Rapid calibration method for six-dimensional force sensor based on transfer learning
By constructing a model knowledge base and fingerprint feature matching, a transfer learning method is used to calibrate a six-dimensional force sensor, which solves the problem of low efficiency in traditional calibration methods and achieves an efficient and accurate calibration process.
Patent Information
- Application Number
- CN202511666347.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-14
- Publication Date
- 2026-02-13
AI Technical Summary
Traditional six-dimensional force sensor calibration methods are extremely inefficient and rely on expensive six-dimensional combined loading equipment, which limits the expansion of calibration capabilities and the efficiency of the production process.
A rapid calibration method based on transfer learning is adopted. By constructing a model knowledge base, the decoupled calibration model is trained using the fingerprint features of the sample's six-dimensional force sensor, and rapid calibration is performed through similarity matching and hierarchical transfer learning.
It significantly improves calibration efficiency, reduces reliance on high-cost equipment, shortens calibration time, ensures calibration accuracy, and has scalability and adaptability, achieving a balance between speed and accuracy.
Smart Images
Figure CN121521353A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of six-dimensional force sensors, in particular to a six-dimensional force sensor rapid calibration method based on transfer learning. BACKGROUND
[0002] The six-dimensional force sensor is a core component for realizing high-precision force control in the fields of robots, precision assembly, aerospace, etc., which can simultaneously measure three forces (Fx, Fy, Fz) and three moments (Mx, My, Mz) in space. The accuracy and reliability of the sensor are highly dependent on accurate calibration, i.e., establishing an accurate mapping relationship between its output voltage and the force / moment.
[0003] Currently, the calibration of the six-dimensional force sensor mainly uses the traditional least squares method. The core process of this method is: for each individual sensor, a high-precision six-dimensional combined loading device must be used to apply hundreds of different sizes and types of known forces and moments (including single-dimensional load and composite load) on its six dimensions, and the corresponding output voltage is synchronously collected to construct a complete calibration training data set; then, based on the data set, a fixed decoupling matrix is solved by the least squares fitting algorithm.
[0004] However, this traditional calibration method has a fundamental technical defect: the calibration efficiency is extremely low. This core problem is embodied in the following aspects: first, the process itself is extremely tedious and time-consuming, as hundreds of high-quality load-voltage data points must be collected for each sensor, the calibration period of a single sensor usually lasts for several days, becoming a bottleneck in the production process; second, this method is highly dependent on special and expensive six-dimensional combined loading equipment, and the scarcity and high cost of such equipment further limit the expansion of calibration capacity.
[0005] Therefore, there is an urgent need in the industry for a new method that can significantly improve the calibration efficiency, significantly reducing the amount of data, time and hardware dependence required for the calibration of a single sensor while ensuring calibration accuracy. SUMMARY
[0006] In view of the above part of the defects of the prior art, the technical problem to be solved by the present application is to provide a six-dimensional force sensor rapid calibration method based on transfer learning, aiming to improve the calibration efficiency of the six-dimensional force sensor while ensuring the calibration accuracy.
[0007] To achieve the above-mentioned purpose, the present application provides a six-dimensional force sensor rapid calibration method based on transfer learning, which comprises:
[0008] Step S1, obtaining N sample six-dimensional force sensors; applying a plurality of known forces or torques to the six dimensions of the sample six-dimensional force sensor, and collecting the corresponding output voltage to construct a calibration training data set for training the decoupling calibration model; collecting sample fingerprint characteristics of the sample six-dimensional force sensor; wherein N is an integer greater than 2, the calibration training data set includes forces and their corresponding output voltages or torques and their corresponding output voltages, and the fingerprint characteristics include at least one of zero position characteristic fingerprint, sensitivity and coupling fingerprint, dynamic response fingerprint, micro load response fingerprint, and temperature sensitivity fingerprint;
[0009] Step S2, training a corresponding decoupling calibration model according to the calibration training data set of each sample six-dimensional force sensor; establishing a mapping relationship between the decoupling calibration models of all sample six-dimensional force sensors and the sample fingerprint characteristics to construct a model knowledge base; wherein the decoupling calibration model is used to output the corresponding force or torque according to the output voltage;
[0010] Step S3, collecting a to-be-calibrated fingerprint characteristic of a to-be-calibrated six-dimensional force sensor; performing similarity matching between the to-be-calibrated fingerprint characteristic and each sample fingerprint characteristic in the model knowledge base to obtain a target sample fingerprint characteristic with the highest similarity and a corresponding target decoupling calibration model; wherein the to-be-calibrated six-dimensional force sensor is of the same model as the sample six-dimensional force sensor;
[0011] Step S4, using the target decoupling calibration model as a pre-training model, adjusting the pre-training model based on the to-be-calibrated fingerprint characteristic to obtain a final calibration model special for the to-be-calibrated six-dimensional force sensor.
[0012] Optionally, the step S1 of collecting the sample fingerprint characteristics of the sample six-dimensional force sensor comprises:
[0013] Under the condition that the sample six-dimensional force sensor does not bear external force, the voltage output of each channel within a first preset time period is continuously collected to obtain the zero position characteristic fingerprint of the sample six-dimensional force sensor;
[0014] A force or torque of a preset size is sequentially applied to the six dimensions of the sample six-dimensional force sensor, and the voltage output during the process is recorded; the ratio of the voltage output of each dimension to the applied force or torque when the corresponding force is applied is extracted to obtain the main sensitivity; the voltage change of other dimensions when the corresponding force is applied is extracted to obtain the cross-coupling coefficient; and the sensitivity and coupling fingerprint is obtained according to the main sensitivity and the cross-coupling coefficient;
[0015] When the sample six-dimensional force sensor is subjected to or removed from the load, sampling is performed for a second preset time period, and the entire process from voltage change to stability is recorded to obtain the dynamic response fingerprint;
[0016] applying forces or torques from small to large gradually to the six dimensions of the sample six-dimensional force sensor in sequence, recording the threshold of force or torque corresponding to the first deviation of the voltage output of each channel from its zero value, to obtain the micro-load response fingerprint;
[0017] placing the sample six-dimensional force sensor in a temperature-controllable environment, controlling the temperature change, recording the relationship between the temperature change and the voltage output, and obtaining the temperature sensitivity fingerprint.
[0018] Optionally, in the step S3, the method further comprises:
[0019] in response to the similarity of the to-be-calibrated fingerprint feature and each sample fingerprint feature in the model knowledge base being lower than a first similarity threshold, applying a plurality of known forces or torques to the to-be-calibrated six-dimensional force sensor, and collecting the corresponding output voltage to obtain a corresponding new training data set;
[0020] training a corresponding new decoupling calibration model according to the corresponding new training data set of the to-be-calibrated six-dimensional force sensor; establishing a mapping relationship between the to-be-calibrated fingerprint feature and the new decoupling calibration model, and adding it to the model knowledge base.
[0021] Optionally, the similarity matching in the step S3 is realized by a weighted comprehensive similarity algorithm, different weight coefficients are configured for different types of fingerprint features, the weight of the sensitivity and coupling fingerprint is configured to be the highest, the weight of the dynamic response fingerprint and the zero characteristic fingerprint is configured to be the second, and the weight of the micro-load response fingerprint and the temperature sensitivity fingerprint is configured to be the lowest.
[0022] Optionally, the adjustment of the pre-trained model in the step S4 is hierarchical transfer learning, comprising:
[0023] freezing the basic feature extraction layer in the pre-trained model representing general physical laws;
[0024] based on the to-be-calibrated fingerprint feature, adjusting the parameters of the intermediate adaptation layer in the pre-trained model responsible for environmental adaptability;
[0025] based on the to-be-calibrated fingerprint feature, retraining the output layer in the pre-trained model responsible for individualized calibration.
[0026] Optionally, after the step S4, further comprising:
[0027] associating the to-be-calibrated fingerprint feature with the final calibration model and adding it to the model knowledge base.
[0028] Optionally, the calibration training data set constructed in step S1 contains more load types and quantities than the load types and quantities applied when collecting the sample fingerprint features.
[0029] Optionally, before performing similarity matching between the to-be-calibrated fingerprint features and each sample fingerprint feature in the model knowledge base, a step of normalizing and preprocessing the to-be-calibrated fingerprint features is further included.
[0030] Advantages of the present application: 1. The present application changes the calibration paradigm of six-dimensional force sensors by constructing a "model knowledge base", and changes the calibration process of the to-be-calibrated sensor from the traditional "training from scratch" to "fine-tuning based on prior knowledge". For most to-be-calibrated sensors, instead of hundreds or thousands of multi-dimensional and composite loads, only a small amount of key "fingerprint features" need to be collected to complete the adaptation of the high-precision model. This greatly shortens the calibration time of a single sensor and improves the efficiency by more than an order of magnitude, greatly meeting the needs of large-scale production and rapid on-site maintenance. 2. Since full-load calibration of each sensor is no longer required, the present application significantly reduces the dependence on and occupation time of high-precision, high-cost six-dimensional joint loading equipment. 3. The core of the present application is not to sacrifice accuracy for speed. On the contrary, by migrating the "baseline decoupling calibration model" trained from a large amount of data, the to-be-calibrated sensor has a high-precision model basis at the start. The subsequent fine-tuning process is a "personalized" adaptation of the high-precision model based on its own physical characteristics, rather than learning from scratch. Therefore, the "final calibration model" obtained ultimately can approach the level of traditional full calibration in terms of accuracy, achieving the unity of speed and accuracy. 4. The present application has natural scalability and adaptability. First of all, the "model knowledge base" is a system that can evolve continuously, and as more sensor data is accumulated, the coverage of the knowledge base and the accuracy of model matching will become higher and higher.
[0031] In summary, the present application fundamentally changes the calibration paradigm of six-dimensional force sensors, successfully solves the bottleneck of traditional methods in efficiency, cost and flexibility by introducing transfer learning and fingerprint feature matching mechanism, and lays a solid technical foundation for large-scale calibration of six-dimensional force sensors. BRIEF DESCRIPTION OF DRAWINGS
[0032] Figure 1 is a flowchart of a six-dimensional force sensor rapid calibration method based on transfer learning provided by an embodiment of the present application;
[0033] Figure 2 is a work flow diagram of a six-dimensional force sensor rapid calibration method based on transfer learning in a specific application of the present application. DETAILED DESCRIPTION
[0034] The application discloses a six-dimensional force sensor rapid calibration method based on transfer learning, and technical personnel in the field can learn from the content and appropriately improve technical details to realize it. It is particularly pointed out that all similar substitutions and changes are obvious to technical personnel in the field, and they are regarded as being included in the application. The method and application of the application have been described through the preferred embodiments, and relevant personnel can obviously change or appropriately change and combine the method and application described herein without departing from the content, spirit and scope of the application to realize and apply the technical field of the application.
[0035] The applicant found that the traditional calibration method has a fundamental technical defect: the calibration efficiency is extremely low. This core problem is embodied in the following aspects: first, the process itself is extremely tedious and time-consuming, because hundreds of high-quality load-voltage data points must be collected for each sensor, the calibration period of a single sensor usually lasts for several days, which becomes a bottleneck in the production process; second, the method is severely dependent on special and expensive six-dimensional joint loading equipment, and the scarcity and high cost of such equipment further limit the expansion of calibration capacity.
[0036] Therefore, the embodiment of the application provides a six-dimensional force sensor rapid calibration method based on transfer learning, as shown in the method, which comprises the following steps: Figure 1
[0037] Step S1, obtaining N sample six-dimensional force sensors; applying a plurality of known forces or torques to the six dimensions of the sample six-dimensional force sensors, and collecting the corresponding output voltages to construct a calibration training data set for training a decoupling calibration model; and collecting sample fingerprint characteristics of the sample six-dimensional force sensors.
[0038] Wherein, N is an integer greater than 2, the calibration training data set includes forces and their corresponding output voltages or torques and their corresponding output voltages, and the fingerprint characteristics include at least one of zero position characteristic fingerprint, sensitivity and coupling fingerprint, dynamic response fingerprint, micro load response fingerprint and temperature sensitivity fingerprint.
[0039] It should be noted that step S1 is the basis for all subsequent "rapid" calibration. By calibrating a small number of representative samples, the initial model in the knowledge base is ensured to be highly accurate and reliable. The "fingerprint characteristics" associate a complex high-precision model with a set of simplified data that is easy to measure and characterize individual characteristics, thereby creating a rapid indexing mechanism for subsequent calibration.
[0040] In this specific embodiment, the collection of sample fingerprint characteristics of the sample six-dimensional force sensors in step S1 comprises:
[0041] Under the condition that the sample six-dimensional force sensor does not bear external force, the voltage output of each channel in a first preset time duration is continuously collected to obtain a zero characteristic fingerprint of the sample six-dimensional force sensor;
[0042] A preset size of force or torque is sequentially applied to six dimensions of the sample six-dimensional force sensor, and the voltage output during the process is recorded; the ratio of the voltage output of each dimension to the applied force or torque when the corresponding force is applied is extracted to obtain a main sensitivity; the voltage change of other dimensions when the corresponding force is applied is extracted to obtain a cross-coupling coefficient; and the sensitivity and coupling fingerprint are obtained according to the main sensitivity and the cross-coupling coefficient;
[0043] When the sample six-dimensional force sensor is applied or unloaded, sampling is performed for a second preset time duration, and the entire process from voltage change to stability is recorded to obtain a dynamic response fingerprint;
[0044] A gradually changing force or torque from small to large is sequentially applied to the six dimensions of the sample six-dimensional force sensor, and the threshold value of the force or torque corresponding to the voltage output of each channel first deviating from its zero value is recorded to obtain a micro-load response fingerprint;
[0045] The sample six-dimensional force sensor is placed in a temperature controllable environment, the temperature change is controlled, the relationship between the temperature change and the voltage output is recorded, and a temperature sensitivity fingerprint is obtained.
[0046] It should be noted that the zero characteristic fingerprint establishes a reference point for measurement, ensuring the initial accuracy of the system; the sensitivity and coupling fingerprint accurately describes the core linear mapping relationship and main interference source of the sensor input and output, which is the basis for decoupling; the dynamic response fingerprint reveals the intrinsic characteristics (such as damping and natural frequency) of the mechanical structure of the sensor, ensuring the reliability of the model under dynamic working conditions; the micro-load response fingerprint defines the effective measurement range and resolution limit of the sensor, improving the accuracy of small signal measurement; and the temperature sensitivity fingerprint quantifies the influence of environmental temperature on the output, providing key environmental adaptive compensation ability for the model. These characteristics, with the least test data, represent the key performance parameters of the sensor in static, dynamic, linear, nonlinear and environmental adaptability to the greatest extent, laying a solid data foundation for subsequent precise matching and efficient fine-tuning.
[0047] In this specific embodiment, the calibration training data set constructed in step S1 contains more types and quantities of loads than the types and quantities of loads applied when the sample fingerprint characteristics are collected.
[0048] It should be noted that in this way, the workload of subsequent rapid calibration is less than that of full calibration of the sample six-dimensional force sensor, and the full calibration is more accurate.
[0049] Step S2, according to the calibration training data set of each sample six-dimensional force sensor, a corresponding decoupling calibration model is trained; a mapping relationship between the decoupling calibration model of all sample six-dimensional force sensors and the sample fingerprint characteristics is established, and a model knowledge base is constructed.
[0050] The decoupling calibration model is used to output corresponding force or torque according to the output voltage.
[0051] It should be noted that the calibration knowledge (i.e. the complex mapping relationship from input voltage to output force / torque) obtained through a long time is encapsulated in a trained decoupling calibration model. The mapping relationship between the model and the fingerprint is to systematically organize the scattered model into a queryable and matchable "knowledge base". This realizes the leap from "having multiple independent models" to "having an intelligent retrieval system", which is the core architecture for realizing fast matching and migration.
[0052] Step S3, the fingerprint characteristics of the to-be-calibrated six-dimensional force sensor are collected; the to-be-calibrated fingerprint characteristics are matched with each sample fingerprint characteristics in the model knowledge base, and the target sample fingerprint characteristics with the highest similarity and the corresponding target decoupling calibration model are obtained.
[0053] The to-be-calibrated six-dimensional force sensor is of the same type as the sample six-dimensional force sensor.
[0054] It should be noted that step S3 realizes fast diagnosis: for the to-be-calibrated sensor, full calibration is not required, only fast and low-cost fingerprint collection is required. This operation greatly saves time and equipment cost, and is the first key point for realizing fast calibration. And intelligent matching: through the similarity matching algorithm, a benchmark model with the most matched teaching method and own characteristics is automatically found for the current to-be-calibrated sensor. This process fully utilizes historical data and knowledge for fast calibration.
[0055] In this specific embodiment, in step S3, the method further includes:
[0056] In response to the similarity between the to-be-calibrated fingerprint characteristics and each sample fingerprint characteristics in the model knowledge base being lower than a first similarity threshold, a plurality of known forces or torques are applied to the to-be-calibrated six-dimensional force sensor, and corresponding output voltages are collected to obtain corresponding new training data sets;
[0057] According to the new training data set corresponding to the to-be-calibrated six-dimensional force sensor, a corresponding new decoupling calibration model is trained; a mapping relationship between the to-be-calibrated fingerprint characteristics and the new decoupling calibration model is established, and is added to the model knowledge base.
[0058] It should be noted that the role of this embodiment is to introduce a closed-loop feedback and self-evolution mechanism for the entire rapid calibration system, and its significance lies in that when a "strange" sensor that is too different from all the benchmark models in the knowledge base appears, the system will not fail or provide low-precision results, but can automatically start the backup plan, establish an accurate model for the sensor through a one-time traditional full-calibration process, and store the "fingerprint-model" mapping relationship as new knowledge in the knowledge base. This not only guarantees the final reliability of a single calibration task, but also realizes the continuous expansion and self-optimization of the knowledge base, so that the system can continuously adapt to new batches of sensors or variant individuals, thereby forming a virtuous cycle of becoming more intelligent, and fundamentally enhancing the robustness and long-term applicability of the method.
[0059] In a specific application, after adding step S3, the workflow of the embodiment of the present application can be as shown in Figure 2
[0060] In this specific embodiment, the similarity matching in step S3 is realized by a weighted comprehensive similarity algorithm, different weight coefficients are configured for different types of fingerprint characteristics, the weight of the sensitivity and coupling fingerprint is configured to be the highest, the weight of the dynamic response fingerprint and the zero characteristic fingerprint is in the second place, and the weight of the micro-load response fingerprint and the temperature sensitivity fingerprint is the lowest.
[0061] It should be noted that this embodiment configures different weight coefficients for different types of fingerprint characteristics, and explicitly sets the weight of the sensitivity and coupling fingerprint to be the highest. This scheme guides the matching algorithm to focus on the most core and essential linear mapping relationship and coupling interference in force measurement, thereby ensuring that the selected basic model for the sensor to be calibrated is the most suitable in terms of core performance. This weighting mechanism not only significantly improves the accuracy and reliability of model matching, avoiding mis-matching caused by secondary feature interference, but also solidifies technical knowledge in this field at the algorithm level, making the rapid calibration method have both engineering practicality and theoretical optimality.
[0062] In this specific embodiment, before the similarity matching between the fingerprint characteristics to be calibrated and the fingerprint characteristics of each sample in the model knowledge base in step S3, a step of normalizing and preprocessing the fingerprint characteristics to be calibrated is further included.
[0063] It should be noted that the normalization processing establishes a fair and unified comparison benchmark for subsequent similarity matching. Due to the great difference in physical meaning and numerical dimension of different fingerprint features (such as zero voltage, sensitivity coefficient, dynamic response time, etc.), direct similarity calculation will lead to the fact that the features with large numerical range (such as sensitivity) dominate the matching results, while the key features with small numerical range (such as micro load threshold) are ignored. Through normalization processing, the bias caused by different dimensions and orders of magnitude of each feature can be eliminated, ensuring that the weighted similarity algorithm can fairly consider the differences in each feature dimension, thereby significantly improving the accuracy and reliability of model matching, and avoiding the selection of inappropriate reference models for the sensor to be calibrated due to data scale problems.
[0064] Step S4, taking the target decoupling calibration model as a pre-training model, adjusting the pre-training model based on the fingerprint features to be calibrated, and obtaining a final calibration model special for the six-dimensional force sensor to be calibrated.
[0065] It should be noted that the adjustment in step S4 is fine-tuning, and the fine-tuning is the key of this step. It is not a mechanical application, but a targeted and small-range parameter optimization of the reference model according to the "fingerprint features" (individual differences) of the new sensor. It not only ensures the rapid calibration, but also makes the pre-training model adapt to the six-dimensional force sensor to be calibrated through fine-tuning, ensuring the accuracy of the calibration.
[0066] The adjustment of the pre-training model in step S4 is hierarchical transfer learning, including:
[0067] Freezing the basic feature extraction layer in the pre-training model representing the general physical law;
[0068] Based on the fingerprint features to be calibrated, the parameter adjustment is performed on the intermediate adaptation layer in the pre-training model responsible for environmental adaptability;
[0069] Based on the fingerprint features to be calibrated, the output layer in the pre-training model responsible for individualized calibration is retrained.
[0070] It should be noted that the embodiment freezes the basic feature extraction layer in the pre-trained model representing general physical laws, which reliably retains the universal knowledge such as strain-voltage relationship and mechanical principles learned by the benchmark model from massive data, preventing the forgetting of valuable core knowledge; by adjusting the parameters of the intermediate adaptation layer responsible for environmental adaptability based on the to-be-calibrated fingerprint features, the model can flexibly adapt to the specific working environment of the new sensor; finally, by retraining the output layer responsible for individual calibration, precise adaptation to the unique sensitivity, zero position and coupling characteristics of the new sensor is realized. This hierarchical processing mechanism perfectly balances the contradiction between knowledge inheritance and individual adaptation, ensuring the stability and generalization ability of the migrated model, and realizing the precision optimization for specific sensors, which is the key technical architecture to ensure the success of rapid calibration.
[0071] In this specific embodiment, after step S4, it further includes:
[0072] The to-be-calibrated fingerprint features are associated with the final calibration model and added to the model knowledge base.
[0073] It should be noted that the embodiment further enriches the model knowledge base.
[0074] The embodiment of the application changes the calibration process of the to-be-calibrated sensor from the traditional "training from scratch" to "fine-tuning based on prior knowledge" by constructing a "model knowledge base". For most to-be-calibrated sensors, hundreds or thousands of multi-dimensional and composite loads are no longer needed, and only a small amount of key "fingerprint features" need to be collected to complete the adaptation of the high-precision model. This greatly shortens the calibration time of a single sensor and improves the efficiency by more than an order of magnitude, greatly meeting the needs of large-scale production and rapid on-site maintenance.
[0075] Since full-load calibration is no longer needed for each sensor, the embodiment of the application significantly reduces the dependence on and occupation time of high-precision, high-cost six-dimensional joint loading equipment.
[0076] The core of the embodiment of the application is not to sacrifice precision for speed. On the contrary, by migrating the "benchmark decoupling calibration model" trained from a large amount of data, the to-be-calibrated sensor has a high-precision model basis at the start. The subsequent fine-tuning process is a "personalized" adaptation of the high-precision model based on its own physical characteristics, rather than learning from scratch. Therefore, the "final calibration model" obtained finally can approach the level of traditional full calibration in terms of precision, realizing the unity of speed and precision.
[0077] The embodiment of the application has natural scalability and adaptability. First of all, the "model knowledge base" is a system that can evolve continuously, and as more sensor data is accumulated, the coverage of the knowledge base and the accuracy of model matching will be higher and higher.
[0078] In summary, the embodiment of the present application fundamentally changes the calibration paradigm of the six-dimensional force sensor, successfully solves the bottleneck of traditional methods in efficiency, cost and flexibility by introducing transfer learning and fingerprint feature matching mechanism, and lays a solid technical foundation for large-scale calibration of six-dimensional force sensors.
[0079] It should be noted that, in this article, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.
[0080] Each embodiment in the specification is described in a relevant manner, and the same and similar parts between each embodiment can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.
[0081] The above is only a preferred embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A rapid calibration method for a six-dimensional force sensor based on transfer learning, characterized in that, The method includes: Step S1: Obtain N sample six-dimensional force sensors; apply multiple known forces or torques to the six dimensions of the sample six-dimensional force sensors and collect the corresponding output voltages to construct a calibration training data set for training the decoupling calibration model; collect the sample fingerprint features of the sample six-dimensional force sensors; wherein, N is an integer greater than 2, the calibration training data set includes forces and their corresponding output voltages or torques and their corresponding output voltages, and the fingerprint features include at least one of zero-position characteristic fingerprints, sensitivity and coupling fingerprints, dynamic response fingerprints, micro-load response fingerprints, and temperature sensitivity fingerprints; Step S2: Based on the calibration training data set of each sample's six-dimensional force sensor, train the corresponding decoupling calibration model; establish the mapping relationship between the decoupling calibration model of all samples' six-dimensional force sensors and the sample fingerprint features, and construct a model knowledge base; wherein, the decoupling calibration model is used to output the corresponding force or torque based on the output voltage; Step S3: Collect the fingerprint features of the six-dimensional force sensor to be calibrated; perform similarity matching between the fingerprint features to be calibrated and the fingerprint features of each sample in the model knowledge base to obtain the fingerprint features of the target sample with the highest similarity and its corresponding target decoupling calibration model; wherein, the six-dimensional force sensor to be calibrated is the same model as the sample six-dimensional force sensor; Step S4: Using the target decoupling calibration model as a pre-trained model, adjust the pre-trained model based on the fingerprint features to be calibrated to obtain the final calibration model specifically for the six-dimensional force sensor to be calibrated.
2. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, The step S1 of collecting the sample fingerprint features of the six-dimensional force sensor includes: Under the condition that the sample six-dimensional force sensor is not subjected to external force, the voltage output of each channel is continuously collected within a first preset time period to obtain the zero-position characteristic fingerprint of the sample six-dimensional force sensor; A preset force or torque is applied sequentially to each of the six dimensions of the sample's six-dimensional force sensor, and the voltage output during the process is recorded. The ratio of the voltage output of each of the six dimensions when stress is applied to the applied force or torque is extracted to obtain the main sensitivity. The voltage changes of other dimensions when stress is applied are extracted to obtain the cross-coupling coefficient. Based on the main sensitivity and the cross-coupling coefficient, the sensitivity and coupling fingerprint are obtained. When applying or removing load to the six-dimensional force sensor of the sample, sampling is performed for a second preset duration to record the entire process of voltage change to stability, thereby obtaining a dynamic response fingerprint. Gradually increasing forces or torques are applied sequentially to the six dimensions of the sample's six-dimensional force sensor. The threshold of the force or torque corresponding to the first deviation of the voltage output of each channel from its zero value is recorded to obtain the micro-load response fingerprint. The sample six-dimensional force sensor was placed in a temperature-controlled environment, and the temperature change was controlled. The relationship between the temperature change and the voltage output was recorded to obtain a temperature-sensitive fingerprint.
3. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, In step S3, the method further includes: In response to the fact that the similarity between the fingerprint feature to be calibrated and the fingerprint features of each sample in the model knowledge base is lower than the first similarity threshold, multiple known forces or torques are applied to the six-dimensional force sensor to be calibrated, and the corresponding output voltages are collected to obtain the corresponding new training data set. Based on the newly added training data set corresponding to the six-dimensional force sensor to be calibrated, a corresponding newly added decoupling calibration model is trained; a mapping relationship between the fingerprint features to be calibrated and the newly added decoupling calibration model is established and added to the model knowledge base.
4. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, The similarity matching in step S3 is achieved through a weighted comprehensive similarity algorithm. Different weight coefficients are configured for different types of fingerprint features. The weight of sensitivity and coupled fingerprints is configured to be the highest, followed by dynamic response fingerprints and zero-position characteristic fingerprints, and microload response fingerprints and temperature-sensitive fingerprints have the lowest weights.
5. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, The adjustment of the pre-trained model in step S4 is a hierarchical transfer learning process, including: Freeze the basic feature extraction layer representing general physical laws in the pre-trained model; Based on the fingerprint features to be calibrated, the parameters of the intermediate adaptation layer responsible for environmental adaptation in the pre-trained model are adjusted. Based on the fingerprint features to be calibrated, the output layer responsible for personalized calibration in the pre-trained model is retrained.
6. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, Following step S4, the method further includes: The fingerprint features to be calibrated are associated with the final calibration model and added to the model knowledge base.
7. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, The calibration training data set constructed in step S1 contains more types and numbers of payloads than the types and numbers of payloads applied when collecting the fingerprint features of the sample.
8. The rapid calibration method for a six-dimensional force sensor based on transfer learning according to claim 1, characterized in that, In step S3, before performing similarity matching between the fingerprint features to be calibrated and the fingerprint features of each sample in the model knowledge base, a normalization preprocessing step is also included for the fingerprint features to be calibrated.