High-precision neuromorphic object identification method based on HfZrO4 memristor

Abstract

The invention discloses a high-precision neuromorphic object recognition method based on an HfZrO4 memristor, which comprises the HfZrO4 memristor and a plurality of temperature-pressure dual-mode sensors, the plurality of temperature-pressure dual-mode sensors are uniformly arranged on a glove, the glove is worn on a hand to grab target objects with different shapes and temperatures, and the temperature-pressure dual-mode sensors and the temperature-pressure dual-mode sensors are connected with the temperature-pressure dual-mode sensors to obtain the high-precision neuromorphic object recognition method based on the HfZrO4 memristor. Sensing signals collected by the plurality of temperature-pressure bimodal sensors are fed back to the HfZrO4 memristor so as to carry out bionic neuromorphic calculation processing, and machine learning is carried out on the processed sensing signals so as to identify target objects with different shapes and temperatures. According to the artificial perception neuromorphic recognition system disclosed by the invention, bionic neuromorphic processing is carried out on the sensing signals through the HfZrO4 memristor, the processed signals have more contrast ratios and differences, then the signals processed by the memristor are processed and recognized through machine learning, and the recognition accuracy is high.

Description

Technical Field

[0001] The present invention relates to the field of neuromorphic recognition technology, and in particular to a high-precision neuromorphic object recognition method based on HfZrO4 memristor. Background Art

[0002] The human body interacts with the surrounding environment to enable us to sense and recognize objects. The human skin contains a variety of receptors, such as pressure and temperature receptors, which respond to external stimuli and transmit stimulus information to the brain through electrical spikes, thus giving us a delicate sense of touch for object recognition, texture discrimination and sensory feedback. This biological sensory system has significant advantages such as compactness, parallelism, event-driven computing and noise resistance, which enriches tactile perception and effectively solves many complex and unstructured real-world problems.

[0003] The biological sensory system has three key components, including receptor cells, nerve channels, and brain information processing centers. Therefore, the corresponding artificial sensory nervous system also contains three core components: a temperature-pressure dual-modal sensor that converts external physical stimuli into electrical signal pulses, a memristor that integrates and converts signals into excitatory postsynaptic currents (EPSCs), and artificial axons that connect them. The artificial sensory nervous system can perceive external signals with lower power consumption and realize data processing and memory functions by adjusting synaptic weights. Therefore, the artificial sensory nervous system is a new paradigm for developing the next generation of low-power and highly bionic neuromorphic systems.

[0004] The most advanced artificial spiking neurons are usually made of dozens of transistors and complex circuits interconnected, which are bulky, energy-intensive, require software assistance, and are not suitable for large-scale integration and efficient implementation. Therefore, exploring emerging electronic components with simple structures and naturally exhibiting rich neuronal spiking behaviors using their inherent physical properties will form compact and powerful neuromorphic systems. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a high-precision neuromorphic object recognition method based on HfZrO4 memristor, which can use the same polarity voltage to increase and decrease the resistance.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A high-precision neuromorphic object recognition method based on HfZrO4 memristor, comprising an HfZrO4 memristor and a plurality of temperature-pressure dual-modal sensors, wherein the plurality of temperature-pressure dual-modal sensors are evenly arranged on a glove, and the glove is worn on the hand to grasp targets of different shapes and temperatures, and the sensing signals collected by the plurality of temperature-pressure dual-modal sensors are fed back to the HfZrO4 memristor for bionic neuromorphic computing processing, and machine learning is performed on the processed sensing signals to identify targets of different shapes and temperatures;

[0008] The HfZrO4 memristor includes a Pt top electrode, a HfZrO4 resistive layer and a Pt bottom electrode arranged from top to bottom. The Pt bottom electrode and the Pt top electrode form a loop with the HfZrO4 resistive layer after being conductive, and the resistance of the loop is changed by the electrical signal applied to the HfZrO4 resistive layer.

[0009] The specific steps of performing machine learning on the processed sensor signals to identify targets of different shapes and temperatures are as follows:

[0010] S1: The electrical signal processed by the HfZrO4 memristor is collected by a Keithley 4200A-SCS parameter analyzer to obtain a current signal, and the obtained Excel current table is exported, the header is removed, and only the two columns of time and current data are retained;

[0011] S2: Use the tsfresh library to extract features from the two columns of data. The features extracted by tsfresh include: mean, variance, kurtosis, skewness, zero crossing rate, number of peaks, main frequency, and spectrum energy;

[0012] S3: classify the extracted features by labels, and convert each feature data for label classification into a standard normal distribution with a mean of "0" and a standard deviation of "1";

[0013] S4: Use PCA to reduce the dimension of feature data by finding the direction with the largest variance in the feature data and projecting the feature data to the direction with the largest variance to achieve dimensionality reduction;

[0014] S5: Finally, the random forest algorithm is used to obtain the accuracy of the identified objects.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The artificial perception neuromorphic recognition system disclosed in the present invention performs bionic neuromorphic processing on the sensing signal through the HfZrO4 memristor, and the processed signal has more contrast and difference. Then, through machine learning, the signal processed by the memristor is processed and recognized with high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a layout diagram of the present invention;

[0018] Figure 2 It is a schematic diagram of the structure of HfZrO4 memristor;

[0019] Figure 3 This is a graph of the voltage-current cycle measurement results of the HfZrO4 memristor;

[0020] Figure 4 This is a diagram showing the results of a high- and low-resistance cycle measurement of the HfZrO4 memristor;

[0021] Figure 5 This is a graph of the excitatory postsynaptic current generated by the current peak when a set of -5V voltage is applied during the bionic performance measurement of the HfZrO4 memristor;

[0022] Figure 6 This is a graph of inhibitory postsynaptic current generated by the current peak when a set of -3V voltage is applied during the bionic performance measurement of HfZrO4 memristor;

[0023] Figure 7 It is a flowchart of the present invention for identifying a target object;

[0024] Figure 8 This is a diagram showing the recognition result of the target object according to the present invention.

[0025] Figure symbols: 11. Pt top electrode, 12. HfZrO4 resistive layer, 13. Pt bottom electrode, 14. Insulating layer, 2. Temperature-pressure dual-mode sensor, 3. Gloves. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0027] Example 1

[0028] A high-precision neuromorphic object recognition method based on HfZrO4 memristor, such as Figure 1As shown, it includes HfZrO4 memristor and multiple temperature-pressure dual-mode sensors 2, multiple temperature-pressure dual-mode sensors 2 are evenly arranged on the glove 3, and the glove 3 is worn on the hand to grab targets of different shapes and temperatures. The sensor signals collected by the multiple temperature-pressure dual-mode sensors 2 are fed back to the HfZrO4 memristor for bionic neuromorphic computing processing, and the processed sensor signals are machine-learned to identify targets of different shapes and temperatures. The recognition model is a random forest algorithm, which is composed of multiple decision trees, each of which is independently trained and outputs results when predicting. For classification tasks, the random forest selects the most common category as the final prediction result through a "voting" mechanism; for regression tasks, the average of the prediction values ​​of all trees is calculated.

[0029] When training each decision tree, Random Forest uses the Bagging (Bootstrap Aggregating) method. That is, samples are randomly extracted from the training data with replacement to generate multiple different sample sets, so that each tree is trained on a different data set. This method can reduce the variance of the model and thus improve the generalization performance.

[0030] When constructing the nodes of each tree, random forest randomly selects a subset in the feature space for splitting instead of using all features. This randomness reduces the correlation between trees, making the final ensemble more stable while avoiding overfitting problems.

[0031] Specific steps are as follows Figure 7 As shown:

[0032] S1: The electrical signal of the HfZrO4 memristor is collected by Keithley4200A-SCS parameter analyzer to obtain the current signal, and the obtained Excel current table is exported, the table header is removed, and only the two columns of time and current data are retained;

[0033] S2: Use the tsfresh library to extract features from the two columns of data. It can automatically extract statistical, morphological, frequency and other features from time series data, and filter these features to ensure that only the most important features are selected. The features extracted by tsfresh include: mean, variance, kurtosis, skewness, zero crossing rate, number of peaks, main frequency, spectrum energy and other 783 features;

[0034] S3: label and classify the extracted features. Specifically, label multiple targets in order and convert each feature data for label classification into a standard normal distribution with a mean of "0" and a standard deviation of "1", so as to avoid the feature dominating the training process of the recognition model due to excessively large or small values;

[0035] S4: Use PCA to reduce the dimension of feature data. Dimension reduction is achieved by finding the direction with the largest variance in the feature data and projecting the feature data to the direction with the largest variance. In this algorithm, n_components=0.9, then at least 90% of the data variance is retained. Specifically, PCA is the abbreviation of Principal Component Analysis (PCA), which is a commonly used dimensionality reduction algorithm that aims to reduce the dimension of data by projecting high-dimensional data into a lower-dimensional space while retaining the main information of the original data as much as possible. Each principal component of PCA is a linear combination, which represents the direction of the largest variance in the data. The first principal component is the direction with the largest variance, the second principal component is the direction with the second largest variance, and is orthogonal to the first principal component. By analogy, PCA retains as much information as possible by selecting principal components with larger variances, reduces the data dimension, and achieves the effect of compressing data. The mathematical expression is as follows:

[0036] Assume the data matrix is ​​X and the covariance matrix is Solve the eigenvalues ​​and eigenvectors of the covariance matrix to obtain the matrix V of the first X principal components k , the data after dimension reduction is Z = XV k ;

[0037] PCA is an effective linear dimensionality reduction method that can compress features by maximizing data variance.

[0038] S5: Use the random forest algorithm to obtain the accuracy of the identified object. Specifically, the random forest algorithm has good performance when dealing with high-dimensional data, nonlinear relationships, and classification problems. Among them, a single decision tree is prone to overfitting the training data, while the random forest can reduce the risk of overfitting by constructing multiple trees and combining their prediction results, especially when the data volume is small or the noise is large. The specific implementation steps are as follows:

[0039] Model parameter setting: Before training the model, set the number of decision trees of the random forest, the maximum depth of each tree, the minimum number of samples required to split a node, and the minimum number of samples required for a leaf node;

[0040] Training model: Use training set data for training so that the random forest model can learn the relationship between category labels and features;

[0041] Grid search: It is an automated hyperparameter optimization method that uses the maximum depth of the tree, the number of trees, and the minimum number of samples for a node to continue splitting to find the optimal parameters. In this algorithm, the number of decision trees is set to 50, 100, and 200, the maximum depth is set to 10, 20, 30, and None, the minimum number of samples for a node to continue splitting is set to 2, 5, and 10, and the minimum number of samples required for a node is set to 1, 2, and 4. After arranging and comparing different parameters, the best parameters are: the number of decision trees is 50, the maximum depth is 10, the minimum number of samples for a node to continue splitting is 2, and the minimum number of samples required for a node is 1, which ultimately gives the highest accuracy.

[0042] The above-mentioned HfZrO4 memristor can realize the synaptic function of excitatory postsynaptic current and inhibitory postsynaptic current, indicating that the HfZrO4 memristor has synaptic plasticity and has the ability to realize neuromorphic computing. The specific structure is as follows Figure 2As shown, it includes a Pt top electrode 11, a HfZrO4 resistive switching layer 12 and a Pt bottom electrode 13 arranged in sequence from top to bottom. The Pt bottom electrode 13 and the Pt top electrode 11 form a loop with the HfZrO4 resistive switching layer 12 after being conductive, and the resistance of the loop is changed by the electric signal applied to the HfZrO4 resistive switching layer 12, wherein the electric signal is generated by the temperature-pressure dual-mode sensor 2 during operation; the Pt top electrode 11 and the Pt bottom electrode 13 are made of metal Pt, and their main function is to conduct electricity. The work function of metal Pt is the highest, which is 5.36eV, which is comparable to that of HfZrO4. The combination of HfZrO4 resistive switching layer 12 will form a gold-semiconductor contact with a huge barrier height. At this time, the resistance of HfZrO4 resistive switching layer 12 is relatively high, so that the resistance of HfZrO4 memristor is also relatively high; the above-mentioned HfZrO4 resistive switching layer 12 contains a large number of oxygen vacancies. When the temperature-pressure dual-mode sensor 2 applies voltage to the Pt top electrode 11 and the Pt bottom electrode 13, the oxygen vacancies in the HfZrO4 resistive switching layer 12 move to the contact surface of the semiconductor HfZrO4 resistive switching layer 12 and the Pt top electrode 11 or the Pt bottom electrode 13, the contact barrier will be reduced, thereby making The resistance of the HfZrO4 resistive layer 12 is reduced. As the voltage is applied, the oxygen vacancies in the HfZrO4 resistive layer 12 continue to be affected by the electric field, move and accumulate to form conductive filaments. At this time, the resistance of the HfZrO4 memristor is also reduced. In summary, using metal Pt as an electrode can maximize the high-to-low resistance ratio of the HfZrO4 memristor; the resistance of the HfZrO4 resistive layer 12 changes with the history of the voltage applied to the circuit. Specifically, when a pulse voltage is applied to the Pt top electrode 1 or the Pt bottom electrode 13, the current of the HfZrO4 resistive layer 12 will slowly increase, and the current The resistance will decrease. When the voltage is increased, a conductive channel will be formed in the HfZrO4 resistive layer 12, which is the conductive filament mentioned above. At this time, the resistance of the HfZrO4 resistive layer 12 is the lowest. Then, the voltage is continuously increased, and the oxygen vacancies in the HfZrO4 resistive layer 12 continue to move. One end of the conductive filament begins to break. At this time, the resistance of the HfZrO4 resistive layer 12 increases, and the resistance first decreases and then increases. This phenomenon is in line with the performance characteristics of complementary memristors, and then a single HfZrO4 memristor can be used to realize bionic neuromorphic computing without the need for complex circuit components.

[0043] As a preferred solution of the above embodiment, the Pt bottom electrode 13 is arranged on the insulating layer 14 , and the insulating layer 14 plays a role in preventing the Pt bottom electrode 13 from leaking current and short circuit.

[0044] Example 2

[0045] On the basis of Example 1, the manufacturing method of the HfZrO4 memristor comprises the following steps:

[0046] S1, cutting a silicon wafer having a silicon oxide layer on the surface as a substrate, and the silicon oxide layer as an insulating layer 14;

[0047] S2, after cleaning the insulating layer 14, a layer of metal Pt is magnetron sputtered as the Pt bottom electrode 13. Specifically, when cleaning the insulating layer 14, water, ethanol, and acetone are ultrasonically applied for 20 minutes each. Ultrasonication can effectively loosen the attachments on the insulating layer 14, and water, ethanol, and acetone remove surface pollutants and organic matter to ensure that the surface of the insulating layer 14 is clean;

[0048] S3, using a plasma cleaning machine to clean and activate the Pt bottom electrode 3. Specifically, when the plasma cleaning machine cleans and activates the Pt bottom electrode 13, oxygen is passed through the plasma cleaning machine at a power of 50W for 5 minutes, which can remove pollutants on the surface of the Pt bottom electrode 13, increase the surface activity of the Pt bottom electrode 13, and introduce hydroxyl groups to prepare for the deposition and growth of the HfZrO4 resistive layer 12.

[0049] S4, performing multiple atomic layer depositions on the cleaned and activated bottom electrode 3 in the order of Hf→H2O→Zr→H2O to form a HfZrO4 resistive switching layer 12;

[0050] S5. A plurality of metal Pts are magnetron sputtered as Pt top electrodes 1 on the HfZrO4 resistive layer 2 to form a HfZrO4 memristor. Preferably, the formed HfZrO4 memristor is placed in a rapid annealing furnace, argon is passed through, and annealed at 500°C for 1 minute to crystallize the grown HfZrO4 resistive layer 12 and reduce the defect concentration. Good crystal quality can significantly improve the switching performance, storage density and durability of a high-precision neuromorphic object recognition method based on the HfZrO4 memristor.

[0051] 3. Verification Example

[0052] The thickness of the Pt top electrode 11 of the HfZrO4 memristor is 20 nm, the thickness of the HfZrO4 resistive layer 12 is 10 nm, the thickness of the Pt bottom electrode 13 is 20 nm, and the thickness of the insulating layer 14 is 50 nm.

[0053] 1. The electrical performance of the HfZrO4 memristor is measured. The results are as follows: Figure 3 As shown, a voltage of 0V→3V→0V→-5V→0V is applied to the memristor through Keithley2636B and a probe. In the same polarity voltage region, the resistance first decreases and then increases.

[0054] Here’s how it works:

[0055] When voltage is initially applied, oxygen vacancies move under the influence of the electric field to form conductive filaments, and the device resistance decreases. As voltage continues to be applied, oxygen vacancies continue to move in a direction, which in turn causes the conductive filament to rupture and the resistance to increase. This phenomenon is consistent with the characteristics of complementary memristors.

[0056] 2. The HfZrO4 memristor was tested with high and low resistance cycles. The results are as follows: Figure 4 As shown, the ratio of high to low resistance states is about 100, and 200 cycles were performed, indicating the potential of HfZrO4 memristor in information storage applications and its good cyclability.

[0057] 3. The bionic performance of the HfZrO4 memristor is measured by applying a series of equal pulse voltages to the memristor using the Keithley4200A-SCS parameter analyzer and probe station, as follows:

[0058] like Figure 5 As shown in , when a set of -5V voltage is applied, the current is gradually increased, indicating that the excitatory postsynaptic current (EPSC) generated by the latter current peak is larger than that of the previous one; Figure 6 As shown, when a set of -3V voltage is applied, the current obtained gradually decreases, indicating that the inhibitory postsynaptic current (DPSC) generated by the latter current peak is smaller than the previous one. These phenomena are consistent with the way signals are processed in vivo.

[0059] When the applied voltage pulse stops, the HfZrO4 memristor will have a current opposite to the voltage value. This is because the HfZrO4 resistive layer 12 is a high dielectric constant material, which makes the HfZrO4 memristor itself have parasitic capacitance, that is, when the voltage of the same polarity is used, the current can be reduced or increased.

[0060] 4. The HfZrO4 memristor performs bionic processing on the signal from the temperature-pressure dual-mode sensor 2, and then performs machine learning on the processed signal, such as Figure 7As shown in the figure, firstly, the signal is preprocessed, then the tsfresh library is used to collect data features, then the features are labeled, standardized and the data dimension is reduced, and finally the grid search is used to obtain the optimal parameters of the random forest model to determine the shape and temperature of the grasped object. Specifically, in this verification example, targets of different temperatures and shapes are grasped 100 times, that is, each target has 100 data sets, of which 70% of the data sets are used as training sets and 30% of the data sets are used as test sets. The targets include but are not limited to squares at 25°C. The recognition results of the objects with different shapes and temperatures are shown in Figure 2. The recognition results of the objects with different shapes and temperatures are shown in Figure 2. The recognition results of the objects with different shapes and temperatures are shown in Figure 2. Figure 8 , it can be seen that the true values ​​and predicted values ​​of these 30 test sets are the same, and the accuracy rate reaches 100%, thus realizing high-precision neuromorphic object recognition based on HfZrO4 memristor.

[0061] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, technicians familiar with the field may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A high-precision neuromorphic object recognition method based on HfZrO4 memristor, characterized in that: It comprises a HfZrO4 memristor and a plurality of temperature-pressure dual-mode sensors (2), wherein the plurality of temperature-pressure dual-mode sensors (2) are evenly arranged on a glove (3), and the glove (3) is worn on the hand to grasp objects of different shapes and temperatures, and the sensing signals collected by the plurality of temperature-pressure dual-mode sensors (2) are fed back to the HfZrO4 memristor for bionic neuromorphic processing, and machine learning is performed on the processed sensing signals to identify objects of different shapes and temperatures; The HfZrO4 memristor comprises a Pt top electrode (11), a HfZrO4 resistive switching layer (12) and a Pt bottom electrode (13) which are arranged in sequence from top to bottom; the Pt bottom electrode (13) and the Pt top electrode (11) form a loop with the HfZrO4 resistive switching layer (12) after being conductive, and the resistance of the loop is changed by an electrical signal applied to the HfZrO4 resistive switching layer (12).

2. A high-precision neuromorphic object recognition method based on HfZrO4 memristor according to claim 1, characterized in that: The processed sensor signals are machine-learned to identify objects of different shapes and temperatures through the following steps: S1: The electrical signal processed by the HfZrO4 memristor is collected by a Keithley 4200A-SCS parameter analyzer to obtain the current signal, and the obtained Excel current table is exported, the table header is removed, and only the two columns of time and current data are retained; S2: Use the tsfresh library to perform feature extraction on the two columns of data, where the features extracted by tsfresh include: mean, variance, kurtosis, skewness, zero crossing rate, number of peaks, main frequency, and spectrum energy; S3: Label the extracted features and convert each feature data after label classification into a standard normal distribution with a mean of 0 and a standard deviation of 1; S4: Use PCA to reduce the dimension of feature data by finding the direction with the largest variance in the feature data and projecting the feature data to the direction with the largest variance to achieve dimensionality reduction; S5: Finally, the random forest algorithm is used to obtain the accuracy of the identified objects.

Images