Angle identification method and device for data enhancement of thin film bending sensor

Abstract

The invention discloses an angle identification method and device for data enhancement of a thin film bending sensor, and belongs to the field of deep learning. The invention provides an electrical impedance tomography (EIT) inspired angle identification method for data enhancement of a thin film bending sensor, which comprises the following steps of: arranging a series of metal electrodes on the edge of a microstructure etching pattern to detect impedance values pairwise; therefore, the output signal of the thin film sensor is enhanced into a feature vector reflecting the impedance value distribution condition, and finally angle prediction is carried out based on the feature vector through a deep learning network. According to the invention, the impedance change on the thin film bending sensor can be monitored and analyzed more accurately, and the angle identification performance of the bending sensor is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and in particular relates to an angle recognition method and device for enhancing film bending sensor data. Background Art

[0002] Film-based bending sensors are usually made of lightweight and flexible carbon-based materials such as carbon nanotubes and graphene. These materials are known for their excellent conductivity and flexibility. In addition, their chemical functionalization potential and large-scale production capabilities make it easy for sensors made of these materials to adapt to different forms of bending and be easily integrated into various wearable devices and flexible mechanisms.

[0003] However, due to the non-uniformity of thin film materials, tiny differences in thickness and quality may be introduced during the manufacturing process. These differences will affect the sensor's response to bending or pressure changes. In addition, thin film bending sensors rely on the microstructure of their surface to detect bending. These microstructures may be damaged or worn during long-term use, thus affecting the performance and accuracy of the sensor. The electrical signal processing and conversion mechanism of traditional thin film bending sensors uses the overall impedance of the film to sense the angle changes of the sensor. This type of solution has difficulty identifying detailed changes on the film caused by the influence of physical quantities, and the output value is easily affected by external factors, making it difficult to guarantee accuracy. Therefore, how to achieve the perception of the global impedance changes of the film and thereby achieve higher-precision bending angle detection is a technical problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to solve the above technical problems in the prior art and to provide an angle recognition method and device for enhancing film bending sensor data.

[0005] The specific technical solutions adopted by the present invention are as follows:

[0006] In a first aspect, the present invention provides an angle recognition method for data enhancement of a thin film bending sensor, wherein the thin film bending sensor is a piezoresistive thin film sensor, and a microstructure etching pattern is symmetrically etched with a rotation axis as the center line at the bending position of the film, and a series of metal electrodes are arranged around the edge of the microstructure etching pattern; the angle recognition method comprises:

[0007] S1. When the thin film bending sensor is at a bending angle to be detected, all metal electrodes are connected to a voltage divider circuit in pairs, so as to measure the impedance value between each pair of metal electrodes under a constant voltage;

[0008] S2. The impedance values ​​between all metal electrode pairs are combined into an impedance feature vector, and the impedance feature vector is input into a supervised training deep learning model to obtain a bending angle prediction value of the thin film bending sensor.

[0009] Preferably, as in the first aspect above, the deep learning model consists of an input layer, a backbone network, a classification module, a regression module, and an output layer;

[0010] The input layer is used to subtract the impedance feature vector from the zero-angle feature vector to obtain a non-zero feature vector, and then splice the impedance feature vector and the non-zero feature vector as the input of the backbone network; the zero-angle feature vector is the impedance feature vector obtained by the thin-film bending sensor when the bending angle is zero;

[0011] The backbone network is composed of cascaded multi-layer deep neural networks, and is used to extract deep features from the input spliced vector and input them into the classification module and the regression module;

[0012] The classification module performs classification prediction with a resolution of 1 degree within the bending angle value range based on the deep features, and outputs the integer part of the bending angle;

[0013] The regression module performs regression estimation within the range of 0 to 1 degree based on the deep features, and outputs the decimal part of the bending angle;

[0014] The output layer is used to add the integer part and the decimal part of the bending angle and output the final bending angle.

[0015] Further, both the classification module and the regression module adopt fully connected networks.

[0016] Preferably, as in the first aspect above, the deep learning model consists of an input layer, a super-resolution reconstruction module, an angle regression model, and an output layer;

[0017] The impedance feature vector is sent into the super-resolution reconstruction module through the input layer. The super-resolution reconstruction module reconstructs the impedance feature vector to increase its dimension into a higher-resolution reconstructed feature vector. Then, the angle regression model performs regression estimation based on the reconstructed feature vector, and finally outputs the bending angle prediction value through the output layer.

[0018] Further, the super-resolution reconstruction module is composed of cascaded several multi-layer perceptrons. First, multiple multi-layer perceptrons perform upsampling layer by layer, and then multiple multi-layer perceptrons perform downsampling layer by layer, and finally output a reconstructed feature vector with a resolution higher than the original input impedance feature vector; the angle regression model adopts a deep neural network.

[0019] As a preferred embodiment of the first aspect, the film bending sensor is composed of a first carbon fiber board, a first adhesive layer, a polyimide film, a second adhesive layer and a second carbon fiber board, and the first carbon fiber board and the second carbon fiber board are both disconnected at the rotation axis position, but each carbon fiber board is staggered and interlocked by edge serrations on both sides of the disconnection position, and a microstructure etching pattern symmetrical with the rotation axis as the center line is formed on the polyimide film by laser etching; a series of metal electrodes are arranged around the edge of the microstructure etching pattern, and each metal electrode is connected to a transfer base for connecting an external circuit through a wire; the microstructure etching pattern (12) and all metal electrodes (9) are completely covered by the carbon nanotube film (11) to form a piezoresistive film sensor.

[0020] As a preference of the first aspect above, the number of metal electrodes arranged around the edge of the microstructure etching pattern is 4 to 32.

[0021] In a second aspect, the present invention provides a robot joint angle recognition device, which includes a film bending sensor and a data processing module;

[0022] The film bending sensor is used to be installed at the joint position of the robot, and the rotating axis is consistent with the rotating axis of the robot joint;

[0023] The data processing module is used to implement the angle recognition method for film bending sensor data enhancement as described in any one of the first aspects above for the film bending sensor.

[0024] In a third aspect, the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, can implement the angle recognition method for film bending sensor data enhancement as described in any one of the first aspects above.

[0025] In a fourth aspect, the present invention provides a computer electronic device comprising a memory and a processor;

[0026] The memory is used to store computer programs;

[0027] The processor is used to implement the angle recognition method for film bending sensor data enhancement as described in any one of the first aspects above when executing the computer program.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The present invention provides an angle recognition method for thin-film bending sensor data enhancement inspired by electrical impedance tomography (EIT). A series of metal electrodes are arranged at the edges of the micro-structured etching patterns to detect impedance values in pairs, thereby enhancing the output signal of the thin-film sensor into a feature vector reflecting the impedance value distribution. Finally, an angle prediction is performed through a deep learning network based on this feature vector. Based on this design, the present invention can more accurately monitor and analyze the changes in physical quantities on the thin film, greatly improving the performance of the bending sensor.

[0030] The present invention also further designs two deep learning networks for the proposed angle recognition method for thin-film bending sensor data enhancement to improve the accuracy of angle recognition. With the deep learning network of the present invention, the complete angle change of the sensor can be accurately estimated, and its accuracy has been greatly improved compared with the pure impedance fitting method. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the steps of the angle recognition method for thin-film bending sensor data enhancement;

[0032] Figure 2 Schematic diagram of the structure of the thin-film bending sensor in the embodiment of the present invention;

[0033] Figure 3 For Figure 2 Schematic diagram of the structure when the polyimide film in the shown thin-film bending sensor does not cover the carbon nanotube film;

[0034] Figure 4 For Figure 3 Schematic diagram of the structure after the polyimide film shown is covered with a carbon nanotube film;

[0035] Figure 5 Schematic diagram of the structure of the computer electronic device in the embodiment of the present invention;

[0036] Figure 6 Schematic diagram of the network structure of the first deep learning model TomoNet in the embodiment of the present invention;

[0037] Figure 7 Variation curve of the loss value (LOSS) with the number of iteration steps (Timestamp) in the embodiment of the present invention;

[0038] Figure 8 Variation curve of the mean square error (MSE) with the number of training epochs (Epoch) in the embodiment of the present invention;

[0039] Figure 9 Variation curve of the accuracy (Accuracy) with the number of training epochs (Epoch) in the embodiment of the present invention;

[0040] Figure 10 It is the structural diagram of the second deep learning model SRNet network in the embodiments of the present invention;

[0041] Figure 11 It is the curve of the change of MSE with the number of training rounds (Epoch) in the experimental group with 8 metal electrodes in the embodiments of the present invention;

[0042] Figure 12 It is the curve of the change of MSE with the number of training rounds (Epoch) in the experimental group with 16 metal electrodes in the embodiments of the present invention;

[0043] Figure 13 It is the curve of the change of MSE with the number of training rounds (Epoch) in the experimental group with 32 metal electrodes in the embodiments of the present invention;

[0044] Figure 14 It is the curve of the change of the error MSE with the number of training rounds (Epoch) in the experimental group with 8 metal electrodes after introducing super-resolution reconstruction.

[0045] The reference numerals in the figure are as follows: polyimide film 1, first carbon fiber board 2, second carbon fiber board 3, rotating shaft 4, microstructural etching pattern 101, metal electrode 102, wire 103, carbon nanotube film 104. Detailed implementation manners

[0046] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined correspondingly without conflict.

[0047] In the description of the present invention, it should be understood that when an element is considered to be "connected" to another element, it can be directly connected to the other element or indirectly connected, that is, there is an intermediate element. On the contrary, when an element is referred to as being "directly" connected to another element, there is no intermediate element.

[0048] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" can explicitly or implicitly include at least one of such features.

[0049] The present invention provides an angle recognition method for data enhancement of a thin-film bending sensor inspired by electrical impedance tomography (EIT). EIT is a method of imaging using differences in resistivity or conductivity. It applies a small current or voltage on the surface of a conductive object and measures the resulting voltage or current changes to reconstruct the resistivity or conductivity distribution inside the object. This technique is particularly suitable for identifying the impedance characteristics of tissues in different physiological or pathological states to obtain images of the internal structure. Based on the principle of EIT, the present invention designs a new data enhancement method for the thin-film bending sensor. A series of metal electrodes are arranged along the edge of the micro-structured etching pattern to detect impedance values pairwise, thereby enhancing the output signal of the thin-film sensor, and an angle prediction is performed based on the output signal through a deep learning network. Based on this design, the present invention can more accurately monitor and analyze the changes in physical quantities on the thin film, greatly improving the performance of the bending sensor. The specific technical solutions of the present invention are described in detail below.

[0050] In a preferred embodiment of the present invention, an angle recognition method for data enhancement of a thin-film bending sensor is provided. The thin-film bending sensor is a piezoresistive thin-film sensor, and a micro-structured etching pattern is symmetrically etched with the axis of rotation as the center line at the bending position of the thin film, and a series of metal electrodes are arranged around the edge of the micro-structured etching pattern. As Figure 1 shown, the angle recognition method includes steps S1 and S2, as follows:

[0051] S1. During the period when the above thin-film bending sensor is at the bending angle to be detected, all the metal electrodes are paired and connected to a voltage-dividing circuit in pairs, so as to measure the impedance value between each pair of metal electrodes under a constant voltage.

[0052] It should be noted that the bending angle to be detected mentioned in the present invention refers to the unknown angle at which the thin-film bending sensor is bent by an external force. Since the present invention needs to detect the impedance value between each pair of metal electrodes pairwise, in order to ensure the accuracy of the detection, the impedance values between all pairs of metal electrodes need to be detected during the maintenance period of the same bending angle to be detected. Therefore, if the thin-film bending sensor is in a fast bending operation condition, the impedance value detection for all pairs of metal electrodes also needs to be completed quickly. All the metal electrodes can be synchronously connected to different voltage-dividing circuits for detection, or can be detected sequentially, which can be specifically designed according to actual needs.

[0053] It should be noted that the form of the above voltage dividing circuit is not limited, as long as it can apply a detection voltage to the metal electrode pair. In addition, for the impedance value between each pair of metal electrodes, it can be achieved through a detection circuit or an external detection device. According to Ohm's law, at a constant voltage, the larger the impedance value, the smaller the corresponding current. Therefore, the impedance value between the metal electrodes can be calculated by detecting the current under a constant voltage.

[0054] S2. Compose an impedance feature vector from the impedance values between all metal electrode pairs and input it into a supervised trained deep learning model to obtain the bending angle prediction value of the thin film bending sensor.

[0055] It should be noted that when composing an impedance feature vector from the impedance values between all metal electrode pairs, a fixed order is required. In actual applications, the order of composing the impedance feature vector needs to be the same as the order of the impedance feature vector in the samples during the supervised training of the deep learning model.

[0056] The form of the above deep learning model is not limited, as long as it can predict the bending angle of the thin film bending sensor based on the impedance feature vector composed of the impedance values between all metal electrode pairs. Before the deep learning model is used for actual prediction, it needs to be supervised trained on a labeled dataset by minimizing the loss function.

[0057] In the embodiments of the present invention, two preferred ways of the deep learning model are provided.

[0058] As the first preferred way of the deep learning model, this deep learning model consists of an input layer, a backbone network, a classification module, a regression module, and an output layer. For the convenience of description, the present invention names the first deep learning model TomoNet.

[0059] In the first deep learning model TomoNet, the input layer is used to subtract the above impedance feature vector from the zero-angle feature vector to obtain a non-zero feature vector, and then splice the above impedance feature vector and the above non-zero feature vector as the input of the backbone network.

[0060] It should be noted that the above zero-angle feature vector is the impedance feature vector obtained when the bending angle of the thin film bending sensor is zero. The impedance values between all metal electrode pairs can be measured according to the method in S1 when the bending angle of the thin film bending sensor is zero, and then these impedance values are composed into an impedance feature vector in a specified order.

[0061] In the first deep learning model TomoNet, the backbone network is composed of multiple layers of deep neural networks (DNN) in cascade, and is used to extract deep features from the input spliced vector and input them into the classification module and the regression module.

[0062] In the first deep learning model, TomoNet, the classification module performs classification prediction with a resolution of 1 degree within the bending angle value range based on the above-mentioned deep features, and outputs the integer part of the bending angle. In an embodiment of the present invention, the above-mentioned classification module can be implemented by a fully connected network, and the classification module can output through a Softmax layer after the fully connected network.

[0063] In the first deep learning model, TomoNet, the regression module performs regression estimation within the range of 0 to 1 degree based on the above-mentioned deep features, and outputs the decimal part of the bending angle. In an embodiment of the present invention, the above-mentioned regression module can also be implemented by a fully connected network.

[0064] In the first deep learning model, TomoNet, the output layer is used to add the integer part and the decimal part of the bending angle, and output the final bending angle.

[0065] As a preferred mode of the second deep learning model, this deep learning model is composed of an input layer, a super-resolution reconstruction module, an angle regression model, and an output layer. For the convenience of description, the present invention names the second deep learning model SRNet. In the second deep learning model SRNet, the above-mentioned impedance feature vector is sent into the super-resolution reconstruction module through the input layer. The super-resolution reconstruction module reconstructs the above-mentioned impedance feature vector to increase its dimension into a higher-resolution reconstructed feature vector. Then, the angle regression model performs regression estimation based on this reconstructed feature vector to predict the bending angle, and finally outputs the bending angle prediction value through the output layer. In an embodiment of the present invention, the super-resolution reconstruction module is composed of several cascaded multi-layer perceptrons. First, the front multi-layer perceptrons perform upsampling layer by layer, and then the remaining multi-layer perceptrons perform downsampling layer by layer, and finally output a reconstructed feature vector with a resolution higher than the original input impedance feature vector. In addition, the above-mentioned angle regression model can adopt a deep neural network DNN.

[0066] The thin-film bending sensor in the present invention can theoretically be any thin-film-based bending sensor, which depends on the impedance change of the microstructure on its thin-film surface during the bending process to detect bending.

[0067] In a preferred embodiment of the present invention, a thin-film bending sensor suitable for implementing the angle recognition method of the present invention is provided. As Figure 2 shown, the thin-film bending sensor is composed of a first carbon fiber board 2, a first bonding layer, a polyimide film 1, a second bonding layer, and a second carbon fiber board 3. Both the first carbon fiber board 2 and the second carbon fiber board 3 are disconnected at the position of the rotating shaft 4, but each carbon fiber board is interlocked by edge serrations on both sides of the disconnected position. As Figure 3As shown, a microstructure etching pattern 101 symmetrical with the rotating shaft 4 as the center line is formed on the polyimide film 1 by laser etching; a series of metal electrodes 102 are arranged around the edge of the microstructure etching pattern 101, and each metal electrode 102 is connected to a transfer base for connecting to an external circuit through a wire 103. Figure 4 As shown, after the microstructure etching pattern 101 and all the metal electrodes 102 are processed on the film, a layer of carbon nanotube film 104 can be further completely covered on the microstructure etching pattern 101 and all the metal electrodes 102 to form a piezoresistive film sensor. The principle of the piezoresistive film sensor is that the microstructure etching pattern 101 made on the polyimide film substrate destroys the original uniformity of the film and forms an anisotropic surface morphology. Therefore, compared with the uniform film without microstructure processing, after the external force is applied to the sensor, the polyimide film processed with the microstructure etching pattern 101 will produce a larger deformation amount with the bending of the rotating shaft 4, and the carbon nanotube film 104 coated and cured on the microstructure etching pattern 101 will also produce cracks with correspondingly greater strength. The resistance between each pair of metal electrodes 102 is directly related to the crack strength on the carbon nanotube film 104 between the two electrodes. Therefore, after the microstructure etching pattern 101 is processed, the sensor can sense a greater resistance change, and the external physical signal sensed will also be greater. Therefore, processing the microstructure etching pattern 101 in the region of the rotating shaft 4 where the polyimide film is deformed to destroy the isotropic state of the film can expand the angle sensing range of the sensor and improve its linearity.

[0068] It should be noted that, no matter the first carbon fiber board 2 or the second carbon fiber board 3, its translational freedom is jointly limited by the staggered interlocking state of the edge serrations and the polyimide film 1. That is, for each carbon fiber board, the two parts of the carbon fiber board that are disconnected along the rotation axis 4 are both bonded and fixed to the polyimide film 1, so they cannot be separated from each other in the direction perpendicular to the rotation axis 4. At the same time, because the two parts are kept in a staggered interlocking state by the edge serrations, they cannot be translated relative to each other in the direction parallel to the rotation axis 4. In the embodiment of the present invention, the first adhesive layer and the second adhesive layer can be made of hot pressing tape, and after the five-layer structure is superimposed, the first carbon fiber board 2 and the second carbon fiber board 3 can be bonded and fixed on both sides of the polyimide film 1 through a hot pressing process.

[0069] It should be noted that the carbon nanotube film 104 can theoretically be any carbon nanotube film with high conductivity and cracks during deformation, and its formula can be adjusted according to actual needs. In the embodiment of the present invention, carbon nanotubes, carbon black and resin binder are mixed to form a homogeneous conductive paste. After the homogeneous conductive paste is prepared, the homogeneous conductive paste is applied to the surface of the substrate where the microstructure etching pattern 101 and the metal electrode 102 are located, and then the whole is heated and cured at a temperature that does not damage the performance of the substrate to form a layer of carbon nanotube film 104. The thickness of the carbon nanotube film 104 should be uniform, and the specific thickness can be optimized and adjusted according to actual conditions, and is generally preferably in the micrometer level.

[0070] The form of the microstructure etching pattern 101 in the present invention is theoretically not limited, as long as it can destroy the isotropic state of the film and amplify the deformation of the polyimide film caused by the bending of the hinge. Figure 3 As shown, the outer contour of the microstructure etching pattern 101 is circular, and a continuous serpentine groove is formed inside by laser ablation.

[0071] It should be noted that the adapter base in the present invention can be any component that can be connected to an external signal acquisition device or circuit. In the embodiment of the present invention, a gold finger-type adapter base can be used in accordance with the practice in the traditional SCM hinge. The gold finger can be used to connect an external voltage divider circuit or detection equipment.

[0072] It should be noted that the number of metal electrodes arranged around the edge of the above-mentioned microstructure etching pattern can be adjusted according to actual needs. Figure 3 The example in FIG. 1 shows a form of four metal electrodes arranged around the edge of the microstructure etching pattern, but the number of arranged metal electrodes can be greater. In an embodiment of the present invention, the number of metal electrodes arranged around the edge of the microstructure etching pattern can preferably be 4 to 32.

[0073] In addition, it should be noted that the angle recognition method for film bending sensor data enhancement shown in S1 and S2 in the above embodiments can essentially be implemented in the form of a computer program.

[0074] Therefore, based on the same inventive concept, another embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the angle recognition method for film bending sensor data enhancement as described above.

[0075] In addition, when the logical instructions in the above-mentioned memory are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention.

[0076] Therefore, based on the same inventive concept, in another embodiment of the present invention, there is provided a computer-readable storage medium corresponding to an angle recognition method for thin-film bending sensor data enhancement. A computer program is stored on the storage medium, and when the computer program is executed by a processor, it can implement the angle recognition method for thin-film bending sensor data enhancement as described above.

[0077] Therefore, based on the same inventive concept, as Figure 5 shown, in another embodiment of the present invention, there is also provided a computer electronic device corresponding to the angle recognition method for thin-film bending sensor data enhancement provided in the above embodiment. It includes a memory and a processor;

[0078] The memory is used to store a computer program;

[0079] The processor is used to implement the angle recognition method for thin-film bending sensor data enhancement as described above when executing the computer program.

[0080] Specifically, in the computer-readable storage media of the above three embodiments, the computer programs stored are executed by the processor, and the steps of S1 and S2 described above can be executed.

[0081] It can be understood that the above storage medium may include a random access memory (RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory. At the same time, the storage medium may also be various media such as a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc that can store program codes.

[0082] It can be understood that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0083] In addition, it should be noted that those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described system can refer to the corresponding process in the foregoing method embodiments, and will not be elaborated herein. In each of the embodiments provided in the present application, the division of steps or modules in the system and method is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0084] In addition, based on Figure 2 、 Figure 3 and Figure 4 the thin-film bending sensor shown, and the angle recognition method for thin-film bending sensor data enhancement shown in the above S1 and S2, a robot joint angle recognition system can be constructed, which includes the above thin-film bending sensor and a data processing module; the thin-film bending sensor is used to be installed at the robot joint position, and the rotation axis is consistent with the rotation axis of the robot joint; and the data processing module is used to implement the angle recognition method for thin-film bending sensor data enhancement as described in the above S1 and S2 steps for the thin-film bending sensor.

[0085] As described above, the above data processing module can be a software function module in the form of a computer program, or a computer electronic device running the software function module.

[0086] In order to better demonstrate the specific implementation manner and technical effects of the angle recognition method for thin-film bending sensor data enhancement shown in the above S1 and S2 of the present invention, the following two embodiments respectively use the above two deep learning models to test the angle recognition performance.

[0087] Embodiment 1

[0088] In this embodiment, a structure as Figure 2 、 Figure 3 andFigure 4 The thin-film bending sensor shown has a specific structure as described above. The number of metal electrodes arranged around the edge of the microstructure etching pattern is set to 16, and the 16 metal electrodes are arranged at equal intervals. According to the above two steps S1 and S2, an angle recognition method for data enhancement of the thin-film bending sensor is constructed. The deep learning model among them adopts the aforementioned first deep learning model TomoNet, which consists of an input layer, a backbone network, a classification module, a regression module, and an output layer. The data processing flow of each network layer and module in the model is as described above and will not be elaborated here. In this embodiment, the network structure of the deep learning model TomoNet is as Figure 6 shown. Among them, the 16 metal electrodes evenly distributed along the circular outer contour are numbered from 0 to 15. Therefore, according to the non-repeating pairwise pairing method, a total of 15 + 14 + 13 + 12 + … + 3 + 2 + 1 = 120-dimensional original impedance data can be obtained. These data are processed by the input layer to form a 240-dimensional concatenated vector as the input of the backbone network. The deep learning model TomoNet converts the angle regression task of the sensor into an angle classification task with a 1-degree resolution and an angle deviation regression task within 1 degree. The following describes the specific data acquisition and network prediction processes in detail.

[0089] When performing data measurement, in this embodiment, the two-terminal method is used to measure the impedance between two electrodes pairwise to reduce the circuit measurement complexity. By performing non-repeating pairwise measurements among the 16 metal electrodes, a total of 120 measurements are required. In this embodiment, two 16-channel channel selectors are used to control the connection of different metal electrode pairs to the voltage division circuit, and then a high-speed ADC is used to calculate the impedance value between two metal electrodes. During the test, a servo motor is used to drive the thin-film bending sensor to rotate, and the hinge bending angle is measured through a visual marker arranged on the side of the rotating shaft of the thin-film bending sensor as the true value of the bending angle.

[0090] In this embodiment, the data acquisition, preprocessing, and model prediction processes are as follows: First, at a fixed bending angle, impedance values between metal electrode pairs are collected to obtain 120-dimensional sensor impedance value data features. After data acquisition, data preprocessing is performed. The 120-dimensional data is normalized, and then the 120-dimensional features at the current angle are subtracted from the 120-dimensional data corresponding to a bending angle of zero. Then, the obtained features are concatenated after the 120-dimensional features at the current angle to generate new 240-dimensional features. Next, a three-layer DNN network is used as the backbone network (the output dimensions of the three layers of the DNN are 1024, 2048, and 2048 respectively) to extract features from the 240-dimensional features. Then, two fully connected networks are used as the classification module and the regression module respectively, and angle classification and regression are performed based on the feature extraction results of the backbone network. In the classification module, after the feature dimensions are sequentially reduced to 1024, 512, and 180 through the fully connected layer, label output is performed through Softmax to achieve classification with a resolution of 1°. The bending angle is classified into 180 category labels corresponding to integer angles from -90 degrees to +90 degrees. Then, in the Maxpred layer, the soft label in the form of a 180-dimensional probability distribution is converted into a hard label, that is, the label with the highest probability is used as the integer part of the bending angle; while the regression module is used to regress the fractional part deviation between the classification value and the true value. In the module, the feature dimensions are sequentially reduced to 1024, 512, 512, and 1 through the fully connected layer, and the final single-dimensional regression value is the fractional part of the bending angle. Finally, the results obtained from the classification module and the deviation module are added together to obtain the final bending angle regression result, which is a 1-dimensional angle value with a decimal part.

[0091] In this embodiment, a dataset of 5 identical thin-film bending sensors rotating from -50° to +50° for 100 cycles is collected. All the data of four thin-film bending sensors and the data of the first 30 cycles of the fifth sensor are used as the training set to participate in the training of the deep learning model, and the data of the last 60 cycles of the fifth sensor are used as the test set to participate in the training. The variation of Loss with the number of iteration steps during the training process is as Figure 7 shown, the variation of the mean square error of the last 100 data with the number of training epochs during the training process is as Figure 8 shown, and the variation of the classification accuracy with the number of training epochs during the training process is as Figure 9 shown.

[0092] Meanwhile, for the purpose of comparison, in this embodiment, a linear fitting was also performed on the single impedance data measured by a single metal electrode pair and the actual bending angle. The final results show that the change in the resistance impedance of only a single metal electrode pair of the sensor cannot fit the change in the accurate sensor angle value. However, by forming features with the impedance values of all electrode pairs according to the methods in S1 and S2 of the present invention, the complete angle change of the sensor can be accurately calculated, and its accuracy has been greatly improved compared with the pure impedance fitting method.

[0093] Embodiment 2

[0094] In this embodiment, a thin-film bending sensor as shown in Figure 2 , Figure 3 and Figure 4 was constructed, and the specific structure is as described above. Among them, three numbers of metal electrodes are arranged around the edge of the microstructure etching pattern, which are 8, 16, and 32 respectively, and all metal electrodes are arranged at equal intervals. According to the above two steps of S1 and S2, an angle recognition method for data enhancement of the thin-film bending sensor was constructed. The deep learning model used the aforementioned second deep learning model SRNet, which consists of an input layer, a super-resolution reconstruction module, an angle regression model, and an output layer. The data processing flow of each network layer and module in the model is as described above and will not be elaborated here. In this embodiment, the network structure of the deep learning model SRNet is as shown in Figure 10 . The figure exemplarily shows the case where 8 metal electrodes are evenly distributed along the circular outer contour. The 8 metal electrodes are numbered from 0 to 7. Therefore, according to the non-repeating pairwise pairing method, a total of 7 + 6 + 5 + 4 + 3 + 2 + 1 = 28-dimensional original impedance data can be obtained. These data are processed by the input layer to form a 28-dimensional impedance feature vector as the input of the backbone network. The same applies to other numbers of electrodes, and only the data dimension needs to be changed. In this embodiment, the super-resolution reconstruction module is composed of several cascaded multi-layer perceptrons MLP. First, the previous multi-layer MLP performs upsampling layer by layer, and then the remaining multi-layer MLP performs downsampling layer by layer, and finally outputs a reconstructed feature vector with a resolution higher than the original input impedance feature vector. In addition, the above angle regression model can be implemented using a deep neural network DNN.

[0095] In addition, in order to demonstrate the role of the super-resolution reconstruction module, a comparison model that removes the super-resolution reconstruction module relative to SRNet was also set up in this embodiment. This model only consists of an input layer, an angle regression model, and an output layer. The impedance feature vector input through the input layer is directly sent to the angle regression model composed of a fully connected network. The bending angle prediction value is obtained by regression estimation by the angle regression model and output by the output layer.

[0096] In this embodiment, three experimental groups for training the comparison model are set up. Thin-film bending sensors with 8, 16, and 32 metal electrodes are arranged around the edge of the micro-structured etching pattern, and impedance value data of different quantities are collected respectively to form impedance feature vectors. At the same time, the hinge bending angle is measured by a visual marker arranged on the side of the rotating shaft of the thin-film bending sensor as the true value of the bending angle. For the thin-film bending sensor with 8 metal electrodes, the formed impedance feature vector is 28-dimensional; for the thin-film bending sensor with 16 metal electrodes, the formed impedance feature vector is 120-dimensional; for the thin-film bending sensor with 32 metal electrodes, the formed impedance feature vector is 496-dimensional. Finally, in the three experimental groups with different numbers of metal electrodes, the mean squared error (MSE) is used as the loss function to train the model. The curves of the MSE corresponding to the experimental groups with 8, 16, and 32 metal electrodes respectively as the number of training rounds changes are shown in Figure 11 , Figure 12 and Figure 13 respectively. It can be seen from the comparison in the figure that obviously, the more metal electrodes there are, the better the performance of the trained model can be obtained.

[0097] However, the more metal electrodes there are, the more wires are arranged on the sensor, the higher the processing difficulty of the whole device is, and the greater the cost is. Therefore, the second deep learning model SRNet is adopted in the present invention. By introducing a super-resolution reconstruction module on the basis of the comparison model, the purpose is to obtain a model prediction performance similar to that with more metal electrodes with fewer metal electrodes.

[0098] Taking the experimental group with 8 metal electrodes as an example, by introducing a super-resolution reconstruction module. Continue to refer to Figure 10As shown, in this embodiment, the super-resolution reconstruction module is composed of a cascade of 5 layers of multi-layer perceptrons MLP. The first two layers of MLP are first up-sampled layer by layer, and then the remaining 3 layers of MLP are down-sampled layer by layer, and finally the output resolution is higher than the reconstructed feature vector of the impedance feature vector of the original input. In this embodiment, the input dimension of the first layer of MLP is 28, the output dimension is 1024, and the output dimensions of the subsequent 4 layers of MLP are 2048, 2048, 1024 and 496 respectively, and the dimension of the reconstructed feature vector of the final input angle regression model is 496 dimensions. Therefore, the model SRNet can upgrade the 28-dimensional impedance feature vector input by the input layer through the super-resolution reconstruction module to a 496-dimensional impedance feature vector that can only be generated by 32 metal electrodes, and send the reconstructed 496-dimensional impedance feature vector to the angle regression model composed of the fully connected network. The angle regression model based on DNN obtains a more accurate prediction value of the bending angle through regression estimation, and outputs it by the output layer. Different from TomoNet, the bending angle prediction value obtained by DNN regression in SRNet is directly the bending angle with both integer and decimal parts.

[0099] It should be noted that the super-resolution reconstruction module needs to be trained separately with actual data in advance for the dimensionality increase of the impedance feature vector. When training the super-resolution reconstruction module, the real 496-dimensional impedance feature vector collected from the experimental group of 32 metal electrodes can be used as the sample label, while the sample input uses the real 28-dimensional impedance feature vector collected from the experimental group of 8 metal electrodes. Through training, the super-resolution reconstruction module can accurately reconstruct high-dimensional impedance feature vectors based on low-dimensional impedance feature vectors to achieve feature dimensionality increase. After training, the super-resolution reconstruction module can be combined into the SRNet model, and then the parameters of the super-resolution reconstruction module can be kept fixed to train other learnable modules in the SRNet model. The aforementioned Figure 9 The MSE curve of the comparison model without super-resolution reconstruction based on the 28-dimensional impedance feature vector of 8 metal electrodes changes with the training rounds. Figure 14 The MSE curve changes with the number of training rounds when the super-resolution reconstruction model SRNet is used to predict the angle based on the 28-dimensional impedance feature vector of 8 metal electrodes. Comparison between the two shows that the present invention can significantly improve the accuracy of full angle prediction by using higher-dimensional data for regression prediction through super-resolution reconstruction.

[0100] The above-described embodiments are only some preferred implementations of the present invention, but are not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. An angle recognition method for data enhancement of a thin-film bending sensor, characterized in that, The thin-film bending sensor is a piezoresistive thin-film sensor, and microstructural etching patterns are symmetrically etched with the axis of rotation as the center line at the bending position of the thin film, and a series of metal electrodes are arranged around the edge of the microstructural etching patterns; The angle recognition method includes: S1. During the period when the thin-film bending sensor is at the bending angle to be detected, all the metal electrodes are paired and connected to a voltage-dividing circuit in pairs, so as to measure the impedance value between each pair of metal electrodes under a constant voltage; S2. The impedance values between all pairs of metal electrodes are formed into an impedance feature vector and input into a supervised-trained deep learning model to obtain the predicted bending angle value of the thin-film bending sensor.

2. The angle recognition method for thin-film bending sensor data enhancement according to claim 1, characterized in that, The deep learning model consists of an input layer, a backbone network, a classification module, a regression module, and an output layer; The input layer is used to subtract the impedance feature vector from the zero-angle feature vector to obtain a non-zero feature vector, and then the impedance feature vector and the non-zero feature vector are spliced and used as the input of the backbone network; the zero-angle feature vector is the impedance feature vector obtained when the thin-film bending sensor is at a bending angle of zero; The backbone network is cascaded by multiple layers of deep neural networks and is used to extract deep features from the input spliced vector and input them into the classification module and the regression module; The classification module performs classification prediction with a resolution of 1 degree within the bending angle value range based on the deep features and outputs the integer part of the bending angle; The regression module performs regression estimation within the range of 0 to 1 degree based on the deep features and outputs the decimal part of the bending angle; The output layer is used to add the integer part and the decimal part of the bending angle and output the final bending angle.

3. The angle recognition method for thin-film bending sensor data enhancement according to claim 2, wherein, Both the classification module and the regression module adopt fully connected networks.

4. The angle recognition method for thin film bending sensor data enhancement according to claim 1, characterized in that The deep learning model consists of an input layer, a super-resolution reconstruction module, an angle regression model, and an output layer; The impedance feature vector is sent into the super-resolution reconstruction module through the input layer, and the super-resolution reconstruction module reconstructs the impedance feature vector to make it dimensionally upgraded into a higher-resolution reconstructed feature vector, and then the angle regression model performs regression estimation based on the reconstructed feature vector, and finally the predicted bending angle value is output through the output layer.

5. The angle recognition method for thin film bending sensor data enhancement according to claim 4, characterized in that, The super-resolution reconstruction module is cascaded by several multi-layer perceptrons. First, multiple multi-layer perceptrons perform upsampling layer by layer, and then multiple multi-layer perceptrons perform downsampling layer by layer, and finally output a reconstructed feature vector with a resolution higher than that of the original input impedance feature vector; The angle regression model adopts a deep neural network.

6. The angle recognition method for thin film bending sensor data enhancement according to claim 1, wherein, The thin film bending sensor is composed of a first carbon fiber board, a first bonding layer, a polyimide thin film, a second bonding layer and a second carbon fiber board. Both the first carbon fiber board and the second carbon fiber board are disconnected at the position of the rotating shaft, but each carbon fiber board is interlocked by edge sawteeth on both sides of the disconnected position. A microstructural etching pattern symmetric about the rotating shaft is formed on the polyimide thin film by laser etching; A series of metal electrodes are arranged around the edge of the microstructural etching pattern, and each metal electrode is connected to a transfer base for connecting an external circuit through a wire; The microstructural etching pattern (12) and all the metal electrodes (9) are completely covered by a carbon nanotube thin film (11) to form a piezoresistive thin film sensor.

7. The angle recognition method for thin film bending sensor data enhancement according to claim 1, characterized in that The number of metal electrodes arranged around the edge of the microstructural etching pattern is 4 to 32.

8. A robot joint angle recognition device, characterized in that, It includes a thin film bending sensor and a data processing module; The thin film bending sensor is used to be installed at the position of a robot joint, and the rotating shaft is consistent with the rotating shaft of the robot joint; The data processing module is used to implement the angle recognition method for data enhancement of the thin film bending sensor as described in any one of claims 1 to 7 for the thin film bending sensor.

9. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by a processor, the angle recognition method for data enhancement of the thin film bending sensor as described in any one of claims 1 to 7 can be implemented.

10. A computer electronic device, characterized in that, It includes a memory and a processor; The memory is used to store a computer program; The processor is used to implement the angle recognition method for data enhancement of the thin film bending sensor as described in any one of claims 1 to 7 when executing the computer program.

Images