Non-directional force sensor control method and system based on real-time force trend prediction

Abstract

The embodiment of the invention provides a non-directional force sensor control method and system based on real-time force trend prediction, and the method comprises the steps: collecting a real-time force signal sequence generated when a target mechanical device carries out the contact operation, and inputting the real-time force signal sequence into a pre-trained force trend prediction model; and obtaining a force direction trend prediction result output by the force trend prediction model, generating a multi-dimensional force compensation signal according to the matching degree of the force direction trend prediction result and the current closed-loop control parameter, adjusting a proportion-integral-differential coefficient in the current closed-loop control parameter based on the multi-dimensional force compensation signal, generating an updated closed-loop control parameter, and outputting the updated closed-loop control parameter. The updated closed-loop control parameters are input into an actuator of the target mechanical device, and the contact operation track of the target mechanical device is dynamically adjusted through the actuator according to the updated closed-loop control parameters, so that the multi-dimensional force amplitude data output by the non-directional force sensor are converged to a preset force control target interval. The response speed and stability of the force control system can be improved.

Description

Technical Field

[0001] The present invention relates to the fields of data processing and intelligent control, and in particular to a control method and system for a non-directional force sensor based on real-time force trend prediction. Background Art

[0002] In the field of industrial automation and precision control, force control systems are a core technology for enabling safe interaction between mechanical devices and the external environment. Due to their simple structure and low cost, non-directional force sensors are widely used in scenarios such as robotic arm contact operations and precision assembly. Existing force control methods typically use traditional proportional-integral-derivative (PID) algorithms to adjust closed-loop control parameters based on real-time force amplitude data output by non-directional sensors, thereby driving the actuator to correct the contact trajectory. However, these methods lack the ability to sense changes in force direction and rely solely on current force signal amplitude feedback for parameter adjustment. This makes the control system susceptible to directional misjudgment during dynamic contact operations, causing oscillation or response lag. In particular, in multi-dimensional force coupling scenarios, traditional PID parameter adjustment struggles to distinguish the directional trends of force amplitude changes in each dimension, making it impossible to proactively compensate for directional deviations. This leads to parameter conflicts and steady-state error accumulation, severely limiting force control accuracy and system stability. Achieving real-time prediction and precise compensation of multi-dimensional force directional trends without relying on directional sensors has become a key challenge in improving the performance of sensorless closed-loop control. Summary of the Invention

[0003] In the first aspect, an embodiment of the present invention provides a closed-loop control method for a non-directional force sensor based on real-time force trend prediction, the method comprising: collecting a real-time force signal sequence generated when a target mechanical device performs a contact operation, the real-time force signal sequence comprising multi-dimensional force amplitude data output by the non-directional force sensor; inputting the real-time force signal sequence into a pre-trained force trend prediction model to obtain a force direction trend prediction result output by the force trend prediction model, the force direction trend prediction result being used to characterize the vector distribution characteristics of the multi-dimensional force amplitude change within a preset time window in the future; and determining the closed-loop control method based on the difference between the force direction trend prediction result and the current closed-loop control method. The method comprises the following steps: adjusting the matching degree of the current closed-loop control parameters, generating a multi-dimensional force compensation signal, and correcting the directional deviation component in the current closed-loop control parameters; adjusting the proportional-integral-differential coefficients in the current closed-loop control parameters based on the multi-dimensional force compensation signal, generating updated closed-loop control parameters, and inputting the updated closed-loop control parameters into the actuator of the target mechanical device; dynamically adjusting the contact operation trajectory of the target mechanical device according to the updated closed-loop control parameters through the actuator, so that the multi-dimensional force amplitude data output by the non-directional force sensor converges to a preset force control target range.

[0004] In a second aspect, an embodiment of the present invention provides a computer system, comprising: a memory storing a computer program; and a processor for loading the computer program to implement the above-mentioned closed-loop control method of a non-directional force sensor based on real-time force trend prediction.

[0005] The present invention provides a closed-loop control method for a non-directional force sensor based on real-time force trend prediction. This method inputs a real-time force signal sequence into a pre-trained force trend prediction model to obtain force direction trend prediction results within a future time window. Based on the prediction results, a multi-dimensional force compensation signal is dynamically generated to adjust closed-loop control parameters, enabling the multi-dimensional force amplitude data output by the non-directional force sensor to quickly converge to a preset force control target range. This method innovatively integrates a dynamic coupling mechanism between real-time prediction and closed-loop control, addressing the control lag and oscillation issues caused by the lack of directional perception in traditional methods. By pre-compensating for directional deviation components through a prediction model and combining component decoupling of the multi-dimensional compensation signal, it effectively avoids parameter conflicts caused by directional misjudgment, significantly improving the response speed and stability of the force control system. Furthermore, this method, based on an adaptive architecture designed based on the characteristics of non-directional sensors, directly maps multi-dimensional force amplitude data into a vector distribution feature space without relying on directional sensors or complex coordinate transformations. Through a dual feedback mechanism, control parameter gain weights are dynamically optimized, achieving coordinated optimization of directional trend prediction and force amplitude control in complex contact operation scenarios. This provides a robust and adaptable force interaction control solution for scenarios such as precision assembly of robotic arms. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is a flow chart of a closed-loop control method for a non-directional force sensor based on real-time force trend prediction provided by an embodiment of the present invention.

[0007] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0008] See also Figure 1 , Figure 1 A flowchart of a closed-loop control method for a non-directional force sensor based on real-time force trend prediction provided by an embodiment of the present invention is provided. The closed-loop control method for a non-directional force sensor based on real-time force trend prediction can be executed by a computer system and includes the following steps: Step S100: collecting a real-time force signal sequence generated when the target mechanical device performs a contact operation, where the real-time force signal sequence includes multi-dimensional force amplitude data output by a non-directional force sensor.

[0009] In an embodiment of the present invention, the target mechanical device is a specific mechanical device performing a contact operation, such as the robotic arm of an industrial robot or processing equipment on an automated production line. Contact operation is the process by which the mechanical device comes into contact with an external object and generates an interaction force, such as a robot grasping an object or processing equipment polishing a workpiece. A real-time force signal sequence is a sequence of force signals continuously collected at preset time intervals while the target mechanical device is performing a contact operation. A non-directional force sensor is a sensor that can measure force magnitude but not its direction. The multi-dimensional force amplitude data it outputs represents the magnitude of force in different dimensions. For example, in three-dimensional space, this may include force amplitudes in the X, Y, and Z directions. The real-time force signal sequence is collected, for example, by sampling the force signal at a preset sampling frequency using a non-directional force sensor connected to the target mechanical device. For example, setting the sampling frequency to 100 Hz means that force signal data is collected 100 times per second. Each collected data point represents a multi-dimensional force amplitude data point. These data points are arranged in chronological order of acquisition time to form a real-time force signal sequence.

[0010] Step S200: Input the real-time force signal sequence into the pre-trained force trend prediction model to obtain the force direction trend prediction result output by the force trend prediction model. The force direction trend prediction result is used to characterize the vector distribution characteristics of the multi-dimensional force amplitude changes within a future preset time window.

[0011] The pre-trained force trend prediction model is a pre-trained model used to predict the changing trend of multi-dimensional force amplitudes over a future period of time based on a real-time force signal sequence input. The force direction trend prediction results use vector distribution features to represent the changes in multi-dimensional force amplitudes within a preset future time window. The preset future time window is a pre-set period of time during which changes in multi-dimensional force amplitudes are predicted, for example, the next 5 seconds.

[0012] As an embodiment, step S200, inputting the real-time force signal sequence into a pre-trained force trend prediction model to obtain the force direction trend prediction result output by the force trend prediction model, may specifically include the following steps S210 to S250: Step S210: performing sliding window segmentation processing on the real-time force signal sequence to generate an input data pair including a historical force signal subsequence and a current force signal subsequence.

[0013] Sliding window segmentation involves sliding a fixed-size window across the real-time force signal sequence, capturing the data within the window as a subsequence with each slide. The historical force signal subsequence is a subsequence consisting of force signal data prior to the current time point within the window; it contains force signal information from the past period. The current force signal subsequence is a subsequence consisting of force signal data corresponding to the current time point within the window; it reflects the force signal state at the current moment. The input data pair, consisting of the historical force signal subsequence and the current force signal subsequence, is subsequently input into the different layers of the force trend prediction model for processing.

[0014] For example, a sliding window is set to 10 data points. Starting from the start position of the real-time force signal sequence, the first 9 data points covered by the window are used as the historical force signal subsequence, and the 11th data point is used as the current force signal subsequence to form an input data pair. The window is then slid backward by one data point, and a new historical force signal subsequence and current force signal subsequence are intercepted again to form a new input data pair. This is repeated until the end of the real-time force signal sequence is reached. In the scenario of an industrial robot grasping an object, assuming that the real-time force signal sequence is collected at a sampling frequency of 100 times per second, the window size is set to 10 data points, that is, the time range covered by the window is 0.1 seconds. Through the sliding window segmentation process, input data pairs containing historical force signal subsequences and current force signal subsequences are continuously generated, providing a data basis for subsequent model processing.

[0015] Step S220: Input the historical force signal subsequence into the time series coding layer of the force trend prediction model to extract the time series correlation features corresponding to the historical force signal subsequence. The time series correlation features include the force amplitude fluctuation patterns under different time dimensions.

[0016] The temporal encoding layer of the force trend prediction model encodes the input historical force signal subsequences and extracts temporal correlation features. Temporal correlation features are the fluctuation patterns and correlations of force amplitudes across different time dimensions, such as rapid changes in force amplitude over short periods of time and slow trends over longer periods of time. Specifically, after the historical force signal subsequences are input to the temporal encoding layer, the layer processes them using an encoding algorithm. For example, a long short-term memory (LSTM) network is used as the temporal encoding layer architecture, with a hidden layer dimension of 64, an activation function of Tanh, and an input dimension consistent with the force signal dimension (e.g., 3 for a three-dimensional force signal input). When processing the historical force signal subsequences, the LSTM learns the fluctuation patterns of force amplitude across different time dimensions based on the input force amplitude data and outputs temporal correlation features containing this fluctuation pattern information.

[0017] Step S230: Input the current force signal subsequence into the dynamic weight allocation layer of the force trend prediction model, and generate a weight allocation vector corresponding to the time series correlation feature according to the force amplitude change rate at each time point in the current force signal subsequence.

[0018] The dynamic weight assignment layer of the force trend prediction model is used to assign different weights to time-series correlation features based on the characteristics of the current force signal subsequence. The force amplitude change rate at each time point in the current force signal subsequence is the ratio of the force amplitude change at each time point relative to the previous time point in the current force signal subsequence. The weight assignment vector has the same dimension as the time-series correlation feature, where each element represents the weight of the corresponding time-series correlation feature and is used in subsequent feature fusion operations.

[0019] For example, when the current force signal subsequence is input into the dynamic weight allocation layer, this layer calculates the rate of change of the force amplitude at each time point in the current force signal subsequence. For example, the force amplitude change rate is calculated using a first-order difference method: the force amplitude at the current time point is subtracted from the force amplitude at the previous time point, and then divided by the force amplitude at the previous time point. Based on the calculated force amplitude change rate, a nonlinear mapping function is then used to generate a weight allocation vector corresponding to the time series-related features. For example, a sigmoid function is used as the nonlinear mapping function to map the force amplitude change rate to the interval [0, 1], thereby obtaining the values ​​of each element of the weight allocation vector.

[0020] Step S240: Through the feature fusion layer of the force trend prediction model, the time series correlation feature and the weight distribution vector are subjected to a dot product operation to generate a dynamic weighted time series feature.

[0021] The feature fusion layer of the force trend prediction model is a layer for fusing different features. In an embodiment of the present invention, it fuses the time series correlation features and the weight allocation vector. The dot product operation multiplies the corresponding elements of two vectors and sums them to obtain a new vector. Of course, in other implementations, the element-by-element multiplication method can also be used. The dynamic weighted time series feature is the result obtained by the dot product operation. It is a feature obtained by weighting the time series correlation feature according to the weight allocation vector, which can more accurately reflect the characteristics of the current force signal and the future force change trend.

[0022] It can be understood that when those skilled in the art implement the relevant calculations of the embodiments of the present invention, if the calculation process involves different dimensions, they can rely on their own technical common sense to perform dimensional unification or conflict elimination operations, such as standardization or normalization processing. For example, the subsequent loss calculations involved can be in the form of percentages to eliminate dimensional differences. The subsequent PID parameters, proportional, integral, and differential coefficients are normalized to [0,1] according to type. Before the subsequent calculation of the rate of change, the normalized data is restored, etc. In addition, for two variables with inconsistent dimensions, those skilled in the art can align the dimensions according to well-known technologies before performing calculations. For example, before performing the dot product operation on the temporal correlation features output by the temporal coding layer and the weight vector dimensions generated by the dynamic weight allocation layer, if the dimensions are inconsistent, the dimensions of the two can be aligned first (such as through a fully connected layer) before performing subsequent operations.

[0023] Step S250: Input the dynamic weighted time series features into the trend prediction layer of the force trend prediction model, output the multidimensional force amplitude prediction value at each time point in the future preset time window, and generate the force direction trend prediction result based on the vector change direction of the multidimensional force amplitude prediction value.

[0024] The trend prediction layer of the force trend prediction model is used to predict the multidimensional force amplitude at each time point within a preset future time window based on the input dynamically weighted time series features, and generate a force direction trend prediction result. This layer can be implemented as a fully connected layer. The multidimensional force amplitude prediction value at each time point within the preset future time window is the predicted force magnitude in different dimensions at each time point within the pre-set future time period. The force direction trend prediction result is generated based on the vector change direction of the multidimensional force amplitude prediction value, and it uses vector distribution characteristics to represent the future force change trend. For example, after the dynamic weighted time series features are input to the trend prediction layer, the layer uses its internal prediction algorithm to predict the multidimensional force amplitude within the preset future time window. For example, a fully connected neural network is used as the structure of the trend prediction layer. Based on the input dynamically weighted time series features, the fully connected neural network outputs the multidimensional force amplitude prediction value at each time point within the preset future time window through a series of linear transformations and nonlinear activation functions. Then, according to the vector change direction of these multi-dimensional force amplitude prediction values, for example, the difference of the multi-dimensional force amplitude vector between adjacent time points is calculated to obtain the force change direction, and then generate the force direction trend prediction result.

[0025] Step S300: generating a multi-dimensional force compensation signal according to the matching degree between the force direction trend prediction result and the current closed-loop control parameters. The multi-dimensional force compensation signal is used to correct the direction deviation component in the current closed-loop control parameters.

[0026] The force direction trend prediction result is the result obtained in step S200 that represents the future force change trend. The current closed-loop control parameters are a set of parameters used to control the target mechanical device at the current moment, such as proportional coefficients, integral coefficients, and differential coefficients. The matching degree represents the degree of similarity or conformity between the force direction trend prediction result and the current closed-loop control parameters. The multi-dimensional force compensation signal is a signal generated based on the matching degree between the force direction trend prediction result and the current closed-loop control parameters. Its function is to correct the directional deviation component in the current closed-loop control parameters to make the control of the target mechanical device more accurate. For example, the matching degree between the force direction trend prediction result and the current closed-loop control parameters is calculated. For example, a vector similarity calculation method, such as cosine similarity, can be used to calculate the similarity between the vector distribution characteristics corresponding to the force direction trend prediction result and the vector corresponding to the current closed-loop control parameters. Then, based on the calculated matching degree, a multi-dimensional force compensation signal is generated by a preset compensation signal generation algorithm. In the scenario of an industrial robot grasping an object, if the force direction trend prediction results show that the force change trend of the robot arm in a certain direction is inconsistent with the expected current closed-loop control parameters, it indicates that there is a directional deviation. The multi-dimensional force compensation signal generated at this time will be used to correct the current closed-loop control parameters to make the movement of the robot arm more consistent with the actual force change situation.

[0027] As an embodiment, step S300 generates a multi-dimensional force compensation signal according to the matching degree between the force direction trend prediction result and the current closed-loop control parameters, which may specifically include the following steps S310 to S350: Step S310: Obtain the proportional coefficient vector, integral coefficient vector and differential coefficient vector in the current closed-loop control parameters, and normalize the proportional coefficient vector, integral coefficient vector and differential coefficient vector to generate a closed-loop control parameter characteristic matrix.

[0028] The proportional coefficient vector, integral coefficient vector, and differential coefficient vector in the current closed-loop control parameters are the parameters used for closed-loop control. The proportional coefficient vector adjusts the proportional relationship between the controlled variable and the error; the integral coefficient vector eliminates the system's steady-state error; and the differential coefficient vector predicts the error's changing trend and enables proactive adjustments. As previously mentioned, normalization converts data to a range, such as [0, 1] or [-1, 1], to eliminate dimensional differences between parameters and make the data comparable. The closed-loop control parameter characteristic matrix is ​​a matrix composed of the normalized proportional coefficient vector, integral coefficient vector, and differential coefficient vector. It more comprehensively represents the characteristics of the current closed-loop control parameters.

[0029] For example, the proportional coefficient vector, integral coefficient vector, and differential coefficient vector are extracted from the current closed-loop control parameters. For example, assume the proportional coefficient vector is [p1, p2, p3], the integral coefficient vector is [i1, i2, i3], and the differential coefficient vector is [d1, d2, d3]. These three vectors are then normalized. A minimum-maximum normalization method can be used: for each vector, first find its minimum and maximum values, then subtract the minimum value from each element and divide by the difference between the maximum and minimum values ​​to obtain the normalized vector. Finally, the normalized proportional coefficient vector, integral coefficient vector, and differential coefficient vector are arranged in rows or columns to generate a closed-loop control parameter feature matrix. In scenarios where industrial robots are grasping objects, normalization and generation of the closed-loop control parameter feature matrix can more accurately represent the characteristics of the current closed-loop control parameters, providing a basis for subsequent similarity calculation and compensation signal generation.

[0030] Step S320: performing similarity calculation on the vector distribution characteristics corresponding to the force direction trend prediction result and the closed-loop control parameter characteristic matrix to generate a direction deviation characteristic vector.

[0031] The vector distribution features corresponding to the force direction trend prediction result are the vector features representing the future force change trend obtained in step S200. The closed-loop control parameter feature matrix is ​​the matrix representing the current closed-loop control parameter features generated in step S310. Similarity calculations are performed to determine the degree of similarity between these two features, such as Euclidean distance or cosine similarity. The directional deviation feature vector is generated based on the similarity calculation results and represents the directional deviation between the force direction trend prediction result and the current closed-loop control parameters.

[0032] For example, a similarity calculation is performed between the vector distribution features corresponding to the force direction trend prediction results and the closed-loop control parameter feature matrix. For example, the cosine similarity calculation method is used to calculate the cosine similarity between the vector distribution features and each row vector in the closed-loop control parameter feature matrix. Then, based on the calculated similarity values, a directional deviation feature vector is generated. For example, a preset threshold can be subtracted from the similarity value to obtain the values ​​of each element of the directional deviation feature vector.

[0033] Step S330: inputting the directional deviation feature vector into a fully connected layer of a preset compensation signal generation network, and extracting a preset oscillation component and a preset offset component from the directional deviation feature vector.

[0034] The preset compensation signal generation network is used to generate a multidimensional force compensation signal based on the input directional deviation feature vector. The fully connected layer connects each input neuron to each output neuron. The preset oscillation component is the part of the directional deviation feature vector that changes at a higher frequency, indicating rapid fluctuations or noise in the system. The preset offset component is the part of the directional deviation feature vector that changes at a lower frequency, indicating a slow trend or steady-state error in the system. How to define a higher or lower frequency is set according to the actual situation. For example, a frequency above 100 Hz is defined as a high frequency, and a frequency below 100 Hz is defined as a low frequency.

[0035] During the specific implementation process, for example, the directional deviation feature vector generated in step S320 is input into the fully connected layer of the preset compensation signal generation network. The fully connected layer performs a linear transformation on the input directional deviation feature vector and maps the input feature vector to a new feature space through a series of weight matrices and bias vectors. Then, a filtering algorithm or signal processing method is used to extract the preset oscillation component and the preset offset component from the output feature vector. For example, a low-pass filter can be used to extract the low-frequency offset component, and a high-pass filter can be used to extract the high-frequency oscillation component.

[0036] Step S340: performing cross-channel fusion on the preset oscillation component and the preset offset component through the convolution kernel group of the compensation signal generation network to generate a preliminary compensation signal.

[0037] The convolution kernel group of the compensation signal generation network performs convolution operations on the input features. Cross-channel fusion combines features from different channels to extract richer feature information. The preliminary compensation signal is generated through cross-channel fusion and serves as an intermediate result in the generation of the multidimensional force compensation signal.

[0038] For example, the preset oscillation component and the preset offset component extracted in step S330 are input to the convolution kernel group of the compensation signal generation network. The convolution kernel group performs a convolution operation on the preset oscillation component and the preset offset component, extracting different feature information using different convolution kernels. The results of the convolution operation are then fused across channels, for example, using feature concatenation or feature addition to fuse the features of different channels together to generate a preliminary compensation signal.

[0039] As an embodiment, step S340, performing cross-channel fusion of the preset oscillation component and the preset offset component by the convolution kernel group of the compensation signal generation network to generate a preliminary compensation signal, may specifically include the following steps S341 to S348: Step S341 : Split the preset oscillation component and the preset offset component into multiple parallel processing channels along the channel dimension to generate a preset oscillation component sub-channel group and a preset offset component sub-channel group.

[0040] The channel dimension represents the number of feature channels. Parallel processing channels are multiple independent channels obtained by splitting feature data along the channel dimension, each of which can be processed independently. A preset oscillation component subchannel group is a group of subchannels obtained by splitting the preset oscillation component along the channel dimension. A preset offset component subchannel group is a group of subchannels obtained by splitting the preset offset component along the channel dimension.

[0041] For example, the preset oscillation component and the preset offset component are each split into multiple parallel processing channels based on the number of their channel dimensions. For example, assuming the channel dimension of the preset oscillation component is 3, it is split into 3 preset oscillation component sub-channels; assuming the channel dimension of the preset offset component is 2, it is split into 2 preset offset component sub-channels. This splitting method allows different channels to be processed independently, improving processing efficiency and feature extraction accuracy. Splitting the preset oscillation component and the preset offset component into multiple parallel processing channels can better capture feature information on different channels, providing more detailed data for subsequent convolution operations and feature fusion.

[0042] Step S342: decompose the convolution kernel group into a first sub-convolution kernel set corresponding to the preset oscillation component sub-channel group and a second sub-convolution kernel set corresponding to the preset offset component sub-channel group.

[0043] The first sub-convolution kernel set is the portion of the convolution kernel group corresponding to the preset oscillation component sub-channel group, and the second sub-convolution kernel set is the portion of the convolution kernel group corresponding to the preset offset component sub-channel group. By decomposing the convolution kernel group, different convolution kernels can be used for different sub-channel groups, improving the pertinence and effectiveness of the convolution operation.

[0044] For example, based on the number and characteristics of the preset oscillation component sub-channel group and the preset offset component sub-channel group, the convolution kernel group is decomposed into a first sub-convolution kernel set and a second sub-convolution kernel set. If the preset oscillation component sub-channel group has 3 channels and the preset offset component sub-channel group has 2 channels, the convolution kernels in the convolution kernel group are allocated to the first sub-convolution kernel set and the second sub-convolution kernel set according to a preset rule, so that the number of convolution kernels in the first sub-convolution kernel set matches the number of channels of the preset oscillation component sub-channel group, and the number of convolution kernels in the second sub-convolution kernel set matches the number of channels of the preset offset component sub-channel group. By decomposing the convolution kernel group into the first sub-convolution kernel set and the second sub-convolution kernel set, more efficient convolution operations can be performed on the different characteristics of the preset oscillation component and the preset offset component, thereby extracting more accurate feature information.

[0045] Step S343: performing a multi-scale convolution operation on the preset oscillation component sub-channel group through the first sub-convolution kernel set, extracting the time domain response features corresponding to different frequency fluctuation modes in the preset oscillation component sub-channel group, and generating a preset fusion feature map.

[0046] Multi-scale convolution operations use convolution kernels of different sizes to convolve input features to extract feature information at different scales. Time-domain response features are the response characteristics of the input signal in the time domain, reflecting the signal's changes at different time points. The preset fused feature map is the result of multi-scale convolution operations. It is a feature map obtained by fusing the time-domain response features corresponding to different frequency fluctuation patterns in the preset oscillation component sub-channel group.

[0047] For example, a preset oscillation component sub-channel group is input into a first sub-convolution kernel set, and convolution kernels of different sizes in the first sub-convolution kernel set perform a convolution operation on the preset oscillation component sub-channel group. These convolution results are then fused, for example by feature addition or feature splicing, to generate a preset fusion feature map. Through multi-scale convolution operations and feature fusion, it is possible to extract the time domain response characteristics corresponding to different frequency fluctuation patterns in the preset oscillation component sub-channel group, generate a more comprehensive and accurate preset fusion feature map, and provide important feature information for subsequent feature splicing and compensation signal generation.

[0048] Step S344: performing a long-period convolution operation on the preset offset component sub-channel group through the second sub-convolution kernel set, extracting the cumulative effect features corresponding to the trend offset in the preset offset component sub-channel group, and generating a preset fusion feature map.

[0049] Long-period convolution uses a large kernel size (e.g., 9×9) to convolve the input features, capturing feature information over a longer timeframe. The cumulative effect feature corresponding to the trend shift is the cumulative effect of the slowly changing trend shift in the preset offset component sub-channel group over a longer period of time. The preset fused feature map is the result of the long-period convolution operation. It is a feature map formed by fusing the cumulative effect features corresponding to the trend shift in the preset offset component sub-channel group.

[0050] For example, a preset offset component sub-channel group is input into a second sub-convolution kernel set, and a larger-sized convolution kernel in the second sub-convolution kernel set performs a long-period convolution operation on the preset offset component sub-channel group. The convolution results are then fused, for example, using feature addition or feature splicing methods to generate a preset fused feature map. Through long-period convolution operations and feature fusion, the cumulative effect features corresponding to the trend offset in the preset offset component sub-channel group can be extracted, generating a more accurate preset fused feature map, which provides important feature information for subsequent feature splicing and compensation signal generation.

[0051] Step S345: perform feature splicing on the preset fusion feature map and the preset fusion feature map along the spatial dimension to generate a cross-channel fusion feature tensor.

[0052] The spatial dimension is a dimension of the feature map, representing the spatial location information of the feature map. The cross-channel fused feature tensor is a tensor obtained by concatenating the preset fused feature map with the preset fused feature map along the spatial dimension. It integrates the preset and preset feature information, providing a richer feature × representation for subsequent feature enhancement and compensation signal generation.

[0053] For example, the preset fused feature map generated in step S343 and the preset fused feature map generated in step S344 are spliced ​​in the spatial dimension. For example, if the spatial dimensions of the preset fused feature map and the preset fused feature map are HW×C1 and H×W×C2 respectively (where H and W represent the spatial dimensions, and C1 and C2 represent the number of channels), they are spliced ​​in the channel dimension to obtain a cross-channel fused feature tensor with a spatial dimension of H×W×(C1+C2). Generating a cross-channel fused feature tensor by feature splicing can integrate the preset and preset feature information, providing a more comprehensive feature representation for subsequent feature processing and compensation signal generation.

[0054] Step S346: Perform cross-layer feature enhancement on the cross-channel fusion feature tensor through the residual connection layer of the compensation signal generation network to generate enhanced fusion features.

[0055] The residual connection layer of the compensation signal generation network is a special network layer that adds input features to processed features through residual connections to alleviate the vanishing gradient problem in deep neural networks and enhance the expressiveness of features. Cross-layer feature enhancement uses the residual connection layer to fuse and enhance feature information from different layers, making the features more representative and expressive. Enhanced fused features are features obtained by performing cross-layer feature enhancement on the cross-channel fused feature tensor through the residual connection layer.

[0056] For example, the cross-channel fusion feature tensor generated in step S345 is input into the residual connection layer of the compensation signal generation network. The residual connection layer takes the cross-channel fusion feature tensor as input, processes it through a series of convolutional layers and nonlinear activation functions, and obtains a processed feature. Then, the processed feature is added to the input cross-channel fusion feature tensor to obtain an enhanced fusion feature. The cross-layer feature enhancement of the residual connection layer can alleviate the gradient vanishing problem in the deep neural network, enhance the expressive power of the cross-channel fusion feature tensor, generate more representative and accurate enhanced fusion features, and provide a better feature basis for subsequent feature dimensionality reduction and compensation signal generation.

[0057] Step S347: inputting the enhanced fusion feature into the depthwise separable convolutional layer of the compensation signal generation network to perform feature dimensionality reduction, and generating a fusion feature vector after dimensionality reduction.

[0058] The depthwise separable convolutional layer of the compensation signal generation network is a convolutional layer that decomposes the traditional convolution operation into two steps: depthwise convolution and pointwise convolution. This reduces computational complexity while also achieving feature dimensionality reduction. Feature dimensionality reduction converts high-dimensional feature data into low-dimensional feature data to reduce data complexity and computational complexity. The reduced fused feature vector is obtained by performing feature dimensionality reduction on the enhanced fused features using the depthwise separable convolutional layer.

[0059] For example, the enhanced fusion features generated in step S346 are input into the depthwise separable convolution layer of the compensation signal generation network. The depthwise separable convolution layer first performs depthwise convolution on the enhanced fusion features, that is, convolution operation is performed on each channel separately to extract feature information within the channel. Then, the result of the depthwise convolution is convolved point by point, that is, the result of the depthwise convolution is convolved with a 1×1 convolution kernel to achieve feature dimensionality reduction. Finally, the result of the pointwise convolution is flattened into a vector to obtain a fusion feature vector after dimensionality reduction. Through the feature dimensionality reduction of the depthwise separable convolution layer, the complexity and computational complexity of the data can be reduced, while retaining the important information in the enhanced fusion features, generating a more concise and effective fusion feature vector after dimensionality reduction, and providing suitable feature input for subsequent amplitude constraint and phase alignment.

[0060] Step S348: Perform amplitude constraint and phase alignment on the fusion feature vector after dimension reduction through a nonlinear mapping function to generate a preliminary compensation signal that matches the dimension of the closed-loop control parameter feature matrix.

[0061] A nonlinear mapping function, such as the Sigmoid function or the ReLU function, performs a nonlinear transformation on the input data. Amplitude constraint limits the amplitude of the fused feature vector after dimensionality reduction to a reasonable range. Phase alignment adjusts the phase of the fused feature vector after dimensionality reduction to align with the phase of the closed-loop control parameter feature matrix. The preliminary compensation signal is generated by applying amplitude constraint and phase alignment to the fused feature vector after dimensionality reduction using the nonlinear mapping function. Its dimension matches the closed-loop control parameter feature matrix, providing a suitable signal input for subsequent nonlinear transformation and multidimensional force compensation signal generation. For example, the fused feature vector after dimensionality reduction generated in step S347 is input into the nonlinear mapping function. For example, a Sigmoid function is used to apply amplitude constraint to the fused feature vector after dimensionality reduction, limiting its amplitude to the range [0, 1]. Then, a phase adjustment algorithm is used to adjust the phase of the fused feature vector after dimensionality reduction to align with the phase of the closed-loop control parameter feature matrix. Finally, a preliminary compensation signal is obtained that matches the dimension of the closed-loop control parameter feature matrix. Through the amplitude constraint and phase alignment of the nonlinear mapping function, a preliminary compensation signal that matches the dimension of the closed-loop control parameter characteristic matrix can be generated, providing suitable signal input for the subsequent multi-dimensional force compensation signal generation, ensuring that the compensation signal can accurately correct the directional deviation component in the current closed-loop control parameters.

[0062] As an implementation method, the above-mentioned multi-dimensional force compensation signal generation process further includes the following constraints: When the force amplitudes in at least two dimensions change in opposite directions in the force direction trend prediction results, the compensation signal amplitudes in these two dimensions are preferentially suppressed. This is because when the force amplitudes in at least two dimensions change in opposite directions, this can lead to system instability or overcompensation. By preferentially suppressing the compensation signal amplitudes in these two dimensions, this can be avoided, making the system more stable. For example, if the force direction trend prediction results show that the force amplitudes in the X and Y directions change in opposite directions, the compensation signal amplitudes in the X and Y directions will be preferentially reduced when generating the multi-dimensional force compensation signal.

[0063] When the force amplitude changes in all dimensions in the force direction trend prediction results are consistent, the gain coefficient of the multi-dimensional force compensation signal in the dominant direction of change is increased. The dominant direction of change is the direction in which the force amplitude changes most significantly across all dimensions. When the force amplitude changes in all dimensions are consistent, increasing the gain coefficient of the multi-dimensional force compensation signal in the dominant direction of change can more effectively correct the directional deviation component in the current closed-loop control parameters, resulting in more accurate control of the target mechanical device. For example, if the force direction trend prediction results show that the force amplitudes in the X, Y, and Z directions are all increasing, with the increase in the X direction being the most significant, then the gain coefficient of the multi-dimensional force compensation signal in the X direction will be increased when generating the multi-dimensional force compensation signal.

[0064] When the force direction trend prediction results contain periodic oscillation patterns, phase lag compensation is applied to the multi-dimensional force compensation signal. A periodic oscillation pattern is a periodic fluctuation in force amplitude over time. Applying phase lag compensation to the multi-dimensional force compensation signal allows the compensation signal to better track force changes and reduce system oscillations. For example, if the force direction trend prediction results indicate periodic oscillations in the force amplitude in a certain direction, phase lag compensation is applied to the compensation signal in that direction when generating the multi-dimensional force compensation signal.

[0065] When the confidence level of the force direction trend prediction result is lower than the preset threshold, the multi-dimensional force compensation signal is reset to zero and switched to open-loop control mode. The confidence level of the force direction trend prediction result indicates the reliability of the prediction result. The preset threshold is a pre-set value used to determine whether the confidence level of the prediction result is high enough. When the confidence level of the force direction trend prediction result is lower than the preset threshold, it means that the reliability of the prediction result is low. At this time, the multi-dimensional force compensation signal is reset to zero and switched to open-loop control mode to avoid system instability caused by unreliable prediction results. For example, if the confidence level of the force direction trend prediction result is lower than the preset threshold, the multi-dimensional force compensation signal will be reset to zero and switched to open-loop control mode until the confidence level of the prediction result increases.

[0066] The activation status of the constraints is dynamically controlled by the prediction confidence index output by the force trend prediction model. The prediction confidence index is an indicator output by the force trend prediction model that indicates the confidence level of the prediction results. Based on the value of the prediction confidence index, the constraints can be dynamically activated or deactivated. For example, a high prediction confidence index indicates high reliability of the prediction results, and some constraints can be deactivated. A low prediction confidence index indicates low reliability of the prediction results, and some constraints can be activated to ensure system stability and control accuracy.

[0067] Step S350: Input the preliminary compensation signal into the activation function layer of the compensation signal generation network for nonlinear transformation to generate a multi-dimensional force compensation signal, wherein the output range of the activation function layer matches the adjustable range of the closed-loop control parameter.

[0068] The activation function layer of the compensation signal generation network performs a nonlinear transformation on the input signal. Nonlinear transformations can enrich the signal's expressiveness and improve the system's control performance. The multidimensional force compensation signal is generated by the activation function layer through a nonlinear transformation of the initial compensation signal. It is used to correct the directional deviation component of the current closed-loop control parameters. The output range of the activation function layer matches the adjustable range of the closed-loop control parameters to ensure that the generated multidimensional force compensation signal can effectively function within the adjustable range of the closed-loop control parameters.

[0069] For example, the preliminary compensation signal generated in step S348 is input into the activation function layer of the compensation signal generation network. The activation function layer can adopt different activation functions, such as Sigmoid function, ReLU function, etc. According to the adjustable range of the closed-loop control parameters, a suitable activation function is selected so that the output range of the activation function layer matches the adjustable range of the closed-loop control parameters. For example, if the adjustable range of the closed-loop control parameters is [-1, 1], the Sigmoid function can be selected as the activation function, and the preliminary compensation signal is nonlinearly transformed so that its output range is also within [-1, 1]. Finally, a multi-dimensional force compensation signal is obtained. By generating a multi-dimensional force compensation signal that matches the adjustable range of the closed-loop control parameters through the nonlinear transformation of the activation function layer, the directional deviation component in the current closed-loop control parameters can be more effectively corrected, making the control of the robotic arm more accurate.

[0070] Step S400: adjusting the proportional-integral-derivative coefficients in the current closed-loop control parameters based on the multi-dimensional force compensation signal, generating updated closed-loop control parameters, and inputting the updated closed-loop control parameters into the actuator of the target mechanical device.

[0071] The multi-dimensional force compensation signal is a signal generated in step S300 and used to correct the directional deviation component in the current closed-loop control parameters. The proportional-integral-differential coefficient is a component of the closed-loop control parameters. The proportional coefficient is used to adjust the proportional relationship between the control quantity and the error, the integral coefficient is used to eliminate the steady-state error of the system, and the differential coefficient is used to predict the changing trend of the error and make adjustments in advance. The updated closed-loop control parameters are the parameters obtained by adjusting the proportional-integral-differential coefficient in the current closed-loop control parameters according to the multi-dimensional force compensation signal. The actuator of the target mechanical device is the component responsible for executing the control instructions, such as a motor, cylinder, etc.

[0072] For example, the proportional-integral-differential coefficients in the current closed-loop control parameters are adjusted based on the multi-dimensional force compensation signal. For example, the multi-dimensional force compensation signal is decomposed into proportional, integral, and differential compensation terms, which are then superimposed with the proportional coefficient vector, integral coefficient vector, and differential coefficient vector in the current closed-loop control parameters. The superimposed coefficient vectors are then processed appropriately, such as through gradient constraints, integral saturation suppression, and noise filtering, to ensure that the adjusted coefficient vectors are within a reasonable range. Finally, the processed proportional, integral, and differential coefficient vectors are concatenated to generate updated closed-loop control parameters. These updated closed-loop control parameters are then input into the actuator of the target mechanical device, which then executes the corresponding control action based on the updated closed-loop control parameters. By adjusting the proportional-integral-differential coefficients based on the multi-dimensional force compensation signal and generating updated closed-loop control parameters, which are then input into the joint motors of the robotic arm, the robotic arm's motion can be made more consistent with actual force variations, improving grasping accuracy and stability.

[0073] As an embodiment, step S400, adjusting the proportional-integral-differential coefficients in the current closed-loop control parameters based on the multi-dimensional force compensation signal, may specifically include the following steps S410 to S450: Step S410: Decompose the multi-dimensional force compensation signal into a proportional compensation term, an integral compensation term, and a differential compensation term, and perform superposition operations with the proportional coefficient vector, the integral coefficient vector, and the differential coefficient vector in the current closed-loop control parameters respectively.

[0074] The multidimensional force compensation signal contains information from multiple dimensions. The proportional compensation term is used to adjust the proportional coefficient, the integral compensation term is used to adjust the integral coefficient, and the differential compensation term is used to adjust the differential coefficient. The proportional coefficient vector, integral coefficient vector, and differential coefficient vector in the current closed-loop control parameters are important parameters for closed-loop control. The superposition operation adds the corresponding elements of the proportional compensation term to the proportional coefficient vector, the integral compensation term to the integral coefficient vector, and the differential compensation term to the differential coefficient vector to obtain the adjusted coefficient vector.

[0075] For example, the multi-dimensional force compensation signal generated in step S350 is decomposed into a proportional compensation term, an integral compensation term, and a differential compensation term according to a preset rule. For example, the signal can be decomposed according to its frequency characteristics or time characteristics, with a preset portion as a differential compensation term, a preset portion as an integral compensation term, and an intermediate frequency portion as a proportional compensation term. Then, the proportional compensation term is superimposed on the proportional coefficient vector in the current closed-loop control parameters, the integral compensation term is superimposed on the integral coefficient vector, and the differential compensation term is superimposed on the differential coefficient vector. By decomposing the multi-dimensional force compensation signal into the proportional compensation term, the integral compensation term, and the differential compensation term, and superimposing them with the corresponding coefficient vectors, the closed-loop control parameters can be adjusted in a timely manner according to the actual force changes, thereby improving the control performance of the robotic arm.

[0076] Step S420: performing gradient constraint processing on the superimposed proportional coefficient vector so that the modulus of the adjusted proportional coefficient vector does not exceed a preset maximum proportional gain threshold.

[0077] Gradient constraint processing is a method for constraining vectors by limiting the gradient of the vector so that its modulus does not exceed a preset threshold. The superimposed proportional coefficient vector is the proportional coefficient vector obtained through the superposition operation in step S410. The preset maximum proportional gain threshold is a pre-set value used to limit the modulus of the proportional coefficient vector to prevent excessively large proportional coefficients from causing system instability.

[0078] For example, the modulus of the superimposed proportional coefficient vector is calculated. For example, the modulus of the vector is calculated using the Euclidean distance. Then, it is determined whether the modulus of the vector exceeds the preset maximum proportional gain threshold. If it exceeds, the proportional coefficient vector is adjusted so that its modulus is equal to the preset maximum proportional gain threshold. The adjustment can be performed using a normalization method, that is, each element of the proportional coefficient vector is divided by the modulus of the vector, and then multiplied by the preset maximum proportional gain threshold. By performing gradient constraint processing on the superimposed proportional coefficient vector, it is possible to ensure that the proportional coefficient is within a reasonable range, thereby avoiding unstable movement of the robotic arm due to an excessively large proportional coefficient.

[0079] Step S430: performing integral saturation suppression processing on the superimposed integral coefficient vector, and limiting the integral accumulation rate by dynamically adjusting the integral time constant.

[0080] Integral windup suppression is a method used to prevent excessive accumulation of the integral term. It limits the rate of integral accumulation by dynamically adjusting the integral time constant. The superimposed integral coefficient vector is the integral coefficient vector obtained through the superposition operation in step S410. The integral time constant is an important parameter in integral control, which determines the rate of accumulation of the integral term.

[0081] For example, during integral control, the accumulated value of the integral term is monitored in real time. When the accumulated value approaches or reaches a preset saturation threshold, the integral time constant is dynamically adjusted. For example, increasing the integral time constant slows the accumulation of the integral term and avoids integral saturation. By applying integral saturation suppression to the superimposed integral coefficient vector, excessive accumulation of the integral term can be prevented, potentially leading to deviations in the robotic arm's control.

[0082] Step S440: performing noise filtering on the superimposed differential coefficient vector, and using a sliding average algorithm to eliminate the influence of the preset noise on the differential term.

[0083] Noise filtering is a method for removing noise from a signal. The sliding average algorithm is a noise filtering algorithm that smoothes the signal and eliminates preset noise by averaging multiple consecutive data points of the signal. The superimposed differential coefficient vector is the differential coefficient vector obtained by the superposition operation in step S410. The preset noise is a high-frequency noise component in the signal, which interferes with the calculation of the differential term and affects the control performance of the system. For example, the superimposed differential coefficient vector is used as input and the sliding average algorithm is used for noise filtering. For example, the size of a sliding window is set, such as 5 data points. For each data point, the average value of the 5 data points in the window is calculated as the filtered value of the data point. In this way, the signal can be smoothed and the influence of the preset noise on the differential term can be eliminated. By performing noise filtering on the superimposed differential coefficient vector, the calculation accuracy of the differential term can be improved, making the control of the robotic arm more stable.

[0084] Step S450: performing vector concatenation on the processed proportional coefficient vector, integral coefficient vector, and differential coefficient vector to generate updated closed-loop control parameters.

[0085] The processed proportional coefficient vector is the proportional coefficient vector processed in step S420, the processed integral coefficient vector is the integral coefficient vector processed in step S430, and the processed differential coefficient vector is the differential coefficient vector processed in step S440. Vector concatenation is the process of sequentially concatenating these three vectors to form a new vector, i.e., the updated closed-loop control parameters. By concatenating the processed proportional coefficient vector, integral coefficient vector, and differential coefficient vector to generate updated closed-loop control parameters, the adjusted and processed proportional-integral-differential coefficients can be integrated to provide accurate control parameters for the actuator of the target mechanical device.

[0086] Step S500: dynamically adjusting the contact operation trajectory of the target mechanical device according to the updated closed-loop control parameters through the actuator, so that the multi-dimensional force amplitude data output by the non-directional force sensor converges to a preset force control target range.

[0087] The actuator is the component in the target mechanical device that is responsible for executing control instructions, such as a motor, cylinder, etc. The updated closed-loop control parameters are the parameters generated in step S400 for controlling the target mechanical device. The contact operation trajectory of the target mechanical device is the motion trajectory of the mechanical device when performing the contact operation. Dynamic adjustment is to continuously adjust the motion trajectory of the target mechanical device based on the real-time force signal and the updated closed-loop control parameters. The multi-dimensional force amplitude data output by the non-directional force sensor is the force signal data collected in real time by the non-directional force sensor installed on the target mechanical device. The preset force control target interval is a pre-set force amplitude range, and the goal is to make the multi-dimensional force amplitude data output by the non-directional force sensor converge within this range.

[0088] For example, the actuator drives the joint motor of the target mechanical device to perform corresponding angle adjustment or displacement adjustment based on the updated closed-loop control parameters, thereby changing the contact operation trajectory of the target mechanical device. During the adjustment process, the multi-dimensional force amplitude data output by the non-directional force sensor is monitored in real time and compared with the preset force control target range. If the multi-dimensional force amplitude data is not within the force control target range, the updated closed-loop control parameters are further adjusted according to the deviation, and the contact operation trajectory is continued to be adjusted until the multi-dimensional force amplitude data output by the non-directional force sensor converges to the preset force control target range. By dynamically adjusting the contact operation trajectory of the robot arm based on the updated closed-loop control parameters by the actuator, the force situation of the robot arm when grasping an object can be kept within the preset force control target range, thereby improving the stability and accuracy of the grasping.

[0089] As an embodiment, step S500 dynamically adjusts the contact operation trajectory of the target mechanical device according to the updated closed-loop control parameters by the actuator so that the multi-dimensional force amplitude data output by the non-directional force sensor converges to a preset force control target range. Specifically, the following steps S510-S560 may be included: Step S510: Calculate the real-time position deviation vector between the current contact position of the actuator and the expected contact position according to the proportional coefficient vector in the updated closed-loop control parameters.

[0090] The proportional coefficient vector in the updated closed-loop control parameters is the vector generated in step S400 and used to adjust the proportional relationship between the control variable and the error. The actuator's current contact position is the actuator's current position, while the desired contact position is the position the actuator should reach based on the job requirements. The real-time position deviation vector is the difference between the actuator's current contact position and the desired contact position, reflecting the deviation between the actuator's current position and the desired position.

[0091] For example, the coordinate information of the actuator's current contact position and the coordinate information of the desired contact position are obtained. For example, in three-dimensional space, the coordinates of the actuator's current contact position are (x1, y1, z1), and the coordinates of the desired contact position are (x2, y2, z2). Then, the real-time position deviation vector is calculated, namely (x2-x1, y2-y1, z2-z1). The real-time position deviation vector is proportionally adjusted based on the proportional coefficient vector in the updated closed-loop control parameters. For example, assuming the proportional coefficient vector is [p1, p2, p3], the adjusted real-time position deviation vector is [p1×(x2-x1), p2×(y2-y1), p3×(z2-z1)]. By calculating the real-time position deviation vector between the actuator's current contact position and the desired contact position and proportionally adjusting it based on the proportional coefficient vector, accurate deviation information can be provided for subsequent trajectory adjustments.

[0092] Step S520: performing a weighted integral operation on the historical accumulated value of the real-time position deviation vector in the continuous time window according to the integral coefficient vector in the updated closed-loop control parameters to generate an integral correction vector.

[0093] The integral coefficient vector in the updated closed-loop control parameters is the vector generated in step S400 for eliminating the system's steady-state error. The historical cumulative value of the real-time position deviation vector within a continuous time window is the cumulative sum of the elements of the real-time position deviation vector within a continuous time window. The weighted integral operation multiplies the elements of the real-time position deviation vector at each time point by the corresponding integral coefficients during the integration process and then accumulates the sum. The integral correction vector is a vector obtained through the weighted integral operation and is used to correct the actuator's motion trajectory and eliminate the system's steady-state error.

[0094] For example, historical data of the real-time position deviation vector is recorded within a continuous time window. For example, the time window is set to 10 sampling periods, and the real-time position deviation vector is recorded for each sampling period. Then, a weighted integral operation is performed on the historical cumulative values ​​of the real-time position deviation vector based on the integral coefficient vector in the updated closed-loop control parameters. By performing a weighted integral operation on the historical cumulative values ​​of the real-time position deviation vector within the continuous time window and generating an integral correction vector, steady-state errors during the robot arm's motion can be eliminated, making the robot arm's motion more accurate.

[0095] Assume that a robotic arm performs contact work in three-dimensional space, and a non-directional force sensor collects force signals in real time to adjust the robotic arm's motion. A continuous time window is set to five sampling periods, during which historical data of the real-time position deviation vector is recorded. The real-time position deviation vector has three dimensions corresponding to the X, Y, and Z directions in three-dimensional space. The integral coefficient vector in the updated closed-loop control parameters is generated in step S400 to eliminate the system's steady-state error. For example, after analyzing the system characteristics and operational requirements using an optimization algorithm (such as a genetic algorithm), the integral coefficient vector is determined to be [0.2, 0.3, 0.5]. This means that in the weighted integral operation, the weights of the elements of the real-time position deviation vector in the X direction are 0.2, 0.3 in the Y direction, and 0.5 in the Z direction. Assume that over five consecutive sampling periods, the elements of the real-time position deviation vector in the X direction are 1, -2, 3, -1, and 2; in the Y direction, 2, 1, -1, 3, and -2; and in the Z direction, -1, 2, 1, -3, and 2. Perform a weighted integral calculation on the historical accumulated values ​​in the X direction: 0.2 × (1-2 + 3-1 + 2) = 0.6; in the Y direction: 0.3 × (2+1-1 + 3-2) = 0.9; and in the Z direction: 0.5 × (-1+2+1-3+2) = 0.5. The resulting integral correction vector is [0.6, 0.9, 0.5], which is used to correct the robot's motion trajectory and eliminate the system's steady-state error.

[0096] Step S530: performing a sliding window differential operation on the rate of change of the real-time position deviation vector in adjacent sampling periods according to the differential coefficient vector in the updated closed-loop control parameters to generate a differential suppression vector.

[0097] The differential coefficient vector in the updated closed-loop control parameters is the vector generated in step S400 and used to predict error trends and make adjustments in advance. The rate of change of the real-time position deviation vector within adjacent sampling periods is the difference between the real-time position deviation vector and the sampling period interval. The sliding window differential operation calculates the rate of change of the real-time position deviation vector within a sliding window. The differential suppression vector, derived through the sliding window differential operation, is used to suppress rapid changes in actuator motion and improve system stability.

[0098] For example, set the size of a sliding window to, say, three sampling periods. In each sampling period, calculate the rate of change of the real-time position deviation vector within the adjacent sampling period. For example, assuming the real-time position deviation vector of the current sampling period is [(x2-x1)_n, (y2-y1)_n, (z2-z1)_n], the real-time position deviation vector of the previous sampling period is [(x2-x1)_(n-1), (y2-y1)_(n-1), (z2-z1)_(n-1)], and the time interval between sampling periods is Δt, then the rate of change of the real-time position deviation vector within the adjacent sampling period is [((x2-x1)_n-(x2-x1)_(n-1)) / Δt, ((y2-y1)_n-(y2-y1)_(n-1)) / Δt, ((z2-z1)_n-(z2-z1)_(n-1)) / Δt]. Then, the rate of change of the real-time position deviation vector is differentially adjusted based on the differential coefficient vector in the updated closed-loop control parameters. For example, assuming the differential coefficient vector is [d1, d2, d3], the differential suppression vector is [d1×((x2-x1)_n-(x2-x1)_(n-1)) / Δt, d2×((y2-y1)_n-(y2-y1)_(n-1)) / Δt, d3×((z2-z1)_n-(z2-z1)_(n-1)) / Δt]. By performing a sliding window differential operation on the rate of change of the real-time position deviation vector within adjacent sampling periods to generate the differential suppression vector, rapid changes during the robot arm's motion can be suppressed, thereby improving the stability of the robot arm's motion.

[0099] Step S540: Superimpose the real-time position deviation vector, the integral correction vector, and the differential suppression vector to generate a trajectory adjustment instruction, and convert the trajectory adjustment instruction into a drive signal of the actuator to drive the joint motor of the target mechanical device to perform corresponding angle adjustment or displacement adjustment.

[0100] The real-time position deviation vector is the deviation vector between the current contact position of the actuator and the expected contact position calculated in step S510, the integral correction vector is the vector generated in step S520 for eliminating the steady-state error of the system, and the differential suppression vector is the vector generated in step S530 for suppressing rapid changes in the actuator movement process. Vector superposition is to add or weighted sum the corresponding elements of these three vectors to obtain a new vector, namely the trajectory adjustment instruction. The trajectory adjustment instruction is an instruction for adjusting the contact operation trajectory of the target mechanical device. The actuator drive signal is the signal obtained after converting the trajectory adjustment instruction, which is used to drive the joint motor of the target mechanical device to perform the corresponding angle adjustment or displacement adjustment.

[0101] For example, the real-time position deviation vector, the integral correction vector and the differential suppression vector are vector-superimposed. Then, the trajectory adjustment instruction is converted into a driving signal for the actuator. For example, each element of the trajectory adjustment instruction is converted into a corresponding voltage or current value through a conversion function as a driving signal for the actuator. The actuator drives the joint motor of the target mechanical device to perform corresponding angle adjustment or displacement adjustment according to the driving signal, thereby adjusting the contact operation trajectory of the target mechanical device. By vector-superimposing the real-time position deviation vector, the integral correction vector and the differential suppression vector to generate a trajectory adjustment instruction, and converting it into a driving signal for the actuator, the contact operation trajectory of the robot arm can be accurately adjusted, so that the movement of the robot arm is more in line with the operation requirements.

[0102] As an embodiment, the process of generating the trajectory adjustment instruction further includes the following steps S541 to S545: S541: Real-time monitoring of whether the multi-dimensional force amplitude data output by the non-directional force sensor exceeds the boundary threshold of the force control target range.

[0103] The multi-dimensional force amplitude data output by the non-directional force sensor is force signal data collected in real time by the non-directional force sensor installed on the target mechanical device. The force control target interval is a pre-set force amplitude range, and the boundary thresholds are the upper and lower limits of the force control target interval. Real-time monitoring involves continuously acquiring the multi-dimensional force amplitude data output by the non-directional force sensor during the operation of the target mechanical device and comparing it with the boundary thresholds of the force control target interval. For example, the data acquisition system acquires the multi-dimensional force amplitude data output by the non-directional force sensor in real time. For example, in three-dimensional space, force amplitude data in the X, Y, and Z directions is acquired. The force amplitude data in each direction is then compared with the boundary thresholds of the force control target interval. If the force amplitude data in a particular direction is greater than the upper threshold or less than the lower threshold of the force control target interval, the force amplitude data in that direction is considered to have exceeded the boundary threshold of the force control target interval. By monitoring the multi-dimensional force amplitude data output by the non-directional force sensor in real time to see if it exceeds the boundary thresholds of the force control target interval, abnormal force conditions in the robotic arm can be detected promptly, providing a basis for subsequent emergency response.

[0104] S542: When it is detected that the boundary threshold is exceeded, an emergency braking instruction is generated according to the exceeding direction and amplitude, and the execution of the trajectory adjustment instruction is interrupted.

[0105] The exceedance direction is the direction in which the force amplitude data exceeds the threshold of the force control target interval boundary, such as the X, Y, or Z directions. The exceedance magnitude is the magnitude by which the force amplitude data exceeds the threshold of the force control target interval boundary. An emergency braking command is generated based on the exceedance direction and magnitude, and is used to immediately stop the movement of the target mechanical device to avoid possible danger or damage. Interrupting the execution of a trajectory adjustment command stops the currently executing trajectory adjustment command, causing the target mechanical device to pause.

[0106] For example, when it is detected that the multi-dimensional force amplitude data output by the non-directional force sensor exceeds the boundary threshold of the force control target interval, the direction of the excess must be accurately determined first. Taking the scenario of an industrial robot grasping an object as an example, assuming that the threshold range of the force control target interval in the X direction is [F x_min , F x_max ], the Y direction is [F y_min , F y_max ], in the Z direction is [F z_min ,F z_max ]. When the X-direction force amplitude F collected in real time x Greater than F x_max or less than F x_min When , it is determined that the X direction exceeds the boundary threshold, and the Y and Z directions can be determined similarly.

[0107] After determining the direction of the excess, the excess amplitude is calculated. The excess amplitude can be expressed as the difference between the actual force amplitude and the corresponding boundary threshold, for example, the excess amplitude ΔF in the X direction x =|F x -F x_max |(When F x > F x_max ) or ΔF x =|F x -F x_min |(When F x < F x_min hour).

[0108] An emergency braking command is generated based on the direction and magnitude of the overshoot. This command contains explicit stopping information to control the actuators of the target mechanical device to immediately stop. For example, for the joint motors of an industrial robot, the emergency braking command can be an electrical signal that quickly powers down the motors, thereby stopping the joint movement. The specific parameters of this electrical signal can be adjusted based on the overshoot magnitude. If the overshoot magnitude is large, a stronger braking signal may be required to ensure a quick stop of the mechanical device.

[0109] At the same time, the system will immediately interrupt the trajectory adjustment instruction being executed. This can be achieved through software-level control logic. For example, an interrupt flag can be set in the control system. When a boundary threshold is exceeded, the flag is set to an active state. Upon receiving the flag signal, the execution program of the trajectory adjustment instruction immediately stops sending and executing subsequent instructions, ensuring that the target mechanical device does not continue to move according to the original trajectory adjustment instruction, avoiding further danger or damage.

[0110] S543: After the emergency braking command takes effect, the real-time force signal sequence is re-collected and a fast recalculation process of the force trend prediction model is started.

[0111] After the emergency braking command takes effect, the target mechanical device stops moving. At this point, the real-time force signal sequence needs to be re-collected to obtain the force signal information in its current state. The re-collection process is similar to step S100. The force signal generated by the target mechanical device during contact operation is collected by the non-directional force sensor at a preset sampling frequency to form a new real-time force signal sequence.

[0112] Start the fast recalculation process of the force trend prediction model. Since the previous force trend prediction results may no longer be applicable due to the force amplitude exceeding the boundary threshold, a new prediction is required. The fast recalculation process is basically the same as the normal force trend prediction process, but has higher requirements on the calculation speed. For example, when performing sliding window segmentation processing on the real-time force signal sequence (step S210), the size of the sliding window can be appropriately reduced to reduce the amount of data processing, thereby speeding up the calculation speed. At the same time, in the calculation process of each layer of the force trend prediction model (such as the timing coding layer, the dynamic weight allocation layer, etc.), more efficient algorithms and optimization strategies can be used, such as using parallel computing technology to improve calculation efficiency.

[0113] In scenarios where an industrial robot is grasping an object, recollecting the real-time force signal sequence after emergency braking accurately determines the current force conditions on the robotic arm. Initiating a rapid recalculation process for the force trend prediction model quickly generates new force direction trend predictions, providing a basis for subsequent adjustments, shortening system response time, and improving system stability and reliability.

[0114] S544: Generate a new multi-dimensional force compensation signal based on the updated force direction trend prediction result obtained by rapid recalculation to cover the original compensation signal.

[0115] The updated force direction trend prediction result obtained by rapid recalculation is the result output by the force trend prediction model after the rapid recalculation process, indicating the future force change trend. Based on this updated result, a new multi-dimensional force compensation signal is generated according to the method of step S300 and its subsequent sub-steps.

[0116] First, a multidimensional force compensation signal is generated based on the matching degree between the updated force direction trend prediction result and the current closed-loop control parameters. Specifically, the proportional coefficient vector, integral coefficient vector, and differential coefficient vector from the current closed-loop control parameters are obtained and normalized to generate a closed-loop control parameter feature matrix (step S310). Similarity is then calculated between the vector distribution features corresponding to the updated force direction trend prediction result and the closed-loop control parameter feature matrix to generate a directional deviation feature vector (step S320). The directional deviation feature vector is then input into the fully connected layer of a preset compensation signal generation network to extract a preset oscillation component and a preset offset component (step S330). The preset oscillation component and the preset offset component are then cross-channel fused using the convolution kernel group of the compensation signal generation network to generate a preliminary compensation signal (step S340). During the generation of the preliminary compensation signal, certain constraints must be adhered to. For example, if the force amplitudes in at least two dimensions of the updated force direction trend prediction result change in opposite directions, the compensation signal amplitudes in those two dimensions are preferentially suppressed. Finally, the preliminary compensation signal is input into the activation function layer of the compensation signal generation network for nonlinear transformation to generate a new multi-dimensional force compensation signal (step S350).

[0117] The generated multi-dimensional force compensation signal overwrites the original compensation signal. This is because the original compensation signal was generated based on the previous force trend prediction results. Since the current situation has changed, the new compensation signal can more accurately reflect the current force trend, thereby better correcting the directional deviation component in the current closed-loop control parameters. In the scenario of an industrial robot grasping an object, the new multi-dimensional force compensation signal enables the robot arm to better adapt to the actual force changes during subsequent movements, improving the accuracy and stability of the grasp.

[0118] S545: After the new multi-dimensional force compensation signal takes effect, the emergency braking command is released and the execution of the trajectory adjustment command is resumed.

[0119] When the new multi-dimensional force compensation signal takes effect, it indicates that the system has generated an appropriate compensation signal based on the current force trend and has begun adjusting the closed-loop control parameters. At this point, the emergency brake command can be released and the movement of the target mechanical device can be resumed.

[0120] Releasing the emergency brake command can be implemented through software-level control logic, for example by resetting a previously set interrupt flag, allowing the control system to continue sending and executing trajectory adjustment commands. After resuming execution of the trajectory adjustment command, the actuator dynamically adjusts the contact trajectory of the target mechanical device based on the updated closed-loop control parameters. The specific process includes calculating the real-time position deviation vector between the actuator's current contact position and the desired contact position based on the proportional coefficient vector in the updated closed-loop control parameters (step S510), performing a weighted integration operation on the historical accumulated value of the real-time position deviation vector within a continuous time window based on the integral coefficient vector to generate an integral correction vector (step S520), performing a sliding window differentiation operation on the rate of change of the real-time position deviation vector within adjacent sampling periods based on the differential coefficient vector to generate a differential inhibition vector (step S530), and then superimposing the real-time position deviation vector, the integral correction vector, and the differential inhibition vector to generate a trajectory adjustment command. The trajectory adjustment command is then converted into a drive signal for the actuator, which drives the joint motor of the target mechanical device to perform the corresponding angle adjustment or displacement adjustment (step S540).

[0121] In the scenario where an industrial robot grasps an object, the emergency braking command is released and the execution of the trajectory adjustment command is resumed, so that the robotic arm can continue to move according to the adjusted trajectory and complete the grasping task more accurately under the action of the new multi-dimensional force compensation signal.

[0122] Step S550: Monitor in real time whether the multi-dimensional force amplitude data output by the non-directional force sensor enters the force control target range. If not, dynamically adjust the gain weight of the proportional-integral-differential coefficient in the updated closed-loop control parameter according to the boundary distance between the multi-dimensional force amplitude data and the force control target range to generate a secondary corrected closed-loop control parameter.

[0123] The multi-dimensional force amplitude data output by the non-directional force sensor is monitored in real time to see if it enters the force control target range. This process is similar to step S541. The multi-dimensional force amplitude data output by the non-directional force sensor is acquired in real time by the data acquisition system and compared with the boundary threshold of the force control target range. If the force amplitude data for all dimensions is within the force control target range, the multi-dimensional force amplitude data is considered to have entered the force control target range; otherwise, it is considered not to have entered the force control target range.

[0124] If the multi-dimensional force amplitude data does not enter the force control target range, it is necessary to dynamically adjust the gain weight of the proportional-integral-differential coefficient in the updated closed-loop control parameters according to the boundary distance between the multi-dimensional force amplitude data and the force control target range. For example, in the X direction, if the force amplitude data F x Greater than the upper threshold F of the force control target range x_max , calculate the boundary distance ΔF x =F x -Fx_max If F x Less than the lower threshold F x_min , calculate the boundary distance ΔF x =F x_min -F x According to this boundary distance, the gain weight of the proportional-integral-differential coefficient is dynamically adjusted.

[0125] For the proportional coefficient, if the boundary distance is large, it means that the current control force may not be enough, and the gain weight of the proportional coefficient needs to be increased to enhance the proportional relationship between the control amount and the error, so that the actuator can adjust the position faster. For example, let the original proportional coefficient be p and the gain weight adjustment coefficient be k p , the adjusted proportional coefficient is p'=p×k p , where k p The adjustment can be linear or nonlinear depending on the size of the boundary distance.

[0126] For the integral coefficient, if the boundary distance persists, it means that the system may have a steady-state error, and it is necessary to increase the gain weight of the integral coefficient to speed up the accumulation of the integral term and eliminate the steady-state error. Similarly, let the original integral coefficient be i and the gain weight adjustment coefficient be k i , the adjusted integral coefficient is i'=i×k i .

[0127] For the differential coefficient, if the boundary distance changes quickly, it means that the dynamic changes of the system are large, and the gain weight of the differential coefficient needs to be increased to better predict the change trend of the error and make adjustments in advance. Let the original differential coefficient be d and the gain weight adjustment coefficient be k d , the adjusted differential coefficient is d'=d×k d .

[0128] The adjusted proportional, integral, and differential coefficients are vectored together to generate the quadratic correction closed-loop control parameters. By dynamically adjusting the gain weights of the proportional-integral-differential coefficients, the quadratic correction closed-loop control parameters can be generated. This allows the robot arm to more effectively adjust its motion trajectory when the multi-dimensional force amplitude data does not enter the force control target range, so that the force amplitude data converges within the force control target range as quickly as possible.

[0129] Step S560: Inputting the secondary corrected closed-loop control parameters into the actuator to trigger compensation adjustment of the contact operation trajectory of the target mechanical device until the multi-dimensional force amplitude data remains stable within the force control target range for a preset time period.

[0130] The secondary corrected closed-loop control parameters are input into the actuator, which drives the joint motors of the target mechanical device to perform corresponding angle or displacement adjustments based on these parameters, thereby triggering compensation adjustments to the contact working trajectory of the target mechanical device. The compensation adjustment process is similar to step S540, including calculating the real-time position deviation vector based on the proportional coefficient vector in the secondary corrected closed-loop control parameters, generating an integral correction vector based on the integral coefficient vector, and generating a differential suppression vector based on the differential coefficient vector. These vectors are then superimposed to generate a trajectory adjustment command, which is then converted into a drive signal for the actuator.

[0131] During the compensation adjustment process, the multi-dimensional force amplitude data output by the non-directional force sensor is continuously monitored in real time to ensure it is within the force control target range. When the multi-dimensional force amplitude data enters the force control target range, a timer begins to determine whether it remains stable for a preset duration. The preset duration is a pre-set time value, such as 5 seconds, to ensure that the multi-dimensional force amplitude data remains stable within the force control target range, avoiding misjudgments due to brief fluctuations.

[0132] If the multi-dimensional force amplitude data remains stable within the force control target range for a preset duration, the target mechanical device's contact trajectory is considered successfully adjusted, and the system achieves the desired force control effect. By inputting secondary correction closed-loop control parameters into the actuator, compensatory adjustments to the contact trajectory are made until the multi-dimensional force amplitude data remains stable within the force control target range for a preset duration. This ensures that the force applied to the robotic arm during grasping remains within a preset range, improving grasping quality and stability.

[0133] In one embodiment, the training process of the pre-trained force trend prediction model includes the following steps S10 to S50: Step S10: Collect multiple sets of historical force signal sequences and corresponding actual force direction change data of mechanical devices under various contact working conditions to construct a training data set.

[0134] Mechanical devices generate different force signals under various contact operating conditions. These force signals reflect the force conditions under these conditions. A historical force signal sequence is a sequence of force signal data collected over a period of time. Actual force direction change data is the actual change in force direction corresponding to the historical force signal sequence.

[0135] The process for collecting multiple sets of historical force signal sequences and the corresponding actual force direction change data is as follows: different types of mechanical devices, such as industrial robots and automated processing equipment, are selected and multiple contact operation conditions are set. For industrial robots, different grasping tasks can be set, such as grasping objects of different shapes and weights; for automated processing equipment, different processing processes can be set, such as grinding and drilling. Under each contact operation condition, a non-directional force sensor is used to collect the force signals generated by the mechanical device during the operation. The signals are sampled according to the sampling frequency to form a historical force signal sequence. At the same time, the actual changes in force direction are recorded through other measurement methods (such as high-precision angle sensors and displacement sensors) to form the corresponding actual force direction change data.

[0136] The multiple sets of historical force signal sequences and the corresponding actual force direction change data were collated and annotated to construct a training dataset. This dataset is used to train the force trend prediction model. It contains a wealth of information about force signals and force direction changes, enabling the model to learn the patterns of force variation under different contact working conditions.

[0137] Step S20: performing time alignment and noise filtering processing on the historical force signal sequence to generate a standardized force signal sample.

[0138] Time alignment accurately arranges the data points in the historical force signal sequence in chronological order, ensuring that each data point corresponds to the correct timestamp. Noise filtering removes noise interference from the historical force signal sequence, making the signal purer and facilitating subsequent analysis and processing. Standardized force signal samples, obtained after time alignment and noise filtering, have a uniform time scale and low noise level, which helps improve the training effect of the force trend prediction model.

[0139] As an implementation method, step S20, performing time alignment and noise filtering on the historical force signal sequence to generate a standardized force signal sample, may specifically include the following steps: Step S21: performing baseline drift correction on the historical force signal sequence to eliminate the preset interference component caused by the sensor zero point offset.

[0140] Sensor zero offset occurs when the sensor outputs a non-zero signal value when no force is applied. This zero offset can introduce a pre-determined interference component into the historical force signal sequence, affecting signal accuracy. Baseline drift correction aims to eliminate this pre-determined interference component and restore the signal baseline to its correct position.

[0141] For example, the historical force signal sequence is first analyzed for predetermined characteristics, and a filtering algorithm (such as a low-pass filter) is used to extract predetermined components. The mean or median of these predetermined components is then calculated and used as an estimate of baseline drift. Finally, the baseline drift estimate is subtracted from each data point in the historical force signal sequence to achieve baseline drift correction.

[0142] Step S22: Perform sliding window variance analysis on the corrected historical force signal sequence to identify abnormal noise segments with sudden changes in force amplitude within the window.

[0143] Sliding window variance analysis is a method used to analyze the local variation characteristics of a signal. It slides a fixed-size window across the signal and calculates the variance of the signal within the window. The variance reflects the degree of signal fluctuation within the window, and the variance increases significantly when the amplitude of the force within the window changes suddenly.

[0144] For example, set the size of a sliding window, for example, a window size of 20 data points. Starting from the starting position of the corrected historical force signal sequence, take the 20 data points covered by the window as a window and calculate the variance of the signal within the window. Then slide the window back one data point, calculate the variance of the signal within the new window again, and so on, until it slides to the end of the sequence. Set a variance threshold. When the variance within a certain window exceeds the threshold, it is considered that there is an abnormal noise segment with a sudden change in force amplitude within the window. For example, set the variance threshold to 1. When the variance calculation result within a certain window is 1.5, the window is identified as an abnormal noise segment.

[0145] Step S23: Adopting an adaptive filtering algorithm to perform local smoothing on the abnormal noise segment, and retaining the dynamic characteristics of the non-noise segment in the original signal.

[0146] Adaptive filtering is an algorithm that automatically adjusts filtering parameters based on the local characteristics of a signal. It can remove noise while preserving the signal's dynamic characteristics. The goal of local smoothing of abnormal noise segments is to eliminate the abnormal noise and smooth the signal without affecting the dynamic changes of the non-noise segments in the original signal.

[0147] For example, an adaptive filtering algorithm is employed to process identified abnormal noise segments. For example, an adaptive median filter algorithm is used, which adaptively adjusts the size of the filter window and the filter coefficient based on the median value of the data within the window and the distribution of neighboring data. For each data point in the abnormal noise segment, an appropriate filter coefficient is selected based on the median value and distribution of its neighboring data, resulting in a smoothed signal value. During this processing, only the abnormal noise segments are processed, while the data in the non-noise segments of the original signal remains unchanged, thereby preserving the dynamic characteristics of the original signal.

[0148] Step S24: resampling the smoothed historical force signal sequence with a timestamp so that its sampling interval is consistent with the recording frequency of the actual force direction change data.

[0149] Timestamp resampling adjusts the sampling interval of the smoothed historical force signal sequence to align with the recording frequency of the actual force direction change data, facilitating subsequent matching and analysis. The recording frequency of the actual force direction change data may differ from the original sampling frequency of the historical force signal sequence. Timestamp resampling ensures temporal alignment between the two.

[0150] For example, determine the recording frequency of the actual force direction change data, for example, the recording frequency is 10 times per second. According to this recording frequency, the smoothed historical force signal sequence is timestamped and resampled. The linear interpolation method can be used. In the case where there is no original data at the resampling time point, linear interpolation calculation is performed based on the adjacent original data points to obtain the signal value at the resampling time point. For example, there is no original data at the resampling time point t, and the adjacent original data points are t1 and t2, and the corresponding signal values ​​are x1 and x2. Then the signal value x at the resampling time point t can be calculated using the linear interpolation formula x=x1+(x2-x1)×(t-t1) / (t2-t1).

[0151] Step S25: Normalize the amplitude of the resampled historical force signal sequence by dimension, so that the force amplitudes of each dimension are time-series aligned under the same dimension, and generate standardized force signal samples.

[0152] Amplitude normalization is to convert the force amplitude of each dimension in the resampled historical force signal sequence into a predefined range, usually [0, 1] or [-1, 1], to eliminate the dimensional differences between different dimensions and align the force amplitudes of each dimension in time series under the same dimension.

[0153] For example, for each dimension of the resampled historical force signal sequence, find the maximum and minimum values ​​on that dimension. For example, for the force amplitude sequence in the X direction, find the maximum value F x_max and the minimum value F x_min Then, for each data point F in this dimension xi , using the normalization formula F xi_normalized =(F xi -F x_min ) / (F x_max -F x_min ) to the range of [0, 1]. The same normalization process is performed on the force amplitude sequences of all dimensions to obtain the historical force signal sequence after amplitude normalization by dimension, that is, the standardized force signal sample.

[0154] Step S30: inputting the standardized force signal sample into the initial force trend prediction model to obtain the predicted force direction trend result output by the initial force trend prediction model.

[0155] The initial force trend prediction model is an untrained or preliminarily trained model, requiring training to optimize its parameters and improve prediction accuracy. Normalized force signal samples are force signal samples that have undergone time alignment, noise filtering, and amplitude normalization. They contain the force signal characteristics of the mechanical device under different contact working conditions.

[0156] After inputting the standardized force signal samples into the initial force trend prediction model, the model processes and analyzes the input signals and, using its internal algorithms and structures, calculates a predicted force direction trend. This predicted force direction trend represents the model's prediction of future force direction trends, expressed in terms of vector distribution features. For example, in a scenario where an industrial robot is grasping an object, the initial force trend prediction model uses the input standardized force signal samples to predict the force trends of the robot arm in different directions over a period of time and outputs the predicted force direction trend results.

[0157] Step S40: Calculate the trend error loss between the predicted force direction trend result and the actual force direction change data, and calculate the closed-loop control error loss of the predicted force direction trend result in a closed-loop control simulation environment.

[0158] The trend error loss is the difference between the predicted force direction trend and the actual force direction change data, reflecting the model's accuracy in predicting the force direction trend. The closed-loop control error loss is the control error of the predicted force direction trend in the closed-loop control simulation environment, reflecting the impact of the model's prediction results on the performance of the closed-loop control system.

[0159] As an embodiment, step S40, calculating the closed-loop control error loss of the predicted force direction trend result in the closed-loop control simulation environment, may specifically include the following steps: Step S41: inputting the predicted force direction trend result into a preset closed-loop control parameter response simulator to generate simulated force control trajectory data including multi-dimensional force adjustment instructions.

[0160] The preset closed-loop control parameter response simulator is a pre-designed simulator that simulates the response of the closed-loop control system based on the input predicted force direction trend results, generating simulated force control trajectory data containing multi-dimensional force adjustment instructions. This simulated force control trajectory data corresponds to the multi-dimensional force adjustment instructions generated based on the predicted force direction trend results within the closed-loop control simulation environment. It contains information about the multi-dimensional force amplitudes and corresponding contact trajectory changes at different time points.

[0161] For example, the predicted force direction trend results are input into a preset closed-loop control parameter response simulator. The simulator simulates the operation of the closed-loop control system based on the predicted force direction trend results and the preset closed-loop control parameters (such as the proportional coefficient, integral coefficient, and differential coefficient). During the simulation, corresponding multi-dimensional force adjustment instructions are generated based on the predicted force direction trend. These instructions are used to adjust the motion trajectory of the mechanical device. The multi-dimensional force amplitudes and corresponding contact trajectory change information at different time points during the simulation are recorded to form simulated force control trajectory data. For example, the preset closed-loop control parameter response simulator simulates the motion process of the robotic arm under closed-loop control based on the predicted force direction trend results, generating simulated force control trajectory data containing force adjustment instructions in the three directions of X, Y, and Z.

[0162] Step S42: extracting real force control trajectory data synchronized with the actual force direction change data from the training data set, where the real force control trajectory data includes a time series of multi-dimensional force amplitudes and corresponding contact trajectory change vectors.

[0163] Real force control trajectory data is collected during actual contact operations and reflects the force conditions and motion trajectory of the mechanical device during actual operation. Real force control trajectory data, synchronized with the actual force direction change data, is extracted from the training dataset to compare the simulated force control trajectory data with the real force control trajectory data and calculate the closed-loop control error loss.

[0164] For example, in the training dataset, the time range corresponding to the actual force direction change data is found, and the real force control trajectory data within this time range is extracted. The real force control trajectory data contains a time series of multi-dimensional force amplitudes, that is, the force amplitudes in different dimensions at each time point, and the corresponding contact trajectory change vectors, that is, the position and posture change information of the mechanical device at each time point. For example, in a scenario where an industrial robot grasps an object, the real force control trajectory data synchronized with the actual force direction change data is extracted from the training dataset, including the time series of the force amplitudes of the robotic arm in the X, Y, and Z directions, as well as contact trajectory change information such as the angle change vectors of the robotic arm joints.

[0165] Step S43: aligning the simulated force control trajectory data with the real force control trajectory data in a windowed manner according to the time dimension, and extracting the instantaneous fluctuation difference and cumulative trend deviation of the multi-dimensional force amplitude in each alignment window.

[0166] Windowed alignment divides the simulated and real force trajectory data into aligned windows based on time. Each window contains the simulated and real force trajectory data for the same timeframe. The instantaneous fluctuation difference is the difference in multi-dimensional force amplitude at each time point between the simulated and real force trajectory data within each alignment window. The cumulative trend deviation is the difference in the cumulative trend of multi-dimensional force amplitude between the simulated and real force trajectory data within each alignment window.

[0167] For example, a window size is set, for example, the window size is 10 time points. Starting from the starting time point of the simulated force-controlled trajectory data and the real force-controlled trajectory data, the data of the 10 time points covered by the window are used as an alignment window. For each time point in each alignment window, the difference between the multi-dimensional force amplitudes in the simulated force-controlled trajectory data and the real force-controlled trajectory data is calculated to obtain the instantaneous fluctuation difference. At the same time, the cumulative change of the multi-dimensional force amplitude in the simulated force-controlled trajectory data and the real force-controlled trajectory data in the window is calculated, and the difference between the two is calculated to obtain the cumulative trend deviation. For example, in an alignment window, for the force amplitude in the X direction, the difference between the simulated force amplitude and the real force amplitude at each time point is calculated to obtain the instantaneous fluctuation difference in the X direction; the cumulative change of the simulated force amplitude and the cumulative change of the real force amplitude in the window are calculated, and the difference between the two is calculated to obtain the cumulative trend deviation in the X direction.

[0168] Step S44: Calculate the dynamic tracking error of the force amplitude in each dimension by the instantaneous fluctuation difference, and calculate the steady-state offset of the force amplitude in each dimension by the cumulative trend deviation.

[0169] The dynamic tracking error is the error in the model's ability to dynamically track force amplitude changes, as reflected by the instantaneous fluctuation differences in the force amplitudes of each dimension within each alignment window. The steady-state offset is the error in the model's ability to control steady-state forces, as reflected by the cumulative trend deviations in the force amplitudes of each dimension within each alignment window. For example, the average sum of the squares of the instantaneous fluctuation differences in the force amplitudes of each dimension within each alignment window is calculated to obtain the dynamic tracking error of the force amplitude in that dimension. For example, for the force amplitude in the X direction, the instantaneous fluctuation difference at 10 time points within an alignment window is ΔF x_1 , ΔF x_2 ,..., ΔF x_10 , then the dynamic tracking error E in the X direction xd ynamic=(ΔF x_1 2 +ΔF x_2 2 +...+ΔF x_10 2) / 10. For the cumulative trend deviation of the force amplitude of each dimension within each alignment window, its absolute value is taken as the steady-state offset of the force amplitude of that dimension.

[0170] Step S45: linearly superimpose the dynamic tracking error and the steady-state offset according to a preset ratio to generate a closed-loop control error loss that represents the direction prediction accuracy and control stability.

[0171] The preset ratio is a pre-set proportional coefficient used to adjust the weight of the dynamic tracking error and steady-state offset in the closed-loop control error loss. Linear superposition is to multiply the dynamic tracking error and steady-state offset by the preset ratio respectively, and then add them to obtain the closed-loop control error loss. The closed-loop control error loss characterizes the directional prediction accuracy and control stability of the model's predicted force direction trend results in the closed-loop control simulation environment. For example, let the preset ratios be α and β, where α represents the weight of the dynamic tracking error, β represents the weight of the steady-state offset, and α+β=1. For the force amplitude of each dimension, the dynamic tracking error E of that dimension is calculated. d ynamic multiplied by α, steady-state offset E steady Multiply by β and then add to get the closed-loop control error loss E of this dimension closed_loop .

[0172] Step S50: Fusion trend error loss and closed-loop control error loss to generate a total training loss value, and update the parameters of the initial force trend prediction model through the back propagation algorithm until the total training loss value converges.

[0173] The trend error loss reflects the model's accuracy in predicting force direction trends, while the closed-loop control error loss reflects the impact of the model's predictions on the performance of the closed-loop control system. Combining the trend error loss and closed-loop control error loss to generate the total training loss comprehensively considers the model's prediction accuracy and control performance.

[0174] For example, set a fusion coefficient γ, where γ represents the weight of the trend error loss and 1-γ represents the weight of the closed-loop control error loss. Multiply the trend error loss by γ, multiply the closed-loop control error loss by 1-γ, and then add them together to get the total training loss value. For example, the total training loss value L total =γ×L trend +(1-γ)×L closed_loop , where L trend represents the trend error loss, L closed_loop Represents the closed-loop control error loss.

[0175] The parameters of the initial force trend prediction model are updated using a backpropagation algorithm. The backpropagation algorithm calculates the gradients of the model parameters based on the total training loss, and then updates the model parameters based on the gradients, gradually reducing the total training loss. In each iteration, the standardized force signal samples are input into the initial force trend prediction model to calculate the predicted force direction trend. The trend error loss and closed-loop control error loss are then calculated and integrated to obtain the total training loss. The backpropagation algorithm then calculates the gradients of the total training loss with respect to the model parameters, and the model parameters are updated based on the gradients. This process is repeated until the total training loss converges, meaning that the total training loss no longer changes significantly over multiple iterations. At this point, the parameters of the initial force trend prediction model are considered optimized, and training is complete, resulting in a pretrained force trend prediction model. In summary, by acquiring real-time force signal sequences and using the pretrained force trend prediction model to predict force direction trends, a multidimensional force compensation signal is generated to adjust the closed-loop control parameters. Ultimately, dynamic adjustment of the contact trajectory of the target mechanical device is achieved, ensuring that the multidimensional force amplitude data output by the non-directional force sensor converges to the preset force control target range. At the same time, by training the force trend prediction model, the prediction accuracy and control performance of the model are improved, ensuring the stability and reliability of the entire closed-loop control system.

[0176] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. The computer system is a system connected to or built into a target mechanical device. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 may be connected via a bus or other means. The processor 101 (also known as the Central Processing Unit (CPU)) is the computing and control core of the computer system, capable of parsing various instructions within the computer system and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi or a mobile communication interface), which can be used to send and receive data under the control of the processor 101. The communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system for storing programs and data. It is understood that the memory 103 here can include both the built-in memory of the computer system and, of course, the extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system, but this is not limited to this in the present invention. In one embodiment, the processor 101 executes the non-directional force sensor closed-loop control method based on real-time force trend prediction provided in the above embodiment of the present invention by running the computer program in the memory 103 .

Claims

1. A closed-loop control method for a non-directional force sensor based on real-time force trend prediction, characterized in that: The method comprises: collecting a real-time force signal sequence generated when the target mechanical device performs a contact operation, wherein the real-time force signal sequence includes multi-dimensional force amplitude data output by a non-directional force sensor; Inputting the real-time force signal sequence into a pre-trained force trend prediction model to obtain a force direction trend prediction result output by the force trend prediction model, wherein the force direction trend prediction result is used to characterize the vector distribution characteristics of the multi-dimensional force amplitude change within a future preset time window; generating a multi-dimensional force compensation signal according to a degree of matching between the force direction trend prediction result and the current closed-loop control parameters, wherein the multi-dimensional force compensation signal is used to correct a directional deviation component in the current closed-loop control parameters; adjusting the proportional-integral-derivative coefficients in the current closed-loop control parameters based on the multi-dimensional force compensation signal to generate updated closed-loop control parameters, and inputting the updated closed-loop control parameters into the actuator of the target mechanical device; The actuator dynamically adjusts the contact operation trajectory of the target mechanical device according to the updated closed-loop control parameters, so that the multi-dimensional force amplitude data output by the non-directional force sensor converges to a preset force control target range.

2. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 1, characterized in that: Inputting the real-time force signal sequence into a pre-trained force trend prediction model to obtain a force direction trend prediction result output by the force trend prediction model includes: Performing a sliding window segmentation process on the real-time force signal sequence to generate an input data pair including a historical force signal subsequence and a current force signal subsequence; Inputting the historical force signal subsequence into the time series encoding layer of the force trend prediction model, extracting the time series correlation features corresponding to the historical force signal subsequence, wherein the time series correlation features include force amplitude fluctuation patterns in different time dimensions; Inputting the current force signal subsequence into the dynamic weight allocation layer of the force trend prediction model, and generating a weight allocation vector corresponding to the time series correlation feature according to the force amplitude change rate at each time point in the current force signal subsequence; Performing a dot product operation on the time series correlation feature and the weight distribution vector through the feature fusion layer of the force trend prediction model to generate a dynamic weighted time series feature; The dynamic weighted time series features are input into the trend prediction layer of the force trend prediction model, the multidimensional force amplitude prediction values ​​at each time point in the future preset time window are output, and the force direction trend prediction result is generated based on the vector change direction of the multidimensional force amplitude prediction value.

3. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 2, characterized in that: Generating a multi-dimensional force compensation signal according to the matching degree between the force direction trend prediction result and the current closed-loop control parameters includes: Obtaining a proportional coefficient vector, an integral coefficient vector, and a differential coefficient vector in the current closed-loop control parameters, and normalizing the proportional coefficient vector, the integral coefficient vector, and the differential coefficient vector to generate a closed-loop control parameter characteristic matrix; Performing similarity calculation on the vector distribution characteristics corresponding to the force direction trend prediction result and the closed-loop control parameter characteristic matrix to generate a direction deviation characteristic vector; Inputting the directional deviation feature vector into a fully connected layer of a preset compensation signal generation network, and extracting a preset oscillation component and a preset offset component from the directional deviation feature vector; Performing cross-channel fusion of the preset oscillation component and the preset offset component through the convolution kernel group of the compensation signal generation network to generate a preliminary compensation signal; The preliminary compensation signal is input into the activation function layer of the compensation signal generation network for nonlinear transformation to generate the multi-dimensional force compensation signal, wherein the output range of the activation function layer matches the adjustable range of the closed-loop control parameter.

4. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 3, characterized in that: The convolution kernel group of the compensation signal generation network performs cross-channel fusion on the preset oscillation component and the preset offset component to generate a preliminary compensation signal, including: Splitting the preset oscillation component and the preset offset component into multiple parallel processing channels along the channel dimension to generate a preset oscillation component sub-channel group and a preset offset component sub-channel group; Decomposing the convolution kernel group into a first sub-convolution kernel set corresponding to the preset oscillation component sub-channel group and a second sub-convolution kernel set corresponding to the preset offset component sub-channel group; performing a multi-scale convolution operation on the preset oscillation component sub-channel group using the first sub-convolution kernel set, extracting time domain response features corresponding to different frequency fluctuation modes in the preset oscillation component sub-channel group, and generating a preset fusion feature map; Performing a long-period convolution operation on the preset offset component sub-channel group using the second sub-convolution kernel set, extracting cumulative effect features corresponding to trend offsets in the preset offset component sub-channel group, and generating a preset fusion feature map; Perform feature splicing on the preset fusion feature map and the preset fusion feature map along the spatial dimension to generate a cross-channel fusion feature tensor; Performing cross-layer feature enhancement on the cross-channel fusion feature tensor through the residual connection layer of the compensation signal generation network to generate enhanced fusion features; Inputting the enhanced fusion feature into the depthwise separable convolutional layer of the compensation signal generation network to perform feature dimensionality reduction, thereby generating a fusion feature vector after dimensionality reduction; A nonlinear mapping function is used to perform amplitude constraint and phase alignment on the fusion feature vector after dimension reduction, so as to generate a preliminary compensation signal that matches the dimension of the closed-loop control parameter feature matrix.

5. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 3, characterized in that: The adjusting the proportional-integral-differential coefficients in the current closed-loop control parameters based on the multi-dimensional force compensation signal includes: Decomposing the multi-dimensional force compensation signal into a proportional compensation term, an integral compensation term, and a differential compensation term, and performing superposition operations with the proportional coefficient vector, the integral coefficient vector, and the differential coefficient vector in the current closed-loop control parameters, respectively; Performing gradient constraint processing on the superimposed proportional coefficient vector so that the modulus of the adjusted proportional coefficient vector does not exceed a preset maximum proportional gain threshold; The integral coefficient vector after superposition is subjected to integral saturation suppression processing, and the integral accumulation rate is limited by dynamically adjusting the integral time constant; The superimposed differential coefficient vector is subjected to noise filtering, and the sliding average algorithm is used to eliminate the influence of noise on the differential term; The processed proportional coefficient vector, integral coefficient vector and differential coefficient vector are vector-concatenated to generate the updated closed-loop control parameters.

6. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 5, characterized in that: The step of dynamically adjusting the contact operation trajectory of the target mechanical device according to the updated closed-loop control parameters by the actuator so that the multi-dimensional force amplitude data output by the non-directional force sensor converges to a preset force control target range includes: Calculating a real-time position deviation vector between a current contact position of the actuator and a desired contact position according to the proportional coefficient vector in the updated closed-loop control parameters; performing a weighted integral operation on the historical accumulated value of the real-time position deviation vector within a continuous time window according to the integral coefficient vector in the updated closed-loop control parameter to generate an integral correction vector; performing a sliding window differential operation on the rate of change of the real-time position deviation vector in adjacent sampling periods according to the differential coefficient vector in the updated closed-loop control parameter to generate a differential suppression vector; Performing vector superposition of the real-time position deviation vector, the integral correction vector, and the differential suppression vector to generate a trajectory adjustment instruction, and converting the trajectory adjustment instruction into a drive signal of the actuator to drive the joint motor of the target mechanical device to perform corresponding angle adjustment or displacement adjustment; monitoring in real time whether the multi-dimensional force amplitude data output by the non-directional force sensor enters the force control target interval; if not, dynamically adjusting the gain weights of the proportional-integral-differential coefficients in the updated closed-loop control parameters according to the boundary distance between the multi-dimensional force amplitude data and the force control target interval to generate a secondary corrected closed-loop control parameter; The secondary corrected closed-loop control parameters are input into the actuator to trigger compensation adjustment of the contact operation trajectory of the target mechanical device until the multi-dimensional force amplitude data remains stable within the force control target range for a preset period of time.

7. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 1, characterized in that: The training process of the pre-trained force trend prediction model includes: Collect historical force signal sequences and corresponding actual force direction change data of multiple sets of mechanical devices under various contact working conditions to construct a training data set; Performing time alignment and noise filtering on the historical force signal sequence to generate a standardized force signal sample; Inputting the standardized force signal sample into an initial force trend prediction model to obtain a predicted force direction trend result output by the initial force trend prediction model; Calculating a trend error loss between the predicted force direction trend result and the actual force direction change data, and calculating a closed-loop control error loss of the predicted force direction trend result in a closed-loop control simulation environment; The trend error loss and the closed-loop control error loss are integrated to generate a total training loss value, and the parameters of the initial force trend prediction model are updated through a back propagation algorithm until the total training loss value converges.

8. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 7, characterized in that: The calculating of the closed-loop control error loss of the predicted force direction trend result in a closed-loop control simulation environment includes: Inputting the predicted force direction trend result into a preset closed-loop control parameter response simulator to generate simulated force control trajectory data including multi-dimensional force adjustment instructions; Extracting real force control trajectory data synchronized with actual force direction change data from the training data set, the real force control trajectory data comprising a time series of multi-dimensional force amplitudes and corresponding contact trajectory change vectors; Window-aligning the simulated force control trajectory data with the real force control trajectory data according to the time dimension, and extracting the instantaneous fluctuation difference and cumulative trend deviation of the multi-dimensional force amplitude in each alignment window; Calculating the dynamic tracking error of the force amplitude in each dimension by using the instantaneous fluctuation difference, and calculating the steady-state offset of the force amplitude in each dimension by using the cumulative trend deviation; The dynamic tracking error and the steady-state offset are linearly superimposed according to a preset ratio to generate a closed-loop control error loss that characterizes the direction prediction accuracy and control stability.

9. The closed-loop control method of a non-directional force sensor based on real-time force trend prediction according to claim 7, characterized in that: The performing time alignment and noise filtering on the historical force signal sequence to generate a standardized force signal sample includes: Performing baseline drift correction on the historical force signal sequence to eliminate a preset interference component caused by a zero point offset of the sensor; Perform sliding window variance analysis on the corrected historical force signal sequence to identify abnormal noise segments with sudden force amplitude changes within the window; Adopting an adaptive filtering algorithm to perform local smoothing on the abnormal noise segment, thereby retaining the dynamic characteristics of the non-noise segment in the original signal; Performing time stamp resampling on the smoothed historical force signal sequence so that its sampling interval is consistent with the recording frequency of the actual force direction change data; The resampled historical force signal sequence is amplitude normalized according to the dimension, so that the force amplitudes of each dimension are time-series aligned under the same dimension, thereby generating the standardized force signal sample.

10. A computer system, characterized in that: include: a memory storing a computer program; A processor is used to load the computer program to implement the closed-loop control method of a non-directional force sensor based on real-time force trend prediction as described in any one of claims 1 to 9.

Images