Fault diagnosis method for strain type six-dimensional force sensor

Abstract

The application relates to the technical field of sensor fault diagnosis, and specifically discloses a strain type six-dimensional force sensor fault diagnosis method. The method takes six measurement channels of a strain type six-dimensional force sensor as independent diagnosis objects, measures the load resistance value between lead wires in each Wheatstone full-bridge circuit, forms a resistance value combination code after normalization, matches a pre-built fault query table, quickly identifies and locates strain gauge short circuit and / or open circuit faults; when no hard fault is detected, voltage time sequence data of each measurement channel of the strain type six-dimensional force sensor under the action of load is collected and input into a trained one-dimensional residual convolutional neural network model, so that soft fault classification and identification are realized; and finally, the diagnosis results of the channels are summarized to obtain the fault state of the strain gauges of the strain type six-dimensional force sensor. The method can consider both the rapid positioning of hard faults and the high-precision identification of soft faults of the strain gauges of the strain type six-dimensional force sensor.

Description

Technical Field

[0001] This invention relates to the field of sensor fault diagnosis technology, and in particular to a fault diagnosis method for a strain gauge six-dimensional force sensor. Background Technology

[0002] Six-dimensional force sensors are a type of multi-dimensional force sensor capable of simultaneously measuring three-dimensional force and three-dimensional torque information. Their main function is to convert external load signals into detectable and processable electrical signals. There are various types of existing six-dimensional force sensors, mainly including piezoelectric, capacitive, and strain gauge types. Among them, strain gauge six-dimensional force sensors have become an important form of precise force sensing for intelligent equipment due to their mature measurement principles, stable structure, and wide range of applications, finding extensive use in intelligent manufacturing, medical rehabilitation, aerospace, and other fields.

[0003] A strain gauge six-dimensional force sensor typically consists of an elastic body, strain gauges, a Wheatstone bridge measurement circuit, and a signal conditioning unit. Generally, a strain gauge six-dimensional force sensor contains multiple measurement channels, each composed of a Wheatstone bridge circuit with strain gauges. When an external force or torque is applied to the elastic body, the elastic body converts the load into strain. The strain gauges sense this deformation, causing a change in resistance, which is then output as a differential electrical signal via the Wheatstone bridge. This signal is then amplified, filtered, and decoupled to obtain independent measurements of the three-dimensional force and three-dimensional torque. Therefore, the operating state of the measurement circuit directly affects the measurement accuracy and system stability of the strain gauge six-dimensional force sensor.

[0004] In practical applications, strain gauge six-dimensional force sensors often operate in complex environments such as mechanical overload, long-term service, vibration and shock, temperature fluctuations, and even deep sea and deep space. They are easily affected by factors such as mechanical wear, material aging, and environmental disturbances, leading to measurement anomalies or even malfunctions. Common faults in strain gauge six-dimensional force sensors can be categorized into two types: hard faults and soft faults. Hard faults mainly manifest as structural failures such as short circuits or open circuits in the strain gauge; soft faults mainly manifest as performance degradation phenomena such as strain gauge resistance deviation and strain gauge delamination. Due to the compact structure, complex internal bridge circuitry, and difficulty in disassembly and maintenance of six-dimensional force sensors, if faults are not detected and accurately located in a timely manner, measurement results can easily become distorted, and in severe cases, even affect the safe operation of the entire system.

[0005] Traditional methods for detecting bridge circuit faults in strain gauge sensors often involve manual measurement of the bridge impedance, input resistance, or output resistance using multimeters and other testing equipment, followed by comparison with standard values ​​to determine the fault type. While this method is intuitive in principle, it often suffers from drawbacks such as being time-consuming, labor-intensive, inefficient, and lacking automation, especially for strain gauge six-dimensional force sensors with numerous measurement channels and complex wiring. These issues hinder practical engineering applications.

[0006] Existing research has explored fault diagnosis for multidimensional force sensors. For example, the consistency between redundant output data has been used to diagnose faults in six-dimensional force sensors, and it has been verified that these sensors maintain a certain level of measurement accuracy even under faults such as wire breakage and strain gauge stripping. A fault diagnosis method based on constant current source excitation and matrix scanning has been proposed, and a corresponding hardware system has been designed to diagnose short-circuit, open-circuit, and virtual short-circuit / virtual open-circuit faults in strain gauge multidimensional force sensors. In the study of six-dimensional accelerometers, fault identification is achieved by combining fault self-diagnosis models and fault tree analysis. For piezoelectric six-dimensional force sensors, channel fault diagnosis is achieved by establishing a mathematical mapping relationship between the output and the actual force / torque and using residual analysis. Furthermore, some studies have improved the availability of six-dimensional force sensors under fault conditions through redundant structural design, fault-tolerant measurement mechanisms, and intelligent compensation algorithms. For example, some six-dimensional force sensors with redundant branches can maintain a certain measurement capability even when some branches fail; other studies have used mathematical models and intelligent algorithms to achieve signal recovery and fault-tolerant compensation under fault conditions. These methods can play a role in improving system reliability, but their research focuses more on fault tolerance or output reconstruction, rather than on the accurate diagnosis of fault types and locations in strain gauge six-dimensional force sensor bridges.

[0007] In summary, existing technologies for fault diagnosis of strain gauge six-dimensional force sensors still have the following shortcomings: First, traditional manual testing methods have low automation levels, making it difficult to meet the rapid diagnostic needs of multi-channel bridge circuits; second, existing methods based on constant current excitation and matrix scanning mainly target hard faults such as short circuits and open circuits, lacking effective identification capabilities for soft faults such as offset faults and strain gauge delamination faults; third, some model-driven or intelligent algorithm methods focus on output anomaly detection, fault tolerance, or diagnosis of other types of six-dimensional force sensors, and have not formed a unified classification and diagnosis scheme for hard and soft faults applicable to strain gauge six-dimensional force sensors. Summary of the Invention

[0008] Based on this, the purpose of this invention is to provide a fault diagnosis method for a strain gauge six-dimensional force sensor, which focuses on strain gauge faults and can simultaneously achieve rapid location of hard faults and high-precision identification of soft faults in strain gauges.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a fault diagnosis method for a strain gauge six-dimensional force sensor, which includes the following steps: S1. Measure the load resistance between leads in the Wheatstone full-bridge circuit of each measurement channel of the strain gauge six-dimensional force sensor; S2. For each measurement channel, the load resistance value is normalized and numbered. The numbers are arranged in a preset order to form a resistance value combination code. The resistance value combination code is matched with a pre-established fault lookup table to determine whether the measurement channel has a strain gauge short circuit fault and / or open circuit fault. If not, proceed to step S3. If yes, output the diagnosis result of the measurement channel and proceed to step S5. S3. Collect voltage timing data of each measurement channel of the strain-type six-dimensional force sensor under load; S4. Input the voltage timing data of each measurement channel into the trained one-dimensional residual convolutional neural network model, and output the diagnostic result of the measurement channel; S5. Summarize the diagnostic results of all measurement channels of the strain gauge six-dimensional force sensor to obtain the strain gauge fault status of the strain gauge six-dimensional force sensor.

[0010] As a further improvement to the above-mentioned solution of the present invention, the Wheatstone full-bridge circuit is composed of four resistive strain gauges, and leads are respectively led out at the connection nodes between two adjacent resistive strain gauges. In step S1, the load resistance is measured using a multimeter. The red and black probes of the multimeter are connected to different leads to measure the load resistance between multiple leads. For each measurement channel, at least five different lead connection states are set to obtain at least five load resistance values.

[0011] As a further improvement to the above-mentioned solution of the present invention, in step S2, the formula for normalizing the load resistance value is as follows:

[0012] In the formula, R This represents the initial resistance value of the resistance strain gauge. R 测 This represents the load resistance between the leads in the Wheatstone full-bridge circuit as actually measured. The normalized resistance values ​​are numbered using a threshold comparison method: When 0≤ When ≤0.05, let =0, numbered 1; When 0.45≤ When ≤0.55, let =0.5, numbered 2; When 0.65≤ When ≤0.7, let =0.67, numbered 3; When 0.72≤ When ≤0.8, let =0.75, numbered 4; When 0.95≤ When ≤1.05, let =1, numbered 5; When 1.95≤ When ≤2.05, let =2, numbered 6; When 2.95≤ When ≤3.05, let =3, numbered 7; when When ≥3.3, let =∞, numbered 8.

[0013] As a further improvement to the above-mentioned solution of the present invention, in step S4, the diagnostic result is one of the following: the resistance strain gauge is in a normal state, a displacement fault state, a partial debonding fault state, or a complete debonding fault state.

[0014] As a further improvement to the above-described scheme of the present invention, in step S4, the one-dimensional residual convolutional neural network model includes an input layer, a first convolutional layer, a pooling layer, multiple residual convolutional blocks, a Dropout layer, a fully connected layer, a Softmax layer, and an output layer; the first convolutional layer is used to perform preliminary feature mapping on the input voltage time-series data; the pooling layer is used to compress the feature dimension; the multiple residual convolutional blocks are used to perform deep feature extraction; the Dropout layer is used to suppress overfitting; the fully connected layer is used to achieve feature fusion; and the Softmax layer is used to output the probability distribution of each fault category. The one-dimensional residual convolutional neural network model includes at least three residual convolutional blocks arranged sequentially, with the number of convolutional channels in the at least three residual convolutional blocks increasing sequentially; each residual convolutional block includes two convolutional layers and a shortcut connection branch, the first convolutional layer of the residual convolutional block is followed by batch normalization and ReLU activation function, the shortcut connection branch is used to add the block input to the output of the convolutional branch, and then output through the ReLU activation function, the shortcut connection branch adjusts the channel dimension through 1×1 convolution; In the one-dimensional residual convolutional neural network model, each convolutional layer is followed by batch normalization and ReLU activation function, and its kernel size is larger than the kernel size of the convolutional layer in the residual convolutional block.

[0015] As a further improvement to the above-described solution of the present invention, in step S4, the training method of the one-dimensional residual convolutional neural network model is as follows: Based on the working principle of the Wheatstone bridge, output voltage models of the resistive strain gauge under normal state, offset fault state, partial debonding fault state, and complete debonding fault state are established, and labeled datasets are generated based on the output voltage models. The dataset is randomly divided into a training set, a validation set, and a test set. The training set and the validation set are used to train and optimize the parameters of the one-dimensional residual convolutional neural network model. The test set is used to test the trained and optimized one-dimensional residual convolutional neural network model.

[0016] As a further improvement to the above-mentioned scheme of the present invention, the output voltage model in each state is established based on the change of the resistive strain gauge of the Wheatstone bridge under load: Based on the normal Wheatstone bridge output, establish the output voltage model under the normal state; A constant offset is superimposed on the initial resistance value of the resistive strain gauge to establish the output voltage model under the offset fault state; By making the effective strain response coefficient of the faulty resistive strain gauge lower than that of the normal resistive strain gauge, an output voltage model under the partial debonding fault state is established. By making the effective strain response coefficient of the fault resistive strain gauge approach zero, an output voltage model is established for all debonding fault states.

[0017] As a further improvement to the above-described solution of the present invention, the output voltage model under normal conditions is as follows:

[0018] In the formula, where V out For output voltage, V in Input voltage, ΔR This represents the change in resistance of a resistive strain gauge under load. And / or, the output voltage model under the offset fault state is:

[0019] In the formula, , , , These represent the offsets of the resistances of the four resistive strain gauges in the Wheatstone full-bridge circuit. , , , Not all are 0; And / or, the output voltage models for the partial debonding fault state and the complete debonding fault state are as follows:

[0020] In the formula, ΔR 1. ΔR 2. ΔR 3. ΔR4 represents the resistance change of the four resistive strain gauges in the Wheatstone full-bridge circuit under load conditions.

[0021] As a further improvement to the above-mentioned solution of the present invention, the dataset is a six-dimensional time-series voltage dataset. Each group of samples includes the time-series voltage signals of the six output terminals of the strain gauge six-dimensional force sensor during the load change process, and corresponds to one of the following labels: normal state, offset fault state, partial debonding fault state, and complete debonding fault state.

[0022] As a further improvement to the above-mentioned solution of the present invention, during the training process, the cross-entropy loss function is used to quantify the difference between the model prediction result and the true label, and the Adam optimization algorithm is used to update the parameters of the one-dimensional residual convolutional neural network model. After each round of training iterations, the performance of the current model is evaluated using the validation set. When the validation accuracy of the current model is higher than the historical best validation accuracy, the current model parameters are saved as the optimal model parameters.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention treats the six measurement channels of a strain gauge six-dimensional force sensor as independent diagnostic objects. First, the load resistance between multiple leads of the Wheatstone full-bridge circuit in each measurement channel is measured. After normalization, a resistance value combination code is generated and matched with a pre-established fault lookup table to identify short-circuit and / or open-circuit faults in the resistive strain gauge and determine the fault location. When no short-circuit or open-circuit fault is detected, output voltage timing data is collected. A trained one-dimensional residual convolutional neural network model is used to achieve soft fault diagnosis. Finally, the diagnostic results of the six measurement channels are summarized to obtain the fault status of the resistive strain gauge in the strain gauge six-dimensional force sensor; thus, a comprehensive diagnosis of both hard and soft faults in the resistive strain gauge of the strain gauge six-dimensional force sensor is achieved.

[0024] This invention achieves rapid identification of short-circuit faults, open-circuit faults, and their locations in resistive strain gauges by combining lead-to-lead load resistance detection with resistance combination code lookup. It constructs output voltage data models under normal, offset, partial debonding, and complete debonding fault states, and uses a one-dimensional residual convolutional neural network to classify and identify time-series voltage signals, thus achieving effective diagnosis of soft faults in resistive strain gauges. The combination of hard and soft fault diagnosis enables this invention to perform hierarchical identification and comprehensive judgment of multiple types of faults in resistive strain gauges of six-dimensional force sensors, offering advantages such as high diagnostic efficiency, strong applicability, and high accuracy. Attached Figure Description

[0025] Figure 1A flowchart of a fault diagnosis method for a strain gauge six-dimensional force sensor provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the Wheatstone full-bridge circuit structure corresponding to a single measurement channel in an embodiment of the present invention; Figure 3 This is a schematic diagram of the one-dimensional residual convolutional neural network model structure in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the training process of a one-dimensional residual convolutional neural network model in an embodiment of the present invention. Detailed Implementation

[0026] To facilitate understanding of the present invention, a more comprehensive description will be given below with reference to specific embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0028] like Figure 1 As shown, this invention provides a fault diagnosis method for a strain gauge six-dimensional force sensor, combined with... Figure 2 Each measurement channel of the strain gauge six-dimensional force sensor consists of four resistive strain gauges R. i1 R i2 R i3 R i4 To form a Wheatstone full-bridge circuit, R i1 =R i2 =R i3 =R i4 Leads are extended from the connection nodes between any two adjacent resistive strain gauges. When a resistive strain gauge experiences a short circuit or open circuit, the equivalent load resistance changes under different lead connection conditions. When a resistive strain gauge experiences misalignment, partial delamination, or complete delamination, the output voltage timing response under load becomes abnormal. Therefore, this invention addresses the fault diagnosis problem of resistive strain gauges in a six-dimensional force sensor by analyzing each of the six measurement channels as an independent diagnostic object, employing a layered diagnostic approach that first detects hard faults and then identifies soft faults.

[0029] I. Hard Fault Diagnosis Stage of Resistance Strain Gauges First, measure the load resistance between multiple leads of the Wheatstone full-bridge circuit in each measurement channel of the strain gauge six-dimensional force sensor.

[0030] Specifically, the load resistance is measured using a multimeter. The red and black probes of the multimeter are connected to different leads respectively to measure the load resistance between multiple leads. For each measurement channel, at least five different lead-to-lead detection states are set; the method is not unique. Combined with... Figure 2 This embodiment sets five detection states, and the detection order is as follows: Method 1: Connect lead one to the red probe and lead two to the black probe of the multimeter; The second method: Connect lead one to the red pen and lead three to the black pen; The third method: Connect lead one to the red pen and lead four to the black pen; The fourth method: Connect lead two to the red pen and lead three to the black pen; Fifth method: Connect lead four to the red pen, and lead one, lead two, and lead three to the black pen.

[0031] Secondly, for each measurement channel, the load resistance values ​​measured under the five detection states are normalized to obtain the corresponding normalized resistance values, which are then numbered. The normalized resistance values ​​can be normalized using a reference resistance value, calculated using the following formula:

[0032] In the formula, R This represents the initial resistance value of the resistance strain gauge. R 测 R represents the measured load resistance between the leads in the Wheatstone full-bridge circuit. i1 =R i2 =R i3 =R i4 =R.

[0033] To facilitate rapid fault identification, this invention employs a threshold comparison method to discretely number the normalized resistance value. When the normalized resistance value is close to 0, 0.5, 0.67, 0.75, 1, 2, 3, and infinity, it is assigned the numbers 1, 2, 3, 4, 5, 6, 7, and 8 respectively. This yields a five-digit number combination for each measurement channel under five detection states, specifically: When 0≤ When ≤0.05, let =0, numbered 1; When 0.45≤ When ≤0.55, let =0.5, numbered 2; When 0.65≤ When ≤0.7, let =0.67, numbered 3; When 0.72≤ When ≤0.8, let =0.75, numbered 4; When 0.95≤ When ≤1.05, let =1, numbered 5; When 1.95≤ When ≤2.05, let =2, numbered 6; When 2.95≤ When ≤3.05, let =3, numbered 7; when When ≥3.3, let =∞, numbered 8.

[0034] Finally, the five-digit numbers obtained are arranged according to the detection order as resistance value combination codes, and matched with the pre-established fault lookup table to complete the identification of short-circuit or open-circuit faults and the determination of fault location for the corresponding measurement channel resistive strain gauge.

[0035] The fault lookup table maps normalized resistance value combination codes to fault types. Fault types include normal state, single short / open circuit faults of one to four resistance strain gauges, and coexisting short / open circuit faults of one to four resistance strain gauges. The fault lookup table shows the correspondence between the number combination and the fault type of the resistance strain gauge, and can be used to determine the location of the resistance strain gauge that has experienced a short circuit or open circuit fault.

[0036] In this embodiment, the fault query table is shown in Table 1 below.

[0037] Table 1 Fault Inquiry Table

[0038] When the resistance combination code of a certain measurement channel is the same as a resistance combination code in the fault lookup table, it is determined that there is a hard fault type in the measurement channel corresponding to the resistance combination code, and the location of the faulty resistance strain gauge can be determined according to the fault lookup table.

[0039] If a measurement channel is determined to have a short-circuit fault and / or open-circuit fault in its resistance strain gauge, the hard fault diagnosis result of the resistance strain gauge in that measurement channel will be directly output.

[0040] When a measurement channel does not detect a short circuit or open circuit fault in the resistive strain gauge, it enters the soft fault diagnosis stage.

[0041] II. Soft Fault Diagnosis Stage of Resistance Strain Gauges After entering the soft fault diagnosis stage, the timing data of the six-channel output voltage of the strain gauge six-dimensional force sensor under load is first collected as the diagnostic data to be used for subsequent soft fault identification.

[0042] The voltage timing data of each measurement channel is input into a trained one-dimensional residual convolutional neural network model, which outputs the diagnostic result for that measurement channel. The diagnostic result is one of the following: the resistance strain gauge is in a normal state, an offset fault state, a partial debonding fault state, or a complete debonding fault state.

[0043] like Figure 3 As shown, the one-dimensional residual convolutional neural network model 1D-ResNet includes an input layer, a first convolutional layer, a pooling layer, multiple residual convolutional blocks, a Dropout layer, a fully connected layer, a Softmax layer, and an output layer. The first convolutional layer performs initial feature mapping on the input voltage time-series data; the pooling layer compresses the feature dimension. Multiple residual convolutional blocks are used for deep feature extraction. The Dropout layer suppresses overfitting. The fully connected layer achieves feature fusion. The Softmax layer outputs the probability distribution of each fault category.

[0044] The one-dimensional residual convolutional neural network model consists of three progressively deeper residual convolutional blocks: Residual Convolutional Block 1, Residual Convolutional Block 2, and Residual Convolutional Block 3. The number of convolutional channels in the three residual convolutional blocks increases sequentially to achieve hierarchical extraction from local to global features. Each residual convolutional block includes two convolutional layers and a shortcut connection branch. The convolutional layers in Residual Convolutional Block 1 are Convolutional Layers 2 and 3; those in Residual Convolutional Block 2 are Convolutional Layers 4 and 5; and those in Residual Convolutional Block 3 are Convolutional Layers 6 and 7. The first convolutional layer of each residual convolutional block is followed by batch normalization and a ReLU activation function. The shortcut connection branch adds the block input to the output of the convolutional branch, and then outputs the result after passing it through the ReLU activation function. When the dimensions of the residual path and the main path are inconsistent, the shortcut connection branch adjusts the channel dimensions through a 1×1 convolution to maintain consistency between the dimensions of the residual path and the main path.

[0045] In a one-dimensional residual convolutional neural network model, batch normalization and ReLU activation function are applied after convolutional layer one. The kernel size of convolutional layer one is larger than the kernel size of all convolutional layers in the residual convolutional block to enhance the ability to perceive the initial features of the original time-series signal.

[0046] The training method for the one-dimensional residual convolutional neural network model in this embodiment is as follows: (1) Based on the working principle of Wheatstone bridge, output voltage models of resistive strain gauges under normal state, offset fault state, partial debonding fault state and complete debonding fault state are established, and labeled datasets are generated based on each output voltage model.

[0047] The output voltage models for each state are established based on the changes in the resistive strain gauges of the Wheatstone bridge under load: Based on the normal output of the Wheatstone bridge, establish the output voltage model under normal conditions; By superimposing a constant offset on the initial resistance value of the resistive strain gauge, an output voltage model under offset fault conditions is established. By making the effective strain response coefficient of the faulty resistive strain gauge lower than that of the normal resistive strain gauge, an output voltage model under partial debonding fault conditions is established. By making the effective strain response coefficient of the fault resistive strain gauge approach zero, an output voltage model is established under all debonding fault conditions.

[0048] The modeling of partial and complete debonding failure states of resistive strain gauges was carried out under load conditions to characterize the strain sensing attenuation or failure caused by debonding of the resistive strain gauges.

[0049] The output voltage model under normal conditions is:

[0050] In the formula, where V out For output voltage, V in Input voltage, R The initial resistance value of the resistance strain gauge. ΔR This represents the change in resistance of a resistive strain gauge under load. The output voltage model under offset fault conditions is as follows:

[0051] In the formula, , , , These represent the offsets of the resistances of the four resistive strain gauges in the Wheatstone full-bridge circuit. , , , Not all are 0; The output voltage models for partial debonding and complete debonding fault states are as follows:

[0052] In the formula, ΔR 1. ΔR 2. ΔR 3. ΔR 4 represents the resistance changes of the four resistive strain gauges in the Wheatstone full-bridge circuit under load conditions.

[0053] To make the established output voltage model closer to the actual measurement environment, a few millivolts can be added to the output voltage model to simulate acquisition noise and environmental disturbances, thereby improving the robustness and generalization ability of the fault diagnosis model.

[0054] Based on the output voltage models under the above fault states, a six-dimensional time-series voltage dataset is constructed. Each sample includes the time-series voltage signals from the six output terminals of a strain gauge six-dimensional force sensor during load changes, and is assigned a label from one of the following states: normal state, offset fault state, partial debonding fault state, and complete debonding fault state.

[0055] (2) The dataset is randomly divided into training set, validation set and test set. The training set and validation set are used to train and optimize the parameters of the one-dimensional residual convolutional neural network model. The test set is used to test the trained and optimized one-dimensional residual convolutional neural network model.

[0056] In this embodiment, the dataset is divided into a training set, a validation set, and a test set in a ratio of 60:25:15. Figure 4 This document describes the entire training process of the one-dimensional residual convolutional neural network model in this embodiment. During model training, the cross-entropy loss function is used to quantify the difference between the model's prediction results and the true labels, and the Adam optimization algorithm is used to update the parameters of the one-dimensional residual convolutional neural network model.

[0057] After each round of training iterations, the performance of the current model is evaluated using the validation set. When the validation accuracy of the current model is higher than the historical best validation accuracy, the current model parameters are saved as the optimal model parameters.

[0058] During the diagnostic phase, the optimal model parameters saved during training are loaded, and the timing data of the six-channel output voltage to be diagnosed are input into a one-dimensional residual convolutional neural network model. After processing by the Softmax layer, the fault category is output, thereby obtaining the soft fault diagnosis result of the corresponding measurement channel.

[0059] Third, the diagnostic results of the six measurement channels are summarized to obtain the fault status of the resistive strain gauge of the strain-type six-dimensional force sensor.

[0060] By comprehensively analyzing the results from the six measurement channels, a final diagnostic result for the fault state of the strain gauge six-dimensional force sensor can be obtained.

[0061] This invention achieves rapid identification of short-circuit faults, open-circuit faults, and their locations in resistive strain gauges of a six-dimensional force sensor by combining lead-to-lead load resistance detection with resistance combination code lookup. By constructing output voltage models under various soft fault states and combining them with a one-dimensional residual convolutional neural network, it achieves effective diagnosis of offset faults, partial delamination faults, and complete delamination faults, thus forming a comprehensive fault diagnosis method that combines hard and soft faults suitable for resistive strain gauges of a six-dimensional force sensor.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A fault diagnosis method for a strain gauge six-dimensional force sensor, characterized in that, It includes the following steps: S1. Measure the load resistance between leads in the Wheatstone full-bridge circuit of each measurement channel of the strain gauge six-dimensional force sensor; S2. For each measurement channel, the load resistance value is normalized and numbered. The numbers are arranged in a preset order to form a resistance value combination code. The resistance value combination code is matched with a pre-established fault lookup table to determine whether the measurement channel has a strain gauge short circuit fault and / or open circuit fault. If not, proceed to step S3. If yes, output the diagnosis result of the measurement channel and proceed to step S5. S3. Collect voltage timing data of each measurement channel of the strain-type six-dimensional force sensor under load; S4. Input the voltage timing data of each measurement channel into the trained one-dimensional residual convolutional neural network model, and output the diagnostic result of the measurement channel; the diagnostic result is one of the following: the resistive strain gauge is in a normal state, an offset fault state, a partial debonding fault state, or a complete debonding fault state; The training method for the one-dimensional residual convolutional neural network model is as follows: Based on the working principle of the Wheatstone bridge, output voltage models of the resistive strain gauge under normal state, offset fault state, partial debonding fault state, and complete debonding fault state are established, and labeled datasets are generated based on the output voltage models. The dataset is randomly divided into a training set, a validation set, and a test set. The training set and the validation set are used to train and optimize the parameters of the one-dimensional residual convolutional neural network model. The test set is used to test the trained and optimized one-dimensional residual convolutional neural network model. The output voltage models for each state are established based on the changes in the resistive strain gauges of the Wheatstone bridge under load: Based on the normal Wheatstone bridge output, establish the output voltage model under normal conditions; the output voltage model under normal conditions is as follows: In the formula, R This represents the initial resistance value of the resistance strain gauge, where V out For output voltage, V in Input voltage, ΔR This represents the change in resistance of a resistive strain gauge under load. By superimposing a constant offset on the initial resistance value of the resistive strain gauge, an output voltage model under the offset fault state is established; the output voltage model under the offset fault state is as follows: In the formula, , , , These represent the offsets of the resistances of the four resistive strain gauges in the Wheatstone full-bridge circuit. , , , Not all are 0; By making the effective strain response coefficient of the faulty resistive strain gauge lower than that of the normal resistive strain gauge, an output voltage model under the partial debonding fault state is established. By making the effective strain response coefficient of the fault resistive strain gauge approach zero, an output voltage model is established for all debonding fault states. The output voltage models for the partial debonding fault state and the complete debonding fault state are as follows: In the formula, ΔR 1. ΔR 2. ΔR 3. ΔR 4 represents the resistance change of the four resistive strain gauges in the Wheatstone full-bridge circuit under load conditions; S5. Summarize the diagnostic results of all measurement channels of the strain gauge six-dimensional force sensor to obtain the strain gauge fault status of the strain gauge six-dimensional force sensor.

2. The fault diagnosis method for a strain gauge six-dimensional force sensor according to claim 1, characterized in that, The Wheatstone full-bridge circuit consists of four resistive strain gauges, with leads extending from the connection nodes between two adjacent resistive strain gauges. In step S1, the load resistance is measured using a multimeter. The red and black probes of the multimeter are connected to different leads to measure the load resistance between multiple leads. For each measurement channel, at least five different lead connection states are set to obtain at least five load resistance values.

3. The fault diagnosis method for a strain gauge six-dimensional force sensor according to claim 2, characterized in that, In step S2, the formula for normalizing the load resistance value is: In the formula, R This represents the initial resistance value of the resistance strain gauge. R 测 This represents the load resistance between the leads in the Wheatstone full-bridge circuit as actually measured. The normalized resistance values ​​are numbered using a threshold comparison method: When 0≤ When ≤0.05, let =0, numbered 1; When 0.45≤ When ≤0.55, let =0.5, numbered 2; When 0.65≤ When ≤0.7, let =0.67, numbered 3; When 0.72≤ When ≤0.8, let =0.75, numbered 4; When 0.95≤ When ≤1.05, let =1, numbered 5; When 1.95≤ When ≤2.05, let =2, numbered 6; When 2.95≤ When ≤3.05, let =3, numbered 7; when When ≥3.3, let =∞, numbered 8.

4. The fault diagnosis method for a strain gauge six-dimensional force sensor according to claim 1, characterized in that, In step S4, the one-dimensional residual convolutional neural network model includes an input layer, a first convolutional layer, a pooling layer, multiple residual convolutional blocks, a Dropout layer, a fully connected layer, a Softmax layer, and an output layer; the first convolutional layer is used to perform preliminary feature mapping on the input voltage time series data; the pooling layer is used to compress the feature dimension; Multiple residual convolutional blocks are used for deep feature extraction; the Dropout layer is used to suppress overfitting; the fully connected layer is used to achieve feature fusion; and the Softmax layer is used to output the probability distribution of each fault category. The one-dimensional residual convolutional neural network model includes three residual convolutional blocks arranged sequentially, with the number of convolutional channels in the three residual convolutional blocks increasing sequentially; each residual convolutional block includes two convolutional layers and a shortcut connection branch, the first convolutional layer of the residual convolutional block is followed by batch normalization and ReLU activation function, the shortcut connection branch is used to add the block input to the output of the convolutional branch, and then output through the ReLU activation function, the shortcut connection branch adjusts the channel dimension through 1×1 convolution; In the one-dimensional residual convolutional neural network model, each convolutional layer is followed by batch normalization and ReLU activation function, and its kernel size is larger than the kernel size of the convolutional layer in the residual convolutional block.

5. The fault diagnosis method for a strain gauge six-dimensional force sensor according to claim 1, characterized in that, The dataset is a six-dimensional time-series voltage dataset. Each sample includes the time-series voltage signals from the six output terminals of a strain gauge six-dimensional force sensor during load changes, and corresponds to one of the following labels: normal state, offset fault state, partial debonding fault state, and complete debonding fault state.

6. The fault diagnosis method for a strain gauge six-dimensional force sensor according to claim 1, characterized in that, During the training process, the cross-entropy loss function is used to quantify the difference between the model prediction results and the true labels, and the Adam optimization algorithm is used to update the parameters of the one-dimensional residual convolutional neural network model. After each round of training iterations, the performance of the current model is evaluated using the validation set. When the validation accuracy of the current model is higher than the historical best validation accuracy, the current model parameters are saved as the optimal model parameters.

Images