Motion recognition method and system based on redox photoelectric memristor, terminal and storage medium

Abstract

The invention discloses a motion recognition method and system based on a redox photoelectric memristor, a terminal and a storage medium, and the method comprises the steps: selecting classical motions based on a human body motion data set, extracting time sequence data, and coding the time sequence data into an optical pulse sequence; constructing a reservoir array composed of a plurality of photoelectric memristors, and expanding an optical pulse sequence signal into a high-dimensional state vector; constructing a supervised training model, solving a weight matrix, and constructing a memristive cross array; and outputting an action classification result through simulation domain operation based on the output current multi-path light current signals in combination with the memristor cross array. According to the method, the motion features can be directly fed into a rear-end classification network for action recognition without depending on a complex digital feature extraction algorithm, so that the transmission and processing overhead of redundant data is fundamentally eliminated; a high-efficiency, low-delay and high-robustness hardware solution is provided for real-time and anti-noise motion recognition in scenes such as intelligent monitoring and man-machine interaction.

Description

Technical Field

[0001] This invention relates to the field of motion recognition technology, and in particular to a motion recognition method, system, terminal, and storage medium based on a redox photomemristor. Background Technology

[0002] Motion recognition is a key technology for realizing human-computer interaction, behavior understanding, and intelligent perception, and it is widely used in fields such as health monitoring, intelligent security, and virtual reality. Current mainstream technologies rely on computer vision, which can achieve non-contact, high-precision motion analysis, but its "perception-transmission-computation" separation architecture brings problems such as high energy consumption, high latency, and privacy leaks, which restricts its deployment in edge computing scenarios.

[0003] In recent years, memristor-based reservoir computing has attracted much attention due to its ability to handle spatiotemporal signals. This architecture only requires training the output layer, and the reservoir is composed of nonlinear dynamic nodes, offering advantages such as high training efficiency and fast response speed. However, existing dynamic reservoirs are mostly constructed based on filamentary memristors, which rely on the physical mechanisms of random ion migration and the formation of conductive filaments. This makes it difficult to precisely control the dynamic response, and it is also difficult to integrate photoresponse dynamics and electrically controlled nonvolatile storage in the same device, limiting the overall energy efficiency and integration of the system.

[0004] Therefore, existing technologies still have shortcomings. Summary of the Invention

[0005] To address the aforementioned deficiencies in the prior art, this invention provides a motion recognition system, method, terminal, and storage medium based on a redox photomemristor. The technical solution adopted by this invention is as follows: In a first aspect, the present invention provides a motion recognition method based on a redox photomemristor, the method comprising: Classic movements are selected based on a human motion dataset, time-series data is extracted based on the selected classic movements, and the time-series data is encoded into a light pulse sequence; A reservoir array consisting of several volatile photoinduced conductive states of redox photoresistors is constructed, and the optical pulse sequence signal is extended into a high-dimensional state vector based on the reservoir array. A supervised training model is constructed based on a high-dimensional state vector, the weight matrix is ​​solved, and a memristor cross array is constructed based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmable redox-type photomemristors with the same structure. Based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross-array, the mapping of high-dimensional features in the current multi-channel photocurrent signals to action categories is realized through analog domain operations, and the action classification results are output.

[0006] In one implementation, classic movements are selected based on a human motion dataset, time-series data is extracted based on the selected classic movements, and the time-series data is encoded into a light pulse sequence, including: Five classic movements were selected based on the human motion dataset: drinking water, throwing, sitting down, standing, and jumping. For each type of classic movement, the three-dimensional coordinate signals of 6 key skeletal nodes of the human body are extracted to form 18 channels of time-series data; The 18-channel timing data is encoded into a sequence of light pulses that can be directly sensed by a redox-type photomemristor that is in a volatile photoinduced conductive state.

[0007] In one implementation, a reservoir array is constructed, consisting of several volatile photo-induced conductive states of redox photoresistors. Based on the reservoir array, the optical pulse sequence signal is extended into a high-dimensional state vector, including: A reservoir array with parallel processing is composed of 18 volatile photo-induced conductive state redox photoresistors. Each redox photoresistor corresponds to a channel, and each redox photoresistor receives the optical pulse sequence signal of the corresponding channel individually. When each redox-type photomemristor in a volatile photoinduced conductive state receives a sequence of light pulses, it generates a photocurrent signal with short-term memory characteristics. During the current decay process of the redox-type photomemristor in each volatile photoinduced conductive state, the current signal is discretely sampled at fixed time intervals. Each sampling point is defined as a virtual node, and 40 virtual nodes are continuously sampled for each channel. The data from the virtual nodes are fused, and the 18-channel optical pulse sequence signals are expanded into a 720-dimensional high-dimensional state vector.

[0008] In one implementation, a supervised training model is constructed based on a high-dimensional state vector, and the weight matrix is ​​solved, including: A supervised training model is constructed by using the ridge regression algorithm, taking a 720-dimensional high-dimensional state vector as input and one-hot encoded labels of five classic actions as output. The weight matrix is ​​obtained by minimizing the loss function with L2 regularization, resulting in a weight matrix of 20 rows and 5 columns.

[0009] In one implementation, a memristor cross-connect array is constructed based on a weight matrix, including: A differential conductance representation scheme is adopted, which maps each weight value in the weight matrix to the conductance difference of a pair of memristor cells; Based on the size of the weight matrix, the difference pairs are determined, and multiple non-volatile electrically programmed photoresistors with the same structure are used to form a memristor cross array to carry the difference pairs. Each weight value in the weight matrix is ​​linearly mapped to a preset target conductance range, and a stepped-incrementing voltage pulse sequence is applied to each photomemristor. The conductivity state is read in real time, and the conductivity value is gradually brought closer to the target value through closed-loop feedback control. The weight matrix is ​​persistently stored in the memristor cross subarray in the form of non-volatile conductivity distribution.

[0010] In one implementation, based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross-connect array, analog domain operations are used to map high-dimensional features in the current multi-channel photocurrent signals to action categories, outputting action classification results, including: The current multi-channel photocurrent signal output by the reservoir array is converted into a voltage signal, and then processed by analog-to-digital conversion and controller to generate a voltage vector corresponding to the current input action frame; The voltage vector is applied to the corresponding row of the memristor cross array through the word line driving circuit, providing an input signal for analog domain operations; Perform analog domain operations and convert the output current signal into a voltage signal through a readout circuit. Each voltage signal corresponds to a classification score for a type of action. The action category with the highest score is identified by a comparison decision circuit, and the action classification result is output.

[0011] In one implementation, performing analog domain operations includes: The input voltage vector is distributed along the word line within the array, and each memristor cell generates a corresponding cell current based on the stored conductance value; According to Kirchhoff's current law, the total output current on each bit line is the sum of the currents of all cells in that column; By using the differential conductance weighting structure, the multiplication and accumulation operations of the input voltage vector and the differential conductance weighting matrix are automatically completed in the analog domain.

[0012] Secondly, embodiments of the present invention also provide a motion recognition system based on a redox photomemristor, wherein the system is used to implement the steps of the motion recognition method based on a redox photomemristor as described in any of the above solutions, and the system includes: The time-series data encoding module is used to select classic movements based on the human motion dataset, extract time-series data based on the selected classic movements, and encode the time-series data into a light pulse sequence. A high-dimensional state extension module is used to construct a reservoir array composed of several volatile photoinduced conductive states of redox photoresistors, and to extend the optical pulse sequence signal into a high-dimensional state vector based on the reservoir array. The memristor cross array construction module is used to build a supervised training model based on a high-dimensional state vector, solve the weight matrix, and construct a memristor cross array based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmable redox-type photomemristors with the same structure. The action recognition module is used to map the high-dimensional features of the current multi-channel photocurrent signals output by the reservoir array to action categories through analog domain operations, in conjunction with the memristor cross array, and output the action classification results.

[0013] Thirdly, embodiments of the present invention also provide a terminal, wherein the terminal includes a memory, a processor, and an entity extraction processing program stored in the memory and executable on the processor. When the processor executes the entity extraction processing program, it implements the steps of the motion recognition method based on a redox photomemristor in any of the above-described schemes.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores an entity extraction processing program, the entity extraction processing program implementing the steps of the motion recognition method based on a redox photomemristor as described in any of the above schemes on the computer-readable storage medium.

[0015] Beneficial Effects: Compared with existing technologies, this invention provides a motion recognition method based on redox photomemristors. First, classic actions are selected from a human motion dataset. Temporal data is extracted from these selected actions and encoded into a light pulse sequence. Then, a reservoir array is constructed, consisting of several redox photomemristors in volatile photoinduced conductive states. Based on this reservoir array, the light pulse sequence signal is extended into a high-dimensional state vector. Next, a supervised training model is constructed based on the high-dimensional state vector. The weight matrix is ​​solved, and a memristor cross array is constructed based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array consists of multiple structurally identical non-volatile photomemristors in non-electrically programmed states. Finally, based on the current multi-channel photocurrent signals output from the reservoir array, and combined with the memristor cross array, analog domain operations are used to map the high-dimensional features in the current multi-channel photocurrent signals to action categories, outputting the action classification result.

[0016] This invention directly processes analog optical signals using a photoelectric memristor array, eliminating the need for complex digital feature extraction algorithms. Motion features can be directly fed into a backend classification network for action recognition. Furthermore, by leveraging the controllable response characteristics of the same memristor under different stimuli, this invention integrates photosensitive dynamic feature extraction with electronically controlled non-volatile weight storage on a single device platform. This allows for end-to-end processing from signal perception to classification computation at the hardware level. This highly integrated architecture of sensing, storage, and computation fundamentally eliminates the overhead of redundant data transmission and processing, providing an efficient, low-latency, and robust hardware solution for real-time, noise-resistant motion recognition in scenarios such as intelligent monitoring and human-computer interaction. Attached Figure Description

[0017] Figure 1 A flowchart of a preferred embodiment of the motion recognition method based on a redox photomemristor provided in this invention.

[0018] Figure 2 This is a schematic diagram of the photoresistor device structure in the motion recognition method based on a redox photoresistor provided in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of the voltage retracement of the redox-type photomemristor in a dark condition, according to an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram of the photoresponse of a redox photomemristor according to an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram of the technical architecture of the motion recognition method based on a redox photomemristor provided in an embodiment of the present invention.

[0022] Figure 6 Using the action of drinking water as an example, this paper demonstrates the signal encoding of the X, Y, and Z axes and the device response waveforms of each channel.

[0023] Figure 7 The schematic diagram of an RC circuit based on a redox photomemristor core is provided for an embodiment of the present invention.

[0024] Figure 8 This is a schematic diagram illustrating the motion recognition rate of the motion recognition method based on a redox photomemristor provided in an embodiment of the present invention.

[0025] Figure 9 The schematic diagram of the motion recognition system based on redox photomemristor provided in the embodiments of the present invention.

[0026] Figure 10 A schematic diagram of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0028] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0029] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0030] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, "first control information" and "second control information" are only used to distinguish different control information and do not limit their order.

[0031] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0032] It should also be understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0033] To address the problems of existing technologies, this invention provides a motion recognition method based on redox-type photomemristors. This method can directly process analog light signals through a photomemristor array, eliminating the need for complex digital feature extraction algorithms, and directly feeds motion features into a backend classification network for action recognition. In specific applications, this embodiment first selects classic actions based on a human motion dataset, extracts time-series data from these actions, and encodes the time-series data into light pulse sequences. Then, a reservoir array composed of several redox-type photomemristors in volatile photoinduced conductive states is constructed. Based on this reservoir array, the light pulse sequence signal is extended into a high-dimensional state vector. Next, a supervised training model is constructed based on the high-dimensional state vector, the weight matrix is ​​solved, and a memristor cross array is constructed based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array of multiple structurally identical non-volatile, electrically programmed redox-type photomemristors. Finally, based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross-array, the mapping of high-dimensional features in the current multi-channel photocurrent signals to action categories is realized through analog domain operations, and the action classification results are output.

[0034] The motion recognition method based on redox photomemristors in this embodiment can be applied to terminals, including intelligent product terminals such as computers, specifically, as shown below. Figure 1 As shown in the figure, the motion recognition method based on redox photomemristor in this embodiment includes the following steps: Step S100: Select classic movements based on the human motion dataset, extract time-series data based on the selected classic movements, and encode the time-series data into a light pulse sequence.

[0035] In practical applications, the redox-type photomemristor in this embodiment is fabricated using the following steps: (1) Weigh 10 mg of polyoxometalate sample and add o-dichlorobenzene organic solvent to prepare a 5 mg / ml solution, which is called solution A. Weigh polystyrene polymer, add o-dichlorobenzene organic solvent, place on a stirring table and heat and stir at 60°C for 30 minutes to prepare a 50 mg / ml solution, which is called solution B. Mix 100 μL of solution A and 200 μL of solution B, and then add o-dichlorobenzene organic solvent to dilute it so that the concentration of polystyrene in the mixed solution is diluted to 20 mg / ml. Then place on a heating table and heat and stir at 60°C for 12 hours. After standing to room temperature, filter to obtain the thin film spin coating solution.

[0036] (2) Take a glass substrate with an ITO (indium tin oxide) electrode pattern etched on it (185 nm thick) and ultrasonically clean it with deionized water for 15 minutes to remove surface contaminants. Then dry the surface with high-purity nitrogen and dry it on a hot plate at 120°C for 120 minutes.

[0037] (3) Using a micropipette, 200 μL of the above-mentioned thin film spin-coating solution was drawn and dropped onto the center of the pretreated glass substrate to form a film. The spin-coating program was set as follows: the first stage was 500 rpm for 10 seconds to ensure the solution spread; the second stage was 1500 rpm for 60 seconds to form a uniform film. The sample was then placed on a hot stage and annealed at 60°C for 30 minutes to remove residual solvent and enhance the film density, finally obtaining a dielectric layer with a thickness of about 60 nm.

[0038] (4) Define the top electrode pattern using a metal mask. Place the mask over the spin-coated substrate and deposit gold electrodes under vacuum conditions using a thermal evaporation deposition apparatus. The vacuum level is controlled below 10⁻⁴ Pa, the thickness is 80 nm, and the deposition rate is controlled at 0.1 Å / s. After deposition, remove the mask to complete the fabrication of the redox-type photoresistor. The structure of the fabricated redox-type photoresistor is shown below. Figure 2 As shown, DDA-SmPW@PS is polyoxometalate-doped polystyrene.

[0039] This embodiment focuses on the performance study of a redox-type photoresistor. The performance study mainly focuses on two aspects: first, the electrically controlled non-volatility of the memristor under dark conditions; second, the optically controlled volatility of the memristor under 375 nm light pulse stimulation. Specific experimental content and results are as follows: (1) The photoelectric performance of the device was tested using a semiconductor parameter analyzer and a light control module. Under dark conditions, a DC voltage of 0V to 5V was applied for a sweep retracement, followed by a sweep retracement of 0V to -5V at a scan rate of 0.05V / s, to monitor its resistive switching behavior. Experimental results showed that the redox-type photomemristor exhibits non-volatile resistive switching characteristics under dark conditions, such as... Figure 3 As shown.

[0040] (2) A semiconductor parameter analyzer was used to apply a continuous reading voltage to the device to detect the current change under light pulse stimulation with a wavelength of 375 nm. Experimental results showed that with continuous light pulse stimulation, the response current of the redox photomemristor gradually accumulated and relaxed, exhibiting light-induced memristor characteristics. Moreover, the degree of current accumulation and relaxation varied under different light intensities, such as... Figure 4 As shown.

[0041] Facing the stringent requirements of real-time response and low power consumption in edge computing scenarios, overcoming the physical limitations of traditional architectures has become a key challenge in building next-generation efficient intelligent sensing systems. While traditional skeletal motion recognition methods are mature, they rely on deep networks to extract temporal features, resulting in computational complexity and low energy efficiency. Therefore, this embodiment provides a motion recognition method based on a redox-based photomemristor. Its core innovation lies in the deep physical integration of sensing, dynamic feature extraction, and computation using isomorphic devices. Specifically, human movement manifests as the evolution of joint spatial positions over time. This invention can encode motion trajectories as temporally varying light intensity or light pulse signals. A volatile photoinduced conductive state redox-based photomemristor can dynamically map the temporal light pulse signal into a high-dimensional state, providing short-term memory and nonlinear transformation. A isomorphic non-volatile electrically programmable state redox-based photomemristor serves as a component of the readout layer, storing training weights and performing matrix-vector multiplication (MVM) in parallel in the analog domain, ultimately achieving motion recognition and classification.

[0042] Specifically, in combination Figure 5 As shown, this embodiment first selects five classic human motion datasets: drinking water, throwing, sitting, standing, and jumping. This dataset contains a large number of real-world human motion samples, demonstrating good versatility and representativeness. Furthermore, the five selected classic motions cover core movement patterns such as limb flexion, extension, displacement, and rising and falling, with significant differences in temporal features, effectively validating the system's recognition capabilities. Next, for each classic motion, the three-dimensional coordinate signals (x, y, z) of six key skeletal nodes of the human body are extracted, as shown... Figure 5 As shown, the six key skeletal nodes include: head, torso, left arm, right arm, left leg, and right leg. The three-dimensional coordinates of each skeletal node directly reflect the spatial position and movement trajectory of that part. Six skeletal nodes × three coordinate axes = 18 channels, thus forming 18 channels of temporal data. The data sampling frequency in this embodiment is 30Hz (i.e., a set of coordinate data is collected every 33.3ms) to ensure the complete capture of the dynamic evolution of the motion.

[0043] Furthermore, in this embodiment, the 18-channel timing data is encoded into a sequence of light pulses that can be directly sensed by a redox-type photomemristor that is in a volatile photoinduced conductive state. The encoding rules of this embodiment are as follows: (1) Optical pulse parameter settings: The wavelength is 375nm, which matches the peak wavelength of the photoresponse of the redox photomemristor and can excite the strongest current response. The pulse width is 10ms, which can ensure that the device responds fully and avoid insufficient response caused by the pulse being too narrow. The pulse interval is 30ms, which is synchronized with the sampling frequency of coordinate data and can ensure that each set of coordinate data corresponds to one optical pulse.

[0044] (2) Mapping relationship between light intensity and coordinate value: In this embodiment, a linear mapping rule is adopted. The light intensity I and the coordinate value S satisfy the formula I=k×S+I0, where k is a proportionality coefficient, which can be calibrated according to the range of coordinate data, and I0 is the reference light intensity. Through this mapping, the numerical difference of spatial coordinates can be transformed into the difference in light intensity, so that the light pulse sequence can simultaneously carry spatial position and temporal evolution information.

[0045] Step S200: Construct a reservoir array consisting of several volatile photoinduced conductive states of redox photoresistors, and expand the optical pulse sequence signal into a high-dimensional state vector based on the reservoir array.

[0046] This embodiment employs a reservoir array composed of 18 volatile photo-induced conductivity (DOC) photoresistors in a parallel processing state. Each DOC photoresistor corresponds one-to-one with a channel, meaning each DOC photoresistor receives a separate optical pulse sequence. Each DOC photoresistor receives the optical pulse sequence signal of its corresponding channel independently. The DOC photoresistors in this embodiment utilize a vertical structure, with a dielectric layer of polyoxometalate-doped polystyrene (DDA-SmPW@PS), gold (Au) as the electrode material, and a glass and ITO composite substrate. Under 375nm optical pulse excitation, the device exhibits a photoinduced conductivity effect, and the conductivity state is volatile; that is, after the optical pulse stimulation stops, the conductivity naturally decays over time. This characteristic provides a physical basis for short-term memory and temporal feature capture.

[0047] When each volatile photo-induced conductive state redox photoresistor receives a sequence of light pulses, it generates an excitatory postsynaptic current (EPSC), or photocurrent signal, based on the redox photoelectric effect, which has short-term memory characteristics. The greater the intensity of the light pulse, the higher the peak current of the EPSC. The duration and interval of the light pulse determine the accumulation and relaxation process of the current, making the EPSC waveform naturally carry the temporal evolution characteristics of the time-series data. For example, the skeletal coordinates of the "jumping" action change drastically, corresponding to large fluctuations in light intensity, and the EPSC waveform exhibits rapid rise, high peak value, and fast relaxation; while the coordinate changes of the "sitting" action are gradual, and the EPSC waveform is relatively flat.

[0048] To extract the spatiotemporal features of motion signals, the current signal is discretely sampled at fixed time intervals during the current decay process of each volatile photoinduced conductivity state redox photoresistor. Each sampling point is defined as a virtual node. Experimental testing showed that the characteristic time constant of the current decay of the volatile photoresistor selected in this invention is approximately 37.5 ms after the light pulse stops. Therefore, the sampling interval is set to 37.5 μs to ensure that sufficient details are captured during the current decay process. In this embodiment, 40 virtual nodes are continuously sampled for each channel, covering a 1.5 ms decay process, thus fully preserving key information of the timing characteristics.

[0049] Finally, the data from the virtual nodes are fused, expanding the 18 optical pulse sequence signals into a 720-dimensional high-dimensional state vector. This vector transforms the original 18 low-dimensional time-series signals into a high-dimensional feature space representation through parallel sampling and fusion at the hardware level. This high-dimensionalization process not only preserves the spatiotemporal information of the original data but also introduces feature enhancement through the nonlinear response of the device, making the features of different actions more significantly separable in the high-dimensional space. Figure 6 Taking the action of "drinking water" as an example, the X, Y, and Z axis signal encoding and the device response waveforms of each channel are shown, which intuitively reflects the mapping relationship from optical signal to state vector.

[0050] Step S300: Construct a supervised training model based on a high-dimensional state vector, solve the weight matrix, and construct a memristor cross array based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmed state redox-type photomemristors with the same structure.

[0051] In this step, the weight matrix is ​​solved through supervised training and physically deployed onto a memristor cross-array composed of non-volatile electrically programmable redox-prototype photomemristors, establishing a mapping bridge between "high-dimensional features and action categories". Specifically, this embodiment uses the ridge regression algorithm, taking a 720-dimensional high-dimensional state vector as input and one-hot encoded labels of five classic action categories as output to construct a supervised training model; in the one-hot encoded labels, the column corresponding to a certain action category has a value of 1, and the other columns have a value of 0. For example, the label for the action "drinking water" is [1,0,0,0,0], and for "throwing" it is [0,1,0,0,0]. Next, the weight matrix is ​​solved by minimizing the loss function containing the L2 regularization term, resulting in a 20-row × 5-column weight matrix. The loss function formula is:

[0052] in, The loss function is the square of the L2 norm, and λ is the regularization parameter (λ=0.01 determined through cross-validation). This loss function ensures the model's ability to fit the training samples while constraining the sparsity of the weight matrix, thus avoiding overfitting. This refers to the high-dimensional state vector matrix of the training samples. This refers to the one-hot encoded label matrix of the training samples. After solving, the weight matrix W is obtained, establishing a linear mapping relationship from the high-dimensional feature space to the action category space. That is, through matrix multiplication Y≈XW, prediction from features to categories can be achieved.

[0053] Next, to suppress the impact of inherent conductance fluctuations (fluctuation range of approximately ±5%) of redox photomemristors and environmental noise (such as conductance drift caused by temperature changes) on calculation accuracy, this embodiment adopts a differential conductance representation scheme, which assigns each weight value of W in the weight matrix to the differential conductance representation. (Reflecting the first) The high-dimensional feature and the first The mapping coefficients for the type of action are mapped to the conductance difference between a pair of memristor cells, i.e.:

[0054] in, This represents the conductance value of the positive conductance memristor unit. This represents the conductance value of the negative conductance memristor cell. When there are conductance fluctuations in the device, and The fluctuation trend is consistent with that of the weights, and the fluctuation of the difference (i.e., the weight value) is largely offset, thereby improving the stability of the calculation.

[0055] Furthermore, this embodiment determines the difference pairs based on the size of the weight matrix (720×5). In this embodiment, each weight value corresponds to one difference pair. and Therefore, a total of 720 × 5 = 3600 difference pairs are needed. Each difference pair contains two non-volatile opto-memristor units, requiring a total of 7200 memristor units. To achieve a compact layout, this embodiment uses multiple non-volatile electrically programmable redox-type opto-memristors with identical structures to form a memristor cross array to carry the difference pairs. This embodiment uses a 40 × 40 memristor cross subarray, with each subarray containing 40 × 40 = 1600 memristor units, for a total of 8000 units across the five subarrays (reserving redundant units for replacing faulty devices). Each subarray carries a portion of the difference pairs, specifically by dividing the 720-dimensional high-dimensional features into five groups (each group has 144 dimensions, 144 × 5 = 720). Each subarray corresponds to a weight mapping between one set of features and five types of actions. The conductance distribution of the subarrays is optimized using a logistic regression algorithm to ensure consistent computational accuracy across all subarrays.

[0056] Furthermore, in this embodiment, the trained weight matrix is ​​written into an array of non-volatile redox-type photomemristors via electrical programming. The programming process employs closed-loop feedback control to ensure precise matching between the conductance values ​​and the target weights. Specifically, in this embodiment, each weight value is linearly mapped to a preset target conductance range (0.1-10 μS), using the following mapping formula:

[0057] in, and These are the positive and negative components of the weight value, respectively. ), , These are the upper and lower limits of conductivity. It is the largest absolute value in the weight matrix.

[0058] Next, in this embodiment, a stepped-incrementing voltage pulse sequence is applied to each opto-memristor. The pulse parameters are: amplitude 0.1-2V (0.05V step), width 50μs, and interval 10ms. After each voltage pulse is applied, the conductance state is read in real time using a semiconductor parameter analyzer. Closed-loop feedback control is used to gradually bring the conductance value closer to the target value. Specifically, this embodiment calculates the error between the current conductance value and the target value. If the error is greater than 1%, the next voltage pulse is applied; if the error is ≤1%, programming stops, and the unit programming is complete; this continues until the conductance value gradually approaches the target value. After all units are programmed, the conductance difference of each differential pair is detected. If the error between the difference and the target weight exceeds 2%, reprogramming is performed; if the error is ≤2%, programming is successful, and the weight matrix is ​​persistently stored in the memristor crossover subarray in the form of a non-volatile conductance distribution.

[0059] Step S400: Based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross array, the high-dimensional features in the current multi-channel photocurrent signals are mapped to action categories through analog domain operations, and the action classification results are output.

[0060] During the hardware inference and classification execution phase, this embodiment can perform analog domain calculations to complete action classification based on a non-volatile memristor cross array with completed weight mapping. The schematic diagram of the RC system circuit of the redox-type photomemristor core in this embodiment is shown below. Figure 7 As shown, Figure 7The ADC in this example is an Analog-to-Digital Converter. Specifically, in this embodiment, the current multi-channel photocurrent signal output from the reservoir array is converted into a voltage signal via a transimpedance amplifier. The gain of the transimpedance amplifier is set to 1MΩ, which can convert nA-level photocurrent into mV-level voltage signals (e.g., 10nA current corresponds to 10mV voltage), meeting the input requirements for subsequent analog-to-digital conversion. Then, through analog-to-digital conversion and controller processing, the detailed features of the signal are captured, generating a 720-dimensional voltage vector corresponding to the current input action frame (each virtual node's sampled value corresponds to one voltage component).

[0061] Next, in this embodiment, a word line driving circuit is used to apply a 720-dimensional voltage vector to the corresponding row (i.e., word line) of the memristor cross array, providing input signals for analog domain operations. The word line driving circuit uses a high-voltage operational amplifier (output voltage range 0-5V, drive current 1mA) to ensure that the voltage signal does not attenuate during long-distance transmission. The 720-dimensional voltage vector is distributed to the word lines of five 40×40 sub-arrays according to the sub-array division: the 40 word lines of each sub-array correspond to 40 components of the 144-dimensional features (the remaining components are processed by time-division multiplexing), ensuring that all feature components are input in parallel.

[0062] Furthermore, this embodiment performs analog domain operations. Specifically, the input voltage vector is distributed along the word lines within the array, and each memristor cell generates a corresponding cell current based on its stored conductance value. and According to Ohm's law, the current in a single cell can be expressed as: , ,in, For the first The input voltage of the bar line, where, The dimension index corresponding to the high-dimensional state vector also corresponds to the first dimension of the memristor crossbar array. bar lines The index corresponding to the action category also corresponds to the first index of the memristor cross array. Therefore, the bar lines. For the first Input voltage of bar lines The effect of the line, number The cell current generated on the corresponding positive conductance memristor cell. For the first Input voltage of bar lines The effect of the line, number The cell current generated on the negative conductance memristor cell corresponding to the column. According to Kirchhoff's current law, the total output current on each bit line is the sum of the cell currents in that column, i.e., the current on the th bit line. The sum of the currents output by all positive conductance memristor units on the bar is expressed as: ;No. The sum of the currents output by all negative conductance memristor units on the bar is expressed as: For the differential weighting structure, the final output current is This embodiment utilizes a differential conductance weighting structure to automatically perform the multiplication and accumulation operation of the input voltage vector and the differential conductance weighting matrix in the analog domain, achieving efficient matrix-vector multiplication. The entire operation is highly parallel, with all units working simultaneously, and the computational latency is less than 100μs, far superior to the serial computation of digital processors.

[0063] Finally, in this embodiment, the output current signal is converted into a voltage signal by a readout circuit (composed of a transimpedance amplifier and a low-pass filter). The cutoff frequency of the low-pass filter is 10kHz to filter out high-frequency noise. Each voltage signal corresponds to a classification score for a type of action. The amplitude of the converted voltage signal is proportional to the action category score. The action category with the highest score is identified by a comparison decision circuit, and the action classification result is output. Specifically, in this embodiment, the comparison decision circuit compares the amplitudes of the voltage signals corresponding to the five types of actions and identifies the voltage signal with the highest score. The corresponding action category is the system's identification result. The comparison decision circuit uses a high-speed comparator (response time ≤ 10ns) to ensure real-time decision-making. Then, the microcontroller receives the signal from the comparison decision circuit, outputs the action category identifier (e.g., numbers 0-4 correspond to drinking water, throwing, sitting down, standing, and jumping, respectively), and can transmit it to external devices via serial port, Bluetooth, or other interfaces.

[0064] Furthermore, this embodiment uses an independent test set containing 200 samples (40 samples per action, none of which were used in the training process) to verify the system's recognition accuracy. Test environment: temperature 25℃, humidity 50%, power supply voltage 3.3V, system power consumption monitoring uses a power analyzer (accuracy 1μW). Test results show that the system's average recognition accuracy is 90.5%, specifically as follows... Figure 8 As shown in the image.

[0065] In other implementations, this embodiment can also design other forms of memristor structures, such as by changing the device structure of the memristor, such as a three-terminal phototransistor, or by changing the electrode material of the photomemristor, such as silver, platinum, copper, etc. This embodiment is not limited to this.

[0066] This invention utilizes polyoxometalates as the core functional material to construct a molecular-type photoelectric memristor. Unlike traditional filamentary mechanisms, the resistance state switching of this device is regulated by the material's well-defined redox state, exhibiting high design flexibility and controllability. Its charge state can be adjusted via electric and optical fields: electrical excitation induces a stable reduced state, enabling non-volatile data storage; optical excitation triggers a reversible redox reaction, causing dynamic conductivity changes with relaxation behavior, endowing the device with optical sensing and signal preprocessing capabilities. This mechanism provides a physical basis for integrating the transient response required for the reservoir layer and the persistent storage required for the readout layer into a single device. Based on this photoelectric memristor, a motion recognition method for skeletal motion signal perception and memory is constructed. When a sequence of light carrying skeletal coordinate information illuminates the device, its conductivity changes volatilely with light intensity and time, manifesting as the accumulation of photocurrent and relaxation effects. Utilizing this light-dependent dynamic electrical response, the memristor can directly map the temporally varying skeletal motion trajectory into a differentiated conductivity state sequence, thereby directly extracting the spatiotemporal characteristics of human motion at the physical level.

[0067] This invention directly processes analog optical signals using an opto-memristor array, eliminating the need for complex digital feature extraction algorithms. Motion features can then be directly fed into a backend classification network for action recognition. This highly integrated architecture of sensing, storage, and computing fundamentally eliminates the overhead of redundant data transmission and processing, providing an efficient, low-latency, and robust hardware solution for real-time, noise-resistant motion recognition in scenarios such as intelligent monitoring and human-computer interaction.

[0068] Based on the above embodiments, the present invention also provides a motion recognition system based on a redox photomemristor, the system being used to implement the steps in the above method embodiments. Figure 9As shown, the system in this embodiment includes: a time-series data encoding module 10, a high-dimensional state expansion module 20, a memristor cross array construction module 30, and an action recognition module 40. Specifically, the time-series data encoding module 10 is used to select classic actions based on a human action dataset, extract time-series data based on the selected classic actions, and encode the time-series data into a light pulse sequence. The high-dimensional state expansion module 20 is used to construct a reservoir array composed of several volatile photo-induced conductive states of redox-type photomemristors, and expand the light pulse sequence signal into a high-dimensional state vector based on the reservoir array. The memristor cross array construction module 30 is used to construct a supervised training model based on the high-dimensional state vector, solve the weight matrix, and construct a memristor cross array based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmed states of redox-type photomemristors with identical structures. The action recognition module 40 is used to map the high-dimensional features of the current multi-channel photocurrent signals output by the reservoir array to action categories through analog domain operations, based on the memristor cross array, and output the action classification results.

[0069] The motion recognition system based on redox photomemristors in this embodiment is based on the same principle as the steps in the above method embodiments, and will not be described in detail here.

[0070] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 10 As shown. The terminal may include one or more processors 100 ( Figure 10 (Only one is shown in the image), memory 101, and computer program 102 stored in memory 101 and executable on one or more processors 100. For example, an entity extraction processing program. When one or more processors 100 execute computer program 102, they can implement various steps in the motion recognition method embodiment based on a redox photomemristor. Alternatively, when one or more processors 100 execute computer program 102, they can implement the functions of various modules / units in the motion recognition system embodiment based on a redox photomemristor, without limitation herein.

[0071] In one embodiment, the processor 100 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0072] In one embodiment, memory 101 may be an internal storage unit of an electronic device, such as a hard drive or RAM. Memory 101 may also be an external storage device of the electronic device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital Card (SD), or Flash Card. Furthermore, memory 101 may include both internal and external storage units. Memory 101 is used to store computer programs and other programs and data required by the terminal. Memory 101 can also be used to temporarily store data that has been output or will be output.

[0073] Those skilled in the art will understand that Figure 10 The block diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0074] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct memory bus RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motion recognition method based on a redox photomemristor, characterized in that, The method includes: Classic movements are selected based on a human motion dataset, time-series data is extracted based on the selected classic movements, and the time-series data is encoded into a light pulse sequence; A reservoir array consisting of several volatile photoinduced conductive states of redox photoresistors is constructed, and the optical pulse sequence signal is extended into a high-dimensional state vector based on the reservoir array. A supervised training model is constructed based on a high-dimensional state vector, the weight matrix is ​​solved, and a memristor cross array is constructed based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmable redox-type photomemristors with the same structure. Based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross-array, the mapping of high-dimensional features in the current multi-channel photocurrent signals to action categories is realized through analog domain operations, and the action classification results are output.

2. The motion recognition method based on a redox photomemristor according to claim 1, characterized in that, Classic movements are selected based on a human motion dataset. Time-series data is extracted from the selected classic movements and encoded into a light pulse sequence, including: Five classic movements were selected based on the human motion dataset: drinking water, throwing, sitting down, standing, and jumping. For each type of classic movement, the three-dimensional coordinate signals of 6 key skeletal nodes of the human body are extracted to form 18 channels of time-series data; The 18-channel timing data is encoded into a sequence of light pulses that can be directly sensed by a redox-type photomemristor that is in a volatile photoinduced conductive state.

3. The motion recognition method based on a redox photomemristor according to claim 2, characterized in that, A reservoir array consisting of several volatile photo-induced conductive states of redox photoresistors is constructed. Based on the reservoir array, the optical pulse sequence signal is extended into a high-dimensional state vector, including: A reservoir array with parallel processing is composed of 18 volatile photo-induced conductive state redox photoresistors. Each redox photoresistor corresponds to a channel, and each redox photoresistor receives the optical pulse sequence signal of the corresponding channel individually. When each redox-type photomemristor in a volatile photoinduced conductive state receives a sequence of light pulses, it generates a photocurrent signal with short-term memory characteristics. During the current decay process of the redox-type photomemristor in each volatile photoinduced conductive state, the current signal is discretely sampled at fixed time intervals. Each sampling point is defined as a virtual node, and 40 virtual nodes are continuously sampled for each channel. The data from the virtual nodes are fused, and the 18-channel optical pulse sequence signals are expanded into a 720-dimensional high-dimensional state vector.

4. The motion recognition method based on a redox photomemristor according to claim 3, characterized in that, A supervised training model is constructed based on a high-dimensional state vector, and the weight matrix is ​​solved, including: A supervised training model is constructed by using the ridge regression algorithm, taking a 720-dimensional high-dimensional state vector as input and one-hot encoded labels of five classic actions as output. The weight matrix is ​​obtained by minimizing the loss function with L2 regularization, resulting in a weight matrix of 20 rows and 5 columns.

5. The motion recognition method based on a redox photomemristor according to claim 4, characterized in that, Constructing a memristor cross-connect array based on a weight matrix includes: A differential conductance representation scheme is adopted, which maps each weight value in the weight matrix to the conductance difference of a pair of memristor cells; Based on the size of the weight matrix, the difference pairs are determined, and multiple non-volatile electrically programmed photoresistors with the same structure are used to form a memristor cross array to carry the difference pairs. Each weight value in the weight matrix is ​​linearly mapped to a preset target conductance range, and a stepped-incrementing voltage pulse sequence is applied to each photomemristor. The conductivity state is read in real time, and the conductivity value is gradually brought closer to the target value through closed-loop feedback control. The weight matrix is ​​persistently stored in the memristor cross subarray in the form of non-volatile conductivity distribution.

6. The motion recognition method based on a redox photomemristor according to claim 5, characterized in that, Based on the current multi-channel photocurrent signals output by the reservoir array, and combined with the memristor cross-connect array, analog domain operations are used to map the high-dimensional features of the current multi-channel photocurrent signals to action categories, outputting action classification results, including: The current multi-channel photocurrent signal output by the reservoir array is converted into a voltage signal, and then processed by analog-to-digital conversion and controller to generate a voltage vector corresponding to the current input action frame; The voltage vector is applied to the corresponding row of the memristor cross array through the word line driving circuit, providing an input signal for analog domain operations; Perform analog domain operations and convert the output current signal into a voltage signal through a readout circuit. Each voltage signal corresponds to a classification score for a type of action. The action category with the highest score is identified by a comparison decision circuit, and the action classification result is output.

7. The motion recognition method based on a redox photomemristor according to claim 6, characterized in that, Performing analog domain operations includes: The input voltage vector is distributed along the word line within the array, and each memristor cell generates a corresponding cell current based on the stored conductance value; According to Kirchhoff's current law, the total output current on each bit line is the sum of the currents of all cells in that column; By using the differential conductance weighting structure, the multiplication and accumulation operations of the input voltage vector and the differential conductance weighting matrix are automatically completed in the analog domain.

8. A motion recognition system based on a redox photomemristor, characterized in that, The system is used to implement the steps of the motion recognition method based on a redox photomemristor as described in any one of claims 1-7, and the system comprises: The time-series data encoding module is used to select classic movements based on the human motion dataset, extract time-series data based on the selected classic movements, and encode the time-series data into a light pulse sequence. A high-dimensional state extension module is used to construct a reservoir array composed of several volatile photoinduced conductive states of redox photoresistors, and to extend the optical pulse sequence signal into a high-dimensional state vector based on the reservoir array. The memristor cross array construction module is used to build a supervised training model based on a high-dimensional state vector, solve the weight matrix, and construct a memristor cross array based on the weight matrix. The weight matrix is ​​used to establish a linear mapping from the high-dimensional feature space to the action category space. The memristor cross array is an array composed of multiple non-volatile electrically programmable redox-type photomemristors with the same structure. The action recognition module is used to map the high-dimensional features of the current multi-channel photocurrent signals output by the reservoir array to action categories through analog domain operations, in conjunction with the memristor cross array, and output the action classification results.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a motion recognition program based on a redox photomemristor stored in the memory and executable on the processor. When the processor executes the motion recognition program based on the redox photomemristor, it implements the steps of the motion recognition method based on a redox photomemristor as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a motion recognition program based on a redox photomemristor, the motion recognition program based on the redox photomemristor implementing the steps of the motion recognition method based on a redox photomemristor as described in any one of claims 1-7 on the computer-readable storage medium.

Images