaptive control mechanism of control-detection-optimization. The collaborative control unit can detect the weight control effect in real time, compare the weight state before and after control with the accuracy of the sensing signal, and optimize the control logic, adjust the threshold parameters and control priority according to the deviation. This solves the problems of fixed weight control logic, no closed-loop optimization and unstable control accuracy in the prior art. At the same time, the control accuracy is further improved by adapting the weight deviation degree through the hierarchical control strategy, realizing the accurate matching of physical quantity signals and memristor synaptic weights, and ensuring the control reliability of the device in dynamic environment.

[0064] (3) This invention solves the defects of independent, low integration and poor adaptability of multi-physical quantity sensing units in the prior art by configuring the multi-physical quantity sensing layer partition and dynamically adjusting the acquisition sensitivity and timing, combined with the hardware-level collaborative design of the device; the device is integrated into the wearable carrier, taking into account both high sensitivity and low power consumption, and can adaptively adjust the acquisition, filtering and control strategies according to the changes in environmental physical quantities, so as to realize the collaborative optimization of multi-physical quantity sensing and brain-like synaptic weight control, and meet the core requirements of real-time perception and precise control in wearable scenarios. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a schematic diagram of the structure of the multi-physical quantity collaborative sensing brain-like synaptic device of the present invention;

[0067] Figure 2 This is a schematic block diagram of the adaptive method of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] This invention provides, for example Figures 1-2 The multi-physical quantity collaborative sensing neuromorphic synaptic device shown includes:

[0070] The multi-physical quantity sensing layer includes a temperature-sensitive unit, a humidity-sensitive unit, and a gas-sensitive unit, which are used to collect temperature, humidity, and gas concentration signals in the environment;

[0071] The memristor synaptic modulation layer employs a reversible conductance-modulated memristor structure to simulate synaptic weights.

[0072] The collaborative control unit is electrically connected to the multi-physical quantity sensing layer and the memristor synapse modulation layer, respectively, and is used to generate weight adjustment instructions based on the signals collected by the multi-physical quantity sensing layer and the current weight state of the memristor synapse modulation layer.

[0073] The read / write control module is electrically connected to the collaborative control unit and the memristor synapse modulation layer, and is used to adjust the weights of the memristor synapse modulation layer according to the weight adjustment instructions;

[0074] The interface and integration module are electrically connected to the collaborative control unit and the read / write control module for communication with external systems.

[0075] All of the above structures are integrated into the wearable carrier, achieving hardware-level coordination of multi-physical quantity sensing and memristor synaptic weighting. The gas-sensitive unit is located on the outer unit, while the temperature-sensitive unit and humidity-sensitive unit are integrated into the inner unit.

[0076] The dynamic environment adaptation method based on multi-physical quantity collaborative sensing specifically includes the following steps:

[0077] S1: Through the temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer, the temperature, humidity, and gas concentration signals in the environment are collected in real time.

[0078] S2: The multi-physical quantity sensing layer performs preliminary filtering on the collected multi-channel physical quantity signals, removes obvious noise, and then transmits them to the collaborative control unit;

[0079] S3: The read / write control module reads the current weight state of the memristor synaptic modulation layer in real time and transmits the weight state to the collaborative control unit;

[0080] S4: The collaborative control unit extracts and analyzes the features of the received physical quantity signals, combines them with the weight state of the memristor synaptic modulation layer, generates targeted weight adjustment instructions, and transmits them to the read / write control module.

[0081] S5: The read / write control module dynamically adjusts the pulse parameters according to the control instructions, and adaptively controls the weight of the memristor synaptic modulation layer to achieve precise matching between physical quantity signals and synaptic weights.

[0082] S6: The collaborative control unit detects the control effect in real time, compares the memristor weight state and the accuracy of the sensing signal before and after control, dynamically optimizes the control logic, and forms a closed-loop adaptive control.

[0083] Specifically, the physical quantity signals in S1 include:

[0084] The temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer adopt a partitioned acquisition configuration. The inner unit is attached to the human skin to collect human-related physical quantities, while the outer unit is exposed to the environment to collect environmental physical quantities, and multiple differential physical quantity signals are collected simultaneously.

[0085] Based on the changing characteristics of environmental physical quantities, the acquisition sensitivity of each unit is dynamically adjusted. The sensitivity of the temperature-sensitive unit is adapted to the temperature change range, the sensitivity of the humidity-sensitive unit is adapted to the humidity fluctuation range, and the sensitivity of the gas-sensitive unit is adapted to the gas concentration threshold.

[0086] After ensuring the integrity of the collected multiple physical quantity signals, they are transmitted to the subsequent processing stage.

[0087] Specifically, the preliminary filtering and transmission in S2 include:

[0088] The multi-physical quantity sensing layer has a built-in hardware-level simple RC filter circuit, which adopts different filtering methods for different physical quantity signals: low-pass filtering is used for temperature-sensitive signals, band-pass filtering is used for humidity-sensitive signals, and high-pass filtering is used for gas-sensitive signals. This is mainly used to remove high-frequency glitches and interference from the signals. At the same time, the original acquisition signal is taken out from the input of the filter circuit and transmitted to the co-control unit together with the filtered signal.

[0089] A sliding window is constructed in the collaborative control unit to receive the original signal before filtering transmitted from the multi-physical quantity sensing layer in real time, calculate the mean, variance, and peak value statistics of the signal within the window, and extract the noise floor; based on the distribution characteristics of the signal within the window, the appropriate clutter removal threshold is dynamically calculated. The threshold is calculated using the mean ± k × standard deviation, where the value of k is adaptively adjusted according to the signal distribution characteristics.

[0090] By combining the coupling relationship between multiple physical quantity signals, a joint threshold judgment mechanism is established: when temperature and humidity change abruptly at the same time, it is judged as a real environmental change, and the clutter removal threshold is relaxed; when a single physical quantity signal changes abruptly while other physical quantity signals remain stable, the threshold sensitivity is increased; a lightweight machine learning classifier is introduced, which uses decision trees, support vector machines or lightweight neural networks to identify whether the signal change is due to environmental change or interference, and then dynamically adjusts the threshold parameters.

[0091] It records historical cases of valid signals and clutter misjudgments, and optimizes the threshold calculation model using a closed-loop feedback mechanism. If, during subsequent control processes, it is found that a segment of signal that was removed is valid information that should have been retained, the threshold calculation parameters are adjusted in reverse according to the proportion of signal misjudgment deviation. Specifically, the larger the misjudgment deviation, the larger the adjustment of the k value. At the same time, the sliding window size can be adjusted synchronously to adapt to different signal fluctuation scenarios. It supports periodic threshold recalibration to accommodate changes in signal characteristics caused by aging adapters or long-term environmental changes.

[0092] The threshold update is implemented using an event-driven model, which only triggers the threshold update when the signal fluctuation exceeds the preset range or at a timed interval, thus avoiding increased power consumption caused by frequent calculations.

[0093] Based on the dynamically adjusted clutter removal threshold, abnormal signals exceeding the threshold are identified as obvious clutter and removed by the filtering circuit, retaining the effective physical quantity signals.

[0094] The filtered multi-channel physical quantity signals are synchronized and aligned to ensure that the transmission timing of each signal is consistent. Then, they are transmitted to the collaborative control unit through a preset bus to avoid signal misalignment.

[0095] Specifically, the original signal is directly output from the input of the RC filter circuit through the lead wire, forming two parallel signals with the filtered signal at the output of the filter circuit, and is synchronously transmitted to the collaborative control unit. The collaborative control unit performs algorithm-level threshold analysis and clutter identification on the original signal, performs validity verification on the filtered signal, and finally selects to directly use the filtered signal (if the filtered signal is normal) or to perform further targeted filtering on the original signal (if the filtered signal still has residual clutter) based on the dynamic clutter removal threshold determination result.

[0096] The window length is selected based on the fluctuation characteristics of the physical quantity signal. The window length for temperature-sensitive signals is set to 100ms, and the window length for humidity-sensitive and gas-sensitive signals is set to 1s. The selection is based on adapting to the fluctuation period of each physical quantity signal to ensure the accuracy of statistical calculation. The mean is calculated by the arithmetic mean of all signal sampling points within the window, the variance is calculated by the sum of the squares of the deviations of the sampling points within the window from the mean, and the peak value is obtained by comparing the amplitude of all sampling points within the window in real time. The k value is adaptively adjusted according to the normality of the signal distribution. The closer the signal distribution is to the normal distribution, the k value is 1~2. When the signal fluctuation is chaotic and deviates from the normal distribution, the k value is 2~3.

[0097] The training data comes from pre-collected typical environmental change samples (such as temperature / humidity gradient changes, stable fluctuations in gas concentration) and typical interference samples (such as electromagnetic interference, sweat interference); the input features of the classifier include core features such as signal change rate, waveform steepness, and correlation of multiple physical quantity signals; the classifier output is the category judgment of whether the signal change belongs to environmental change or interference, and outputs the corresponding judgment probability value; the classifier adopts a lookup table or lightweight inference engine deployment method in the collaborative control unit to reduce hardware resource consumption and ensure real-time response.

[0098] The effectiveness of the signal is verified by the subsequent weight adjustment effect. If a segment of the signal that was removed corresponds to a real environmental event (i.e., the weight state after adjustment does not match the environmental change corresponding to the signal), then the signal is determined to be valid information that should be retained. The adjustment range of the k value adopts a quantitative rule. For every 10% increase in the misjudgment deviation, the k value increases by 0.2, and for every 10% decrease in the misjudgment deviation, the k value decreases by 0.1. The sliding window size is adjusted according to the signal fluctuation period. When the signal fluctuation period becomes longer, the window length is increased by 20%, and when the signal fluctuation period becomes shorter, the window length is reduced by 20% to ensure that it is adapted to the signal fluctuation characteristics.

[0099] The threshold for signal fluctuations exceeding the preset range is set to the historical average ±2σ (σ is the historical signal variance). When the signal amplitude exceeds this range, the threshold is updated immediately. The timed trigger cycle is set to every 5 minutes as a supplement to the event-driven trigger. The switching logic between the two modes is as follows: the event-driven mode is used first. When the signal fluctuation is stable (no event-driven trigger for 3 consecutive minutes), it automatically switches to the timed trigger mode. When the signal fluctuation exceeds the preset range again, it immediately switches back to the event-driven mode.

[0100] Specifically, the current weight state control and transmission in S3 includes:

[0101] The read / write control module uses a pulse detection method to send a detection pulse of preset amplitude to the memristor synapse modulation layer to obtain the memristor conductance response signal;

[0102] Based on the memristor conductance response signal, the current memristor synaptic weight value is calculated using the linear mapping formula W=k×G+b, where W is the memristor synaptic weight value, G is the amplitude of the memristor conductance response signal, k is the proportional coefficient, and b is the offset, all of which are factory preset parameters. Simultaneously, the stable range of the weight is defined, using one of two methods: Method 1 is the historical statistically stable weight value ±3σ (σ is the standard deviation of the historical weight value), and Method 2 is the factory-calibrated weight value ±5%. The weight drift is recorded synchronously; the drift is the difference between the current weight value and the median value of the stable range.

[0103] The current weight value, weight stability range, and drift amount are packaged and transmitted to the cooperative control unit via the SPI / I²C bus to ensure the accuracy and real-time nature of the transmitted data.

[0104] Specifically, during the initial operation phase, factory calibration ±5% is used to ensure the accuracy of the initial control benchmark. After long-term operation, it switches to historical statistics ±3σ to account for the characteristic changes of the adapter device after operation. The switching condition is that the device runs continuously for 30 days, automatically switching from the factory calibration method to the historical statistics method. An example of drift quantification is: current weight = 0.75, midpoint of stable range = 0.70, drift = +0.05. When the drift exceeds the preset threshold, a calibration request is automatically triggered, and the drift and calibration request are simultaneously reported to the collaborative control unit along with the weight data to facilitate subsequent control logic optimization.

[0105] Specifically, S4 generates targeted weight adjustment instructions, including:

[0106] The collaborative control unit performs feature extraction on the received multi-channel physical quantity signals, extracting three core features: the rate of change of the signal, the fluctuation amplitude, and the steady-state duration.

[0107] Based on the extracted signal characteristics, the influence of each physical quantity on the memristor synaptic weight is determined, and the control priority of each physical quantity is dynamically allocated in combination with the application requirements of wearable scenarios.

[0108] Based on the memristor weight status transmitted by the read / write control module, determine whether the current weight deviates from the adaptation range, and generate a targeted weight adjustment command including pulse amplitude, pulse width, and adjustment step size;

[0109] The control commands are encoded and transmitted to the read / write control module to ensure the stability and readability of the command transmission.

[0110] Among them, the S5 weights are adaptively adjusted, specifically including:

[0111] The read / write control module parses the received control commands, extracts core parameters such as pulse amplitude, pulse width, and control step size, and determines the control mode.

[0112] According to the control mode, the parameters of the output pulse are dynamically adjusted. The judgment criteria for small / large weight deviation are adopted by one of the following two methods: Method 1 is quantitative judgment, that is, the current weight deviation from the stable range is <5% as small deviation, and ≥5% as large deviation; Method 2 is dynamic threshold judgment, that is, the deviation judgment threshold is adaptively adjusted according to the historical control effect; small step fine adjustment is adopted for the case of small weight deviation, and large step fast control is adopted for the case of large weight deviation.

[0113] The adjusted pulse signal is output to the memristor synapse modulation layer to monitor the change process of memristor weights in real time and ensure the accuracy of weight control.

[0114] During the control process, the correspondence between pulse parameters and weight changes is recorded simultaneously to provide data support for subsequent control logic optimization.

[0115] The control logic in S6 specifically includes:

[0116] The collaborative control unit receives the memristor weight status after adjustment from the read / write control module, compares it with the weight status before adjustment, calculates the weight adjustment deviation, and determines whether the adjustment effect meets the standard.

[0117] Synchronously collect the physical quantity sensing signals after regulation, compare the signal accuracy before regulation, and analyze the impact of interference factors on the regulation effect.

[0118] If the control effect is not up to standard, the control pulse parameters are optimized and the control logic is adjusted according to the weighted adjustment deviation and signal accuracy changes. Specific adjustments include: priority allocation rule adjustment, prioritizing the control priority of physical quantities that have the greatest impact on signal accuracy. If the gas concentration signal accuracy deviation is the largest, the priority of gas-sensitive control is increased, while the priority of temperature / humidity-sensitive control, which has a smaller impact on accuracy, is decreased; threshold judgment standard adjustment, simultaneously adjusting the weight deviation judgment threshold (e.g., if the original deviation ≥5% is considered large, it can be adjusted to ≥4% or ≥6% depending on the misjudgment situation) and the k value of the clutter removal threshold, optimizing the sliding window size to adapt to the current signal fluctuation characteristics;

[0119] The optimized control logic is stored to form a closed-loop iteration mechanism, ensuring that the subsequent control effect continues to improve.

[0120] Specifically, the implementation logic of each adjustment in the closed-loop optimization includes:

[0121] The initial control priority allocation was temperature = 0.3, humidity = 0.3, and gas = 0.4. When a 10% decrease in gas signal accuracy was detected, it was determined that the gas signal had the greatest impact on the control effect, and the priority was adjusted to temperature = 0.2, humidity = 0.2, and gas = 0.6, prioritizing the control accuracy corresponding to the gas signal. When the decrease in gas signal accuracy resulted in the control effect not meeting the target, various thresholds were adjusted synchronously. The weight deviation judgment threshold was adjusted from 5% to 4% to improve the sensitivity of weight deviation recognition; at the same time, the k value of the clutter removal threshold was adjusted from 2.5 to 2.0 to improve clutter recognition sensitivity, reduce the impact of interference on signal accuracy, and achieve threshold linkage optimization. The initial sliding window length is 100ms. When the detected signal fluctuation period lengthens to 500ms, the sliding window is adjusted to 500ms to adapt to the signal fluctuation characteristics. The collaborative logic between window adjustment and threshold adjustment is as follows: when the window expands, the clutter removal threshold k is appropriately reduced to improve the accuracy of signal statistics calculation and the adaptability of threshold determination; when the window shrinks, the k value is appropriately increased to avoid misjudging valid signals.

[0122] The acquisition process in S1 also includes adaptive adjustment of the sampling timing, specifically including:

[0123] The collaborative control unit detects the rate of change of each physical quantity in real time and classifies it into slow rate and fast rate. The threshold for classifying the slow / fast rate of temperature change can be adaptively adjusted according to seasonal differences, regional environmental characteristics and historical environmental data. The basic classification standard is: a temperature change of ≤0.5℃ per second is a slow rate, and a temperature change of >0.5℃ per second is a fast rate. This threshold can be dynamically fine-tuned according to the actual scenario.

[0124] Based on the rate of change of the physical quantity, the sampling timing of the sensing layer is dynamically adjusted: when the rate of change is slow, the sampling interval is extended to 2-3 s / time; when the rate of change is fast, the sampling interval is shortened to 0.1-0.5 s / time.

[0125] The coordinated control unit synchronizes the timing of the read / write control module to ensure that the sampling of the sensing layer and the control of the read / write module are synchronized, thus avoiding signal misalignment.

[0126] The preliminary filtering process in S2 also includes:

[0127] Based on the material response characteristics of each unit in the multi-physical quantity sensing layer, a feature library of electromagnetic interference, human sweat interference, and external stray gas interference is preset, and historical interference samples are stored in the collaborative control unit.

[0128] The collaborative control unit extracts the frequency and amplitude fluctuation patterns of the signal before filtering, as well as the linkage characteristics of the signals of each unit. It then calculates the similarity between the signal and historical interference samples in the feature library. The similarity calculation adopts one of the following typical methods: Euclidean distance, cosine similarity, or correlation coefficient, which can be adaptively selected according to the signal feature type. At the same time, the feature matching threshold is dynamically adjusted. The higher the similarity, the lower the matching threshold, thereby improving the accuracy of interference classification.

[0129] If the similarity between the extracted signal features and all historical interference samples in the feature library is lower than the preset minimum threshold, it is determined to be a suspected new interference type. The signal features are automatically marked and recorded to complete the automatic discovery of the new interference type. The preset minimum threshold can be a fixed value or dynamically adjusted according to the historical false alarm rate. The higher the false alarm rate, the lower threshold should be appropriately increased to reduce the probability of false judgment.

[0130] The collaborative control unit shares the extracted signal features and similarity calculation results with a lightweight machine learning classifier, reuses the classification results to further verify the type of interference, and then executes differentiated anti-interference strategies according to the type of interference source (including newly discovered interference): when there is electromagnetic interference, the built-in filter circuit is activated to filter high-frequency interference; when there is human sweat interference, the humidity-sensitive signal is baseline-calibrated; when there is stray gas interference and newly discovered interference, the corresponding physical quantity signal is threshold-screened to remove abnormal signals.

[0131] After completing the anti-interference processing, subsequent filtering and signal transmission steps are performed to ensure signal purity. At the same time, the characteristics of newly discovered interference are regularly updated to the feature library to optimize the matching threshold. The regular update cycle can be daily, weekly, or event-driven. When the cumulative number of newly discovered interference types reaches a preset number, an immediate update is triggered.

[0132] Specifically, interference originates from learning and its interaction with machine learning classifiers, including:

[0133] Electromagnetic interference is characterized by 50Hz and its harmonic components, with a signal abrupt change time of <1ms; human sweat interference is characterized by slow drift of the humidity signal, with a drift rate of <0.5%RH / s; stray gas interference is characterized by sudden spikes in the gas signal, with a duration of <2s. All types of features are stored as standard feature vectors for easy similarity comparison. When the signal is mainly characterized by frequency domain features, cosine similarity is used; when the signal is mainly characterized by time domain waveforms, Euclidean distance is used; when the signal is mainly characterized by the correlation of multiple physical quantities, the correlation coefficient is used. The selection is adaptively based on the signal type to improve comparison accuracy. The initial preset minimum threshold is set to 0.7; if the false alarm rate of interference within one week is >5%, the threshold is increased to 0.75 to reduce the probability of false judgment; if the false alarm rate within one week is <1%, the threshold is decreased to 0.65 to avoid missing new interference types. The feature extraction module simultaneously sends the extracted signal feature vectors to both the interference classifier and the lightweight machine learning classifier. If the interference type classification results output by both are consistent, the corresponding differentiated anti-interference strategy is directly executed. If the classification results are inconsistent, arbitration logic is initiated, and the classifier with the higher confidence score is used to ensure accurate interference type determination. When the similarity between the extracted signal features and all historical samples in the feature library is less than the preset minimum threshold, it is automatically marked as a suspected new interference. The complete feature vector, occurrence time, and duration of the signal are recorded simultaneously. If the same feature vector appears ≥3 times within a certain period, it is confirmed as a new interference type and officially added to the interference feature library. At the same time, matching threshold optimization is triggered, and the cluster centers of each feature vector are recalculated to improve the accuracy of subsequent interference identification. The interference feature library can be updated regularly in three modes: automatic update during the low-power period at midnight every day; fixed update every Sunday; and instant update triggered when the cumulative number of newly discovered interference types reaches 5, ensuring that the feature library can adapt to new interference scenarios in a timely manner.

[0134] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A brain-like synaptic device for multi-physical quantity collaborative sensing, characterized in that, include: The multi-physical quantity sensing layer includes a temperature-sensitive unit, a humidity-sensitive unit, and a gas-sensitive unit, which are used to collect temperature, humidity, and gas concentration signals in the environment; The memristor synaptic modulation layer employs a reversible conductance-modulated memristor structure to simulate synaptic weights. The collaborative control unit is electrically connected to the multi-physical quantity sensing layer and the memristor synapse modulation layer, respectively, and is used to generate weight adjustment instructions based on the signals collected by the multi-physical quantity sensing layer and the current weight state of the memristor synapse modulation layer. The read / write control module is electrically connected to the collaborative control unit and the memristor synapse modulation layer, and is used to adaptively adjust the weight of the memristor synapse modulation layer according to the weight adjustment command; An interface and integration module is electrically connected to the collaborative control unit and the read / write control module, and is used to communicate with external systems. All of the above structures are integrated into the wearable carrier to achieve hardware-level collaboration between multi-physical quantity sensing and memristor synaptic weight control; The collaborative control unit also includes: The collaborative control unit performs feature extraction on the received multi-channel physical quantity signals, extracting three core features: the rate of change of the signal, the fluctuation amplitude, and the steady-state duration. Based on the extracted signal characteristics, the influence of each physical quantity on the memristor synaptic weight is determined, and the control priority of each physical quantity is dynamically allocated in combination with the application requirements of wearable scenarios. Based on the memristor weight status transmitted by the read / write control module, determine whether the current weight deviates from the adaptation range, and generate a targeted weight adjustment command including pulse amplitude, pulse width, and adjustment step size; The read / write control module also includes: The read / write control module parses the received control commands, extracts the core control parameters, and determines the control mode; The output pulse parameters are dynamically adjusted according to the control mode. The degree of weight deviation is determined by two methods: quantization or dynamic threshold. Depending on the degree of deviation, small step fine-tuning or large step rapid control are used respectively. The adjusted pulse signal is output to the memristor synaptic modulation layer, and the weight changes are monitored in real time to ensure the accuracy of regulation; During the control process, the correspondence between pulse parameters and weight changes is recorded simultaneously to provide data support for subsequent control optimization.

2. A dynamic environment adaptive method based on multi-physical quantity collaborative sensing, characterized in that, The adaptive method includes the multi-physical quantity collaborative sensing neuromorphic synaptic device as described in claim 1, and specifically includes the following steps: S1: Through the temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer, the temperature, humidity, and gas concentration signals in the environment are collected in real time. S2: The multi-physical quantity sensing layer performs preliminary filtering on the collected multi-channel physical quantity signals, removes obvious noise, and then transmits them to the collaborative control unit; S3: The read / write control module reads the current weight state of the memristor synaptic modulation layer in real time and transmits the weight state to the collaborative control unit; S4: The collaborative control unit extracts and analyzes the features of the received physical quantity signals, combines them with the weight state of the memristor synaptic modulation layer, generates targeted weight adjustment instructions, and transmits them to the read / write control module. S5: The read / write control module dynamically adjusts the pulse parameters according to the control instructions, and adaptively controls the weight of the memristor synapse modulation layer to achieve precise matching between physical quantity signals and synaptic weights; S6: The collaborative control unit detects the control effect in real time, compares the memristor weight state and the accuracy of the sensing signal before and after control, dynamically optimizes the control logic, and forms a closed-loop adaptive control.

3. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The physical quantity signals in S1 specifically include: The temperature-sensitive unit, humidity-sensitive unit, and gas-sensitive unit of the multi-physical quantity sensing layer adopt a partitioned acquisition configuration. The inner unit is attached to the human skin to collect human-related physical quantities, while the outer unit is exposed to the environment to collect environmental physical quantities. The acquisition sensitivity of each unit is dynamically adjusted based on the changing characteristics of environmental physical quantities. After ensuring the integrity of the collected multiple physical quantity signals, they are transmitted to the subsequent processing stage.

4. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The preliminary filtering and transmission in S2 specifically include: Differentiated filtering methods are used to remove high-frequency interference for different physical quantity signals, while the original acquired signal and the filtered signal are transmitted to the collaborative control unit simultaneously. The collaborative control unit acquires the original signal before filtering through a sliding window, extracts the signal statistical features and noise floor, dynamically calculates the clutter removal threshold based on the signal distribution characteristics, and the threshold parameter is adaptively adjusted. A joint threshold judgment mechanism is established by combining the coupling characteristics of multiple physical quantities. A lightweight machine learning classifier is introduced to identify the signal mutation attributes and dynamically adjust the threshold parameters to distinguish between environmental changes and interference. Record signal misjudgment cases, optimize threshold calculation parameters through closed-loop feedback mechanism, adapt to device aging and environmental changes, and support periodic threshold recalibration; Threshold updates are implemented using an event-driven pattern. Abnormal noise is removed based on dynamically adjusted thresholds, while valid physical quantity signals are retained. The filtered multi-channel physical quantity signals are synchronized and aligned in time to ensure consistent transmission timing before being transmitted to the collaborative control unit.

5. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The current weight state control and transmission in S3 specifically includes: The read / write control module sends probe pulses to the memristor synapse modulation layer via pulse detection to obtain the memristor conductance response signal; Based on the memristor conductance response signal, the current memristor synaptic weight value is calculated using a preset algorithm, the weight stability range is defined, and the weight drift is recorded synchronously. The current weight value, weight stability range, and drift amount are packaged and transmitted to the collaborative control unit via a preset bus.

6. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The control logic in S6 specifically includes: The collaborative control unit receives the memristor weight status after adjustment from the read / write control module, compares it with the weight status before adjustment, calculates the weight adjustment deviation, and determines whether the adjustment effect meets the standard. Synchronously collect the physical quantity sensing signals after regulation, compare the signal accuracy before regulation, and analyze the impact of interference factors on the regulation effect. If the control effect is not satisfactory, adjust the control pulse parameters and control logic according to the weighted deviation and signal accuracy changes; The optimized control logic is stored to form a closed-loop iteration mechanism, ensuring that the subsequent control effect continues to improve.

7. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The acquisition process in S1 also includes adaptive adjustment of the sampling timing, specifically including: The collaborative control unit detects the rate of change of each physical quantity in real time and classifies it into slow rate and fast rate. Based on the rate of change of physical quantities, the sampling timing of the sensing layer is dynamically adjusted, with the sampling interval extended for slow rates and shortened for fast rates. The coordinated control unit synchronizes the timing of the read / write control module to ensure that the sampling and control of the sensing layer are synchronized, thus avoiding signal misalignment.

8. The dynamic environment adaptive method for multi-physical quantity collaborative sensing according to claim 2, characterized in that, The preliminary filtering process in S2 also includes: Based on the characteristics of each unit in the multi-physical quantity sensing layer, a pre-defined interference source feature library is established and historical interference samples are stored in the collaborative control unit. The collaborative control unit extracts the features of the signal before filtering, calculates its similarity with historical interference samples in the feature library, and dynamically adjusts the feature matching threshold. If the similarity between the signal and all historical interference samples is lower than a preset minimum threshold, it is determined to be a suspected new type of interference, and its characteristics are automatically marked and recorded to complete the automatic discovery of new interference; the preset minimum threshold can be dynamically adjusted to reduce the probability of false positives. The collaborative control unit shares signal features and similarity results with a lightweight machine learning classifier, reuses its classification results to verify the type of interference, and executes differentiated anti-interference strategies based on the type of interference source to eliminate abnormal signals. After completing the anti-interference processing, subsequent filtering and signal transmission are performed to ensure signal purity. At the same time, newly discovered interference features are updated regularly, the matching threshold is optimized, and the update cycle is adapted to the scenario as needed.

Images