Method for dynamic tension compensation in a liquid bag cutting process

Abstract

The application provides a liquid bag shearing process dynamic tension compensation method, and belongs to the technical field of liquid bags. The application deploys a flow sliding window high-frequency sampling unit on the upper part of a liquid bag shearing device gripper, collects material elongation, instantaneous acceleration and three-dimensional transient tension data, constructs a liquid bag three-dimensional transient tension characteristic tensor, inputs a mixed expert routing tension prediction network, uses spatial orthogonal decoupling, a Fibonacci golden section dynamic routing optimization algorithm and a viscoelastic physical property constraint equation to drive a gated routing, activates the first two expert networks and outputs a tension prediction compensation amount, simultaneously uses a high-dimensional data lossless dimension reduction operator to perform low-order projection on concurrent tension data to eliminate memory saturation, and then calculates a dynamic tension compensation gain adjustment function according to the tension prediction compensation amount and a low-order tension correlation characteristic matrix, and outputs a feedforward tension compensation instruction to the gripper in stages, thereby solving the technical problem that instantaneous tension mutation caused by liquid impact in the liquid bag shearing process cannot be compensated in real time, resulting in shear size deviation exceeding the limit.

Description

Technical Field

[0001] This invention belongs to the field of liquid bag technology, and more specifically, relates to a method for dynamic tension compensation during the shearing process of a liquid bag. Background Technology

[0002] Liquid bag shearing is a critical process in the production lines of flexible packaging, medical infusion, and industrial flexible containers. Traditional clamping force control methods rely on preset fixed parameters or closed-loop adjustment based on force sensor feedback, which can maintain acceptable dimensional accuracy under steady-state shearing conditions. However, the liquid inside the bag generates a nonlinear fluid shock wave the moment the shearing tool intervenes, causing a drastic change in tension in the gripper area within milliseconds. Traditional feedback control, due to its inherent detection delay and actuator response lag, cannot effectively correct the clamping force within the propagation period of the liquid shock wave. In current high-speed continuous shearing conditions, the amplitude and frequency of instantaneous tension changes are highly uncertain due to variations in the liquid bag filling volume, batch fluctuations in material viscoelasticity, and changes in shearing speed. Existing fixed-gain feedforward methods cannot adaptively match the tension disturbance characteristics under different operating conditions, while deep learning prediction methods based on full-volume time-series data face engineering constraints such as edge computing node memory saturation and excessive inference latency. In other words, existing technologies suffer from the technical problem that the instantaneous tension changes caused by liquid impact during liquid bag shearing cannot be predicted and compensated for in real time, leading to excessive shearing dimensional deviations. Summary of the Invention

[0003] In view of this, the present invention provides a method for dynamic tension compensation during the shearing process of a liquid bag, which can solve the technical problem in the prior art that the instantaneous tension change caused by liquid impact during the shearing process of a liquid bag cannot be predicted and compensated in real time, resulting in excessive deviation of shearing dimensions.

[0004] This invention is implemented as follows: This invention provides a method for dynamic tension compensation during the shearing process of a liquid bag, comprising the following steps:

[0005] A high-frequency sampling unit with a flow window is deployed on the gripper of the liquid bag shearing device to collect material elongation, instantaneous acceleration and three-dimensional transient tension data, and to construct the three-dimensional transient tension characteristic tensor of the liquid bag;

[0006] The three-dimensional transient tension feature tensor of the liquid bag is input into the hybrid expert routing tension prediction network. The tension is transformed into a reference vector through spatial orthogonal decoupling. The gated routing is driven by the Fibonacci golden section dynamic routing optimization algorithm, which activates the top two expert networks and outputs the tension prediction compensation amount for the next moment.

[0007] A lossless dimensionality reduction operator for high-dimensional data is constructed. The tension, velocity, and displacement concurrent data acquired by streaming are subjected to a low-order projection of the tension correlation matrix, and the low-order tension correlation feature matrix is ​​extracted to eliminate memory saturation.

[0008] Based on the tension prediction compensation amount and the low-order tension correlation feature matrix, the dynamic tension compensation gain adjustment function is calculated, the gain adjustment value is output, and different levels of feedforward tension compensation commands are fed back to the gripper according to the interval to which the gain adjustment value belongs.

[0009] Optionally, it also includes: the gripper dynamically adjusts the clamping force according to the feedforward tension compensation command to compensate for the instantaneous tension change caused by the liquid impact during the shearing of the liquid bag, so as to realize closed-loop control of the shearing size.

[0010] Among them, the streaming sliding window high-frequency sampling unit refers to the sliding time window sampling module deployed on the edge computing node, which concurrently collects and buffers the outputs of tension sensor, displacement sensor and acceleration sensor in each window to form a continuous streaming data frame.

[0011] Specifically, the three-dimensional transient tension feature tensor of the liquid bag is a high-dimensional numerical matrix formed by aligning the longitudinal elongation, transverse elongation, shear tension, instantaneous acceleration rate of change, and timestamp of the material in three-dimensional space. The dimensional arrangement consists of batch dimension, time dimension, and feature dimension. Tensor.

[0012] Spatial orthogonal decoupling refers to applying an orthogonal transformation matrix along the characteristic dimension to the three-dimensional transient tension characteristic tensor of the liquid bag. By projecting each tension component onto mutually orthogonal base directions, a reference vector is obtained. Orthogonal transformation matrix It is obtained by performing singular value decomposition on historical data.

[0013] The hybrid expert routing tension prediction network includes a gated routing layer, which introduces viscoelastic property constraint equations for the liquid bag material. Calculate the property correction factor The routing features are obtained by performing matrix multiplication with the input features.

[0014] The Fibonacci golden section dynamic routing optimization algorithm maps the gating activation threshold space of four independent expert networks to a one-dimensional interval. It uses the golden section point sequence to perform two-point sampling iterative optimization of the energy distribution of routing features. In each iteration, the interval is shrunk and the conditional branch selection of the gating unit is updated, outputting a sparse weight vector. The calculation is only performed on the expert network for the top two scorers.

[0015] The hybrid expert route tension prediction network also includes an adaptive exponentially decaying sliding window operator to adjust the expert fusion output matrix. Apply weights along the time dimension This causes the feature attention weights of historical time steps to decay exponentially over time, compressing memory usage from the level of the entire sequence to the level of a fixed window.

[0016] Among them, the hybrid expert routing tension prediction network also includes a dynamic backtracking mechanism. When the real-time prediction error rate exceeds a preset threshold, the information flow jumps back to the front-end gating layer to dynamically adjust the routing weight.

[0017] The output of the hybrid expert routing tension prediction network uses a residual dense hop matrix. The original input is directly added to the expert fusion output, that is... Output tension prediction compensation amount .

[0018] Among them, the lossless dimensionality reduction operator for high-dimensional data reduces the size of the tension correlation matrix. Perform truncated singular value decomposition, retaining the first... The projection matrix is ​​formed by the left singular vectors corresponding to the singular values. The low-order tension correlation feature matrix is ​​obtained. The reconstruction error after dimensionality reduction is less than the preset reconstruction error threshold.

[0019] The dynamic tension compensation gain adjustment function is defined as follows: ,in For tension prediction compensation amount Norm, is the spectral radius of the low-order tension correlation characteristic matrix, and both are normalized to make the gain adjustment value a dimensionless scalar.

[0020] Among them, the feedforward tension compensation command is divided into low-gain feedforward compensation mode, medium-gain feedforward compensation mode and high-gain feedforward compensation mode according to the range of the gain adjustment value, and the corresponding clamping force adjustment amounts are respectively , and ,and .

[0021] This also includes an online update step: after each shearing is completed, the three-dimensional transient tension feature tensor of the liquid bag and the corresponding actual cutting size deviation are written into the online update data buffer of the hybrid expert routing tension prediction network, triggering incremental updates of network parameters.

[0022] The online update data buffer is a first-in-first-out queue-style memory area on the edge computing node. When the buffer reaches the capacity threshold, it triggers an incremental update of parameters. The learning rate of the incremental update is a reduction of the initial training learning rate.

[0023] The sampling frequency of the flowing sliding window high-frequency sampling unit is 100–2000 Hz, the window length is 20–200 ms, and the step interval is 0.5–5 ms; the feature dimension of the liquid bag three-dimensional transient tension feature tensor is... The number of iterations for the Fibonacci golden section dynamic routing optimization algorithm is 5–12; the window length for the adaptive exponential decay sliding window operator is 16–128. The attenuation coefficient is 20–80. satisfy ,in The range is 0.5 to 2.0; the preset reconstruction error threshold is 0.01 to 0.05, and the projection dimension is... The gain threshold is 4–32. The value is between 0.15 and 0.35. The gain coefficient is 0.55 to 0.80. The value ranges from 0.8 to 1.2. It ranges from 1.5 to 2.5. The learning rate is 3.0 to 5.0; the online update data buffer capacity is 500 to 5000 frames, and the incremental update learning rate is 0.01 to 0.1 times the initial training learning rate.

[0024] This invention acquires three-dimensional transient tension feature tensors by deploying a high-frequency sampling unit with a flowing sliding window on the gripper. It then uses a hybrid expert routing tension prediction network to output the predicted tension compensation for the next time step. The concurrent data is compressed using a high-dimensional data lossless dimensionality reduction operator, and feedforward tension compensation commands are output in stages using a dynamic tension compensation gain adjustment function. This achieves advanced prediction and graded compensation for instantaneous tension changes caused by liquid impact. The hybrid expert routing tension prediction network employs a Fibonacci golden section dynamic routing optimization algorithm, ensuring that routing decisions are completed within the sampling interval, guaranteeing the real-time performance of the feedforward commands. Spatial orthogonal decoupling eliminates linear coupling between tension components, reducing prediction errors. An adaptive exponential decay sliding window operator compresses memory usage from the full sequence level to the fixed window level, eliminating the risk of memory saturation at edge nodes. In summary, this invention solves the technical problem mentioned in the background art: the instantaneous tension changes caused by liquid impact during liquid bag shearing cannot be predicted and compensated in real time, leading to excessive shearing size deviations. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention.

[0026] Figure 2 This is a diagram illustrating the iterative convergence process of the Fibonacci golden ratio dynamic routing optimization algorithm.

[0027] Figure 3 This is a comparison chart of the measured tension and the predicted tension compensation during continuous shearing. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shown is a flowchart of a method for dynamic tension compensation during the shearing process of a liquid bag provided by the present invention. This method includes the following steps:

[0030] S01. Deploy a flow-type sliding window high-frequency sampling unit on the gripper of the liquid bag shearing device to collect material elongation, instantaneous acceleration and three-dimensional transient tension data at a sampling frequency of 100-2000Hz, and construct the three-dimensional transient tension characteristic tensor of the liquid bag.

[0031] S02. Input the three-dimensional transient tension feature tensor of the liquid bag into the hybrid expert routing tension prediction network. Transform the tension into a reference vector through spatial orthogonal decoupling. Drive the gated routing through the Fibonacci golden section dynamic routing optimization algorithm, activate the top two expert networks, and output the tension prediction compensation amount for the next moment.

[0032] S03. Construct a lossless dimensionality reduction operator for high-dimensional data, perform low-order projection of the tension correlation matrix on the massive concurrent tension, velocity, and displacement data collected by streaming, extract the low-order tension correlation feature matrix, and eliminate memory saturation.

[0033] S04. Based on the tension prediction compensation amount and the low-order tension correlation feature matrix, calculate the dynamic tension compensation gain adjustment function, output the gain adjustment value, and feed back different levels of feedforward tension compensation commands to the gripper according to the interval to which the gain adjustment value belongs.

[0034] S05. The gripper dynamically corrects the clamping force based on the feedforward tension compensation command to compensate for the instantaneous tension change caused by the liquid impact during the shearing of the liquid bag, thereby achieving closed-loop control of the shearing size.

[0035] S06. Optionally, it also includes: after each shearing is completed, writing the collected three-dimensional transient tension feature tensor of the liquid bag and the corresponding actual cutting size deviation into the online update data buffer of the hybrid expert routing tension prediction network, triggering incremental updates of network parameters, and continuously optimizing the accuracy of tension prediction compensation.

[0036] Among them, the streaming sliding window high-frequency sampling unit refers to the sliding time window sampling module deployed on the edge computing node. The window length is 20 to 200 ms and the step interval is 0.5 to 5 ms. Within each window, the outputs of the tension sensor, displacement sensor and acceleration sensor are concurrently acquired and buffered to form a continuous streaming data frame for subsequent feature extraction.

[0037] The three-dimensional transient tension feature tensor of the liquid bag refers to a high-dimensional numerical matrix formed by aligning the longitudinal elongation, transverse elongation, shear tension, instantaneous acceleration rate of change, and timestamp of the material in three-dimensional space. The dimension alignment method is to arrange the above features along the batch dimension, time dimension, and feature dimension. The tensor, in which For batch size, For time steps, For feature dimension, The value is determined by summing the actual number of sensor channels and the number of channels after feature engineering expansion, and the recommended value range is 16 to 128.

[0038] Spatial orthogonal decoupling refers to applying an orthogonal transformation matrix along the characteristic dimension to the three-dimensional transient tension characteristic tensor of the liquid bag. By projecting each tension component onto mutually orthogonal base directions, the linear correlation between different tension components is eliminated, thus obtaining the reference vector. ,in For the input feature vector, The data is obtained by performing singular value decomposition on historical data, and the decomposition sample size is recommended to be no less than 5000 frames.

[0039] The specific structure of the hybrid expert routing tension prediction network is as follows: the network receiver dimension is... The three-dimensional transient tension feature tensor of the liquid bag is transformed into a reference vector matrix by a pre-space orthogonal decoupling layer. The gating routing layer introduces viscoelastic property constraint equations for the liquid bag material. As the independent variable, calculate the property correction factor. , and input features Matrix multiplication is used to obtain routing features. The Fibonacci golden ratio dynamic routing optimization algorithm is nested within the gated routing layer, defining the gate activation threshold space of the four independent expert networks as a one-dimensional interval. Generate a sequence of golden section points using the Fibonacci sequence. For routing features The energy distribution is sampled at two points for iterative optimization. Each iteration shrinks the interval and updates the conditional branch selection of the gating unit. The recommended number of iterations is 5 to 12. This quickly locks the best matching interval between the current feature and the four expert networks, and outputs a sparse weight vector. Activation calculations are performed only on the expert networks of the top two scorers; each of the four expert networks consists of three fully connected layers, with a recommended number of 64–256 neurons in the hidden layers. The neuron weight matrices between adjacent expert networks are initialized using a differentiated method to allow each expert to learn different frequency band tension fluctuation patterns, and the inter-layer activation function uses modified linear units; the output of each expert network is expressed through a sparse weight vector. Weighted fusion yields the expert fusion output matrix Adaptive exponential decay sliding window operator pair Apply weights along the time dimension ,in The attenuation coefficient is... For the current time step, For historical time steps, the feature attention weights of historical time steps decay exponentially with the time span. This operator reduces memory usage from the entire sequence. Compress to a fixed window Window length A value of 20-80 is recommended; when the real-time prediction error rate... Exceeding the preset threshold When this occurs, the hidden layer's dynamic backtracking mechanism is triggered, and the information flow jumps back to the front-end gating layer, dynamically adjusting the routing weights. The 95th quantile of the prediction error distribution is determined by statistical analysis on multiple batch validation sets, with an initial reference range of 0.05–0.15 recommended. A residual-dense jump matrix is ​​used at the output. , the original input Integration with experts for output Direct addition, that is Output the tension prediction compensation amount for the next moment. The connection weights between neurons are stored independently in memory partitions grouped by expert network. The partition size is proportional to the number of experts, with each of the four expert networks occupying 20%–28% of the total memory. The remaining memory is used for the gated routing layer and the output layer. Inter-layer computational data flows are pipelined in parallel scheduling in the batch dimension, allowing for partial overlap of forward computations in adjacent layers on the time axis, reducing inter-layer latency. The window length of the adaptive exponential decay sliding window operator is... The update frequency of the sparse weight vector of the gated routing layer They satisfy a functional relationship ,in To ensure the scaling factor is calibrated experimentally, a reference range of 800–3200 is recommended. This relationship ensures that the window automatically shrinks to save video memory when routing updates are frequent, and appropriately expands the window to utilize longer historical information when routing is stable; batch processing is performed during data looping. The size of the attenuation coefficient in the adaptive exponential attenuation sliding window operator The relationship between them is satisfied ,in The correction factor, adjusted experimentally, is recommended to be within the range of 0.5–2.0. This relationship ensures slower historical weight decay when batch sizes are large, making full use of long historical samples, while accelerating decay when batch sizes are small to reduce historical noise interference. The total number of overall network parameters is recommended to be controlled within a certain range. ~ The scale is adjusted to ensure the feasibility of memory allocation for edge computing nodes.

[0040] The technical effects of the hybrid expert routing tension prediction network are as follows: spatial orthogonal decoupling eliminates linear coupling between tension components, enabling gated routing to obtain a non-redundant input representation and reducing the routing misjudgment rate; viscoelastic property constraint equations, as physical priors, directly participate in the calculation of gated weights, ensuring the network maintains routing stability even when the liquid bag material parameters change; the Fibonacci golden section dynamic routing optimization algorithm replaces fully connected classification calculation with fast-converging interval optimization, significantly compressing the single-step inference time of the routing layer, making the output delay of tension compensation commands much lower than the propagation period of liquid shock waves; the adaptive exponential decay sliding window operator processes streaming time-series data of arbitrary length with a fixed window memory usage, avoiding memory saturation; the dynamic backtracking mechanism automatically corrects routing weights when prediction deviations increase sharply, enhancing the network's robustness under transient liquid shock conditions; the residual dense jump matrix retains the low-frequency trend components of the original tension input, superimposed with the high-frequency shock features extracted by the expert network, enabling the predicted compensation amount to simultaneously cover both slowly varying tension drift and instantaneous shock disturbances.

[0041] The specific steps for establishing the training dataset for the hybrid expert routing tension prediction network include: conducting at least 200 shearing experiments on a liquid bag shearing test bench under different combinations of liquid bag materials (elastic modulus range recommended to cover 0.5–5.0 MPa), different liquid bag filling amounts (filling rate recommended to cover 30%–90%), and different shearing speeds (recommended to cover 0.1–2.0 m / s). For each experiment, a complete three-dimensional transient tension feature tensor sequence of the liquid bag and the corresponding actual cutting size deviation are collected using a flow-through sliding window high-frequency sampling unit. The collected data undergoes abnormal frame removal (using a 3x standard deviation criterion), time axis alignment, and normalization. The normalization method is to subtract the mean and divide the standard deviation for each feature dimension. The training set, validation set, and test set are divided in an 8:1:1 ratio. A preset threshold is determined using the prediction error distribution on the validation set. Reference range.

[0042] The training steps of the hybrid expert routing tension prediction network specifically include: using the weighted sum of the mean square error and the physical property constraint residuals as the loss function, i.e. ,in This is the weighting coefficient for physical property constraints; a suggested reference range is 0.01 to 0.1. The learning rate is determined by a grid search on the validation set, balancing prediction error and physical property residuals; an adaptive moment estimator optimizer is employed, with an initial learning rate range suggested as follows: ~ Batch size The recommended range is 16–64; the recommended number of training epochs is 100–500; training should be stopped early when the validation set loss no longer decreases for 20 consecutive epochs; the real-time prediction error rate should be calculated on the validation set every 10 epochs, and the preset threshold should be updated accordingly. After training, evaluate the tension prediction compensation on the test set. The mean absolute error is used to confirm that the model's generalization ability meets engineering requirements.

[0043] Among them, the lossless dimensionality reduction operator for high-dimensional data refers to applying a topologically preserving linear projection to the feature dimension of the three-dimensional transient tension feature tensor of the liquid bag, specifically through the tension correlation matrix of historical data. ( For the first Wei and Di The Pearson correlation coefficient of the tension characteristics is truncated and singular value decomposition is performed, retaining the first... The projection matrix is ​​formed by the left singular vectors corresponding to the singular values. The low-order tension correlation feature matrix is ​​obtained. The term "lossless" refers to achieving dimensionality reduction while preserving the geometric topology of the multidimensional feature space, meaning that the reconstruction error after dimensionality reduction should be minimized. Less than the preset reconstruction error threshold , Experiments were conducted on the validation set to calibrate the prediction accuracy loss after dimensionality reduction; a suggested reference range is 0.01–0.05. The value is determined by the cumulative contribution rate of singular values. It is recommended to retain the value corresponding to a cumulative contribution rate exceeding 95%. value, The actual reference range is 4 to 32.

[0044] Among them, the dynamic tension compensation gain adjustment function refers to the compensation amount based on the tension prediction. Characteristic matrix associated with low-order tension Scalar functions for comprehensive calculation Specifically defined as ,in For tension prediction compensation amount Norm, The spectral radius (i.e., the maximum singular value) of the low-order tension correlation characteristic matrix is ​​used to unify the dimensions of both. To calibrate the full-scale tension value (unit Normalization, for Normalized by the maximum value of the historical spectral radius, so that It is a dimensionless scalar; when At that time, a low-gain feedforward compensation mode is adopted, and the clamping force adjustment is... ;when At that time, a medium-gain feedforward compensation mode is adopted, and the clamping force adjustment is... ;when At that time, a high-gain feedforward compensation mode is adopted, and the clamping force adjustment is... ;in Gain coefficient , , and threshold , Through iterative experimental calibration of the relationship between cutting size deviation and clamping force response in multiple shearing experiments, it is recommended that... The reference range is 0.15 to 0.35. The reference range is 0.55 to 0.80. The reference range is 0.8 to 1.2. The reference range is 1.5 to 2.5. The reference range is 3.0 to 5.0. All of the above reference ranges need to be confirmed through no less than 3 rounds of iterative experiments under actual production line conditions.

[0045] The principle of the Fibonacci golden section dynamic routing optimization algorithm and its specific implementation in the scheme are as follows: Fibonacci sequence satisfy The ratio of adjacent terms approaches the golden ratio. The algorithm maps the gating activation threshold space of the four expert networks into a one-dimensional interval. At the start of each gated route calculation, based on the current input features energy Take two golden section sampling points within the interval. and The gated response values ​​at the two points are calculated separately. The interval with the larger response is retained and a new sampling point is generated. After 5 to 12 iterations, the system converges to the optimal activation threshold interval. The conditional branch selection of the gated unit is then directly changed to determine the sparse weight vector. Compared to fully connected classifiers that calculate all threshold candidates one by one, the Fibonacci golden section dynamic routing optimization algorithm only requires an order of magnitude of sampling to converge, significantly reducing the single-step inference time of the routing layer. This allows gated routing to complete routing decisions within the sampling interval of the high-frequency sampling unit of the streaming sliding window even in scenarios of transient tension changes caused by liquid impact, ensuring the real-time performance of the feedforward tension compensation command. The algorithm also directly couples the routing search process with the energy distribution of the three-dimensional transient tension feature tensor of the liquid bag, avoiding drastic changes in routing weights due to occasional noise frames in high-concurrency streaming data, and improving the temporal stability of the tension prediction compensation.

[0046] Among them, the feedforward tension compensation command refers to the compensation amount based on the tension prediction before the tension deviation is detected. The clamping force adjustment command issued in advance to the gripper actuator is used to counteract the instantaneous tension change caused by liquid impact, which is different from the feedback compensation method that responds after detecting the deviation.

[0047] The online update data buffer refers to a first-in-first-out queue-style memory area allocated on the edge computing node. The buffer capacity is recommended to be 500 to 5000 frames. When the buffer reaches the set capacity, an incremental update of the parameters of the hybrid expert routing tension prediction network is triggered. The incremental update uses the same loss function as the initial training. The learning rate is recommended to be reduced to 0.01 to 0.1 times the initial training learning rate to prevent catastrophic forgetting caused by overfitting to historical distribution shift.

[0048] Among them, the viscoelastic property constraint equation middle, The measured tension at the current moment. The elastic modulus of the liquid bag material. For the current elongation strain, For strain rate, The viscosity coefficient of the material and The results were obtained through dynamic mechanical analysis experiments on liquid bag material samples. The recommended experimental frequency range is 1–100 Hz, and the temperature range is recommended to cover the actual operating temperature range. Property correction factors were also considered. By embedding the aforementioned material physical priors into network gating calculations, the routing weights are automatically adjusted when material parameters change, reducing the drift in tension prediction compensation caused by batch differences in materials.

[0049] The instantaneous acceleration change rate refers to the scalar obtained by taking the first-order difference of the instantaneous acceleration signal output by the accelerometer along the time axis. It reflects the intensity change of the nonlinear shock wave near the shear point of the liquid bag and is used to adaptively adjust the response sensitivity of the feedforward tension compensation gain.

[0050] As an optional approach, the present invention also provides a computer-based dynamic tension compensation system, wherein the computer is provided with a readable storage medium, and the readable storage medium stores program instructions. When the program instructions are run in the computer, they are used to drive the high-frequency sampling unit of the flowing sliding window, the hybrid expert routing tension prediction network, and the gripper actuator to collaboratively execute the above-mentioned method for dynamic tension compensation in the liquid bag shearing process, thereby realizing the full-process computerized control of tension acquisition, dimensionality reduction processing, compensation amount prediction, and clamping force adjustment.

[0051] The specific implementation of step S01 is as follows: a high-frequency sampling unit with a sliding window is integrated into the gripper mechanical structure of the liquid bag shearing device. The sampling frequency is set in the range of 100-2000Hz, and signals are output concurrently from three types of sensors: tension sensor, displacement sensor, and acceleration sensor. The sliding time window sampling module runs on the edge computing node, with a window length set to 20-200ms and a step interval set to 0.5-5ms. Within each window, the synchronous acquisition and buffering of the above three types of sensors are completed, forming a continuous streaming data frame. Each frame of data includes the longitudinal elongation, transverse elongation, shear tension, instantaneous acceleration rate of change, and a timestamp. The instantaneous acceleration rate of change is obtained by calculating the first-order difference of the acceleration sensor output along the time axis. The above features are arranged along the batch dimension, time dimension, and feature dimension as follows: The three-dimensional transient tension feature tensor of the liquid bag, feature dimension The value is determined by summing the actual number of sensor channels and the number of channels expanded by feature engineering, with a recommended range of 16–128. The purpose of this step is to capture shock precursor information at a sampling rate higher than the propagation frequency of the liquid shock wave, providing sufficient temporal resolution input data for the subsequent prediction network.

[0052] The specific implementation of step S02 is as follows: the three-dimensional transient tension feature tensor of the liquid bag is fed into a hybrid expert routing tension prediction network for inference. The spatial orthogonal decoupling layer at the front end of the network applies an orthogonal transformation matrix to the input tensor along the feature dimension. This matrix is ​​pre-obtained by performing singular value decomposition on the tension correlation matrix of no less than 5000 frames of historical data, and the transformation result is a reference vector matrix. This eliminates the linear correlation between tension components. The gated routing layer uses viscoelastic property constraints equations. Calculate the property correction factor Among them, elastic modulus With viscosity coefficient The property correction factor and reference vector matrix were pre-calibrated through dynamic mechanical analysis experiments. Matrix multiplication is used to obtain routing features. The Fibonacci golden ratio dynamic routing optimization algorithm operates within the gated routing layer, mapping the gating activation threshold space of the four expert networks to a one-dimensional interval. Based on current routing characteristics energy Within the interval, two golden section sampling points are selected, and the gating response values ​​at these two points are calculated. The interval with the larger response is retained, and after 5 to 12 iterations, it converges to the optimal activation threshold, thus determining the sparse weight vector. Only the top two expert networks in terms of score are activated. Each of the four expert networks consists of three fully connected layers, with modified linear unit activation functions used between layers. The weight matrices of adjacent expert networks are differentially initialized. The output of each expert network is processed by a sparse weight vector. Weighted fusion Adaptive exponential decay sliding window operator pair Apply weights along the time dimension attenuation coefficient satisfy Window length satisfy This reduces video memory usage to a fixed window size. When the real-time prediction error rate exceeds a preset threshold... At this time, the dynamic backtracking mechanism will redirect the information flow back to the gating layer to correct the routing weights, with a preset threshold. The reference range is 0.05 to 0.15. The output is processed through a residual dense jump matrix. The original input and the expert fusion output are directly added together to obtain the tension prediction compensation amount. The purpose of this step is to complete routing decisions and tension prediction within the sampling interval, ensuring the real-time performance of feedforward commands.

[0053] The specific implementation of step S03 is to construct a lossless dimensionality reduction operator for high-dimensional data to eliminate the risk of memory saturation at edge computing nodes. First, the tension correlation matrix of historically collected data is analyzed. Perform truncated singular value decomposition. For the first Wei and Di Pearson correlation coefficient of tension characteristics, retaining the first The projection matrix is ​​formed by the left singular vectors corresponding to the singular values. , The value is determined by the sum of singular values ​​when their cumulative contribution rate exceeds 95%, with a suggested range of 4–32. A low-order tension correlation feature matrix is ​​obtained by projecting the concurrent tension, velocity, and displacement data acquired via convection. Reconstruction error after dimensionality reduction It must be less than the preset reconstruction error threshold. A reference range of 0.01 to 0.05 is recommended. This constraint ensures that the geometric topology of the multidimensional feature space is preserved during the dimensionality reduction process. The purpose of this step is to compress high-dimensional concurrent data into a low-dimensional subspace, eliminating memory saturation without losing key tension correlation information.

[0054] The specific implementation of step S04 is to predict the compensation amount based on the tension. Characteristic matrix associated with low-order tension Calculate the dynamic tension compensation gain adjustment function ,in To calibrate the full-scale tension value Normalization Normalized by the maximum value of the historical spectral radius, so that It is a dimensionless scalar. Based on... The corresponding interval outputs different levels of feedforward tension compensation commands: when When using low-gain feedforward compensation mode, the clamping force adjustment is: ;when The medium-gain feedforward compensation mode is adopted, and the clamping force adjustment is... ;when The high-gain feedforward compensation mode is used, and the clamping force adjustment is... Gain coefficient The value ranges from 0.8 to 1.2. It ranges from 1.5 to 2.5. The threshold is 3.0 to 5.0. The value is between 0.15 and 0.35. The values ​​are 0.55 to 0.80, and these parameters are calibrated through at least three rounds of iterative experiments. The purpose of this step is to adaptively match the compensation gain according to the actual magnitude of the impact intensity, avoiding overcompensation or undercompensation with a single fixed gain.

[0055] The specific implementation of step S05 is as follows: After receiving the feedforward tension compensation command, the gripper actuator adjusts the clamping force in advance before detecting the tension deviation to counteract the instantaneous tension change caused by the liquid shock wave. The amount of clamping force adjustment is determined by step S04. After the adjustment is completed, the shearing size data is fed back to the control system, forming a closed-loop control loop for the shearing size. The optional implementation of step S06 is as follows: After each shearing operation, the three-dimensional transient tension feature tensor of the liquid bag and the actual cutting size deviation are written into an online update data buffer. The buffer is a first-in-first-out queue-style memory area with a capacity of 500-5000 frames. After reaching the capacity threshold, the incremental parameter update of the hybrid expert routing tension prediction network is triggered. The incremental update uses the same loss function as the initial training. The learning rate is reduced to 0.01 to 0.1 times the initial training learning rate to prevent catastrophic forgetting. The purpose of this step is to enable the network to continuously adapt to material batch drift and changes in operating conditions, maintaining the long-term accuracy of tension prediction compensation.

[0056] It should be noted that the key technologies of this invention include: a Fibonacci golden section dynamic routing optimization algorithm in a hybrid expert routing tension prediction network, which compresses the computational load of routing decisions from linear to logarithmic through iteration at two golden sections, making the routing delay lower than the propagation period of a liquid shock wave, thus overcoming the inference bottleneck of traditional fully connected routing in high-frequency streaming data scenarios; joint embedding of spatial orthogonal decoupling and viscoelastic property constraint equations, the former eliminating linear coupling between tension components to reduce the routing misjudgment rate, and the latter directly incorporating material physics priors into the gating weight calculation, enabling the prediction network to maintain routing stability without retraining when material batches change; and adaptive exponential decay. The collaborative dimensionality reduction of sliding window operator and lossless dimensionality reduction operator for high-dimensional data achieves the former to replace full sequence accumulation with a fixed window in the time dimension, and the latter to replace full-dimensional concurrent storage with low-rank projection in the feature dimension. Together, they transform the memory usage of edge nodes from linear growth over time to a constant order of magnitude, fundamentally eliminating the risk of memory saturation. At the same time, the dynamic tension compensation gain adjustment function adaptively matches the compensation level based on the product of the tension prediction compensation amount and the spectral radius of the low-order feature matrix. The three-level gain partition ensures the compensation accuracy under both weak and strong impact conditions. The above key technologies work together to form a complete closed-loop technology chain from high-frequency sensing, real-time prediction, memory management to adaptive compensation.

[0057] It should be noted that when the elastic modulus or viscosity coefficient of the liquid bag material drifts significantly due to batch variations, the expert network components in the hybrid expert routing tension prediction network will experience routing mismatch due to input feature distribution shifts. This leads to a systematic drift in the tension prediction compensation, resulting in shear size deviations exceeding the allowable range. The reason for this technical problem is that the expert network learns tension fluctuation patterns based on a fixed material parameter distribution during the initial training phase. When material parameters drift in actual production, the routing features calculated by the gated routing layer based on historical weights no longer match the learning components of each expert network. The expert network originally responsible for high-frequency impact features may be incorrectly activated by low-frequency trend features, causing the prediction results to deviate from the actual tension change pattern. A common solution to this problem is to periodically re-collect data offline and retrain the network to update the expert network's weight distribution. However, full retraining requires stopping production to accumulate sufficient samples, cannot respond promptly to material drift during the training period, and the large learning rate used in full retraining can easily overwrite existing impact pattern knowledge, leading to catastrophic forgetting and causing the network's performance to degrade under the adapted operating conditions. This invention effectively solves this technical problem. The online data buffer update employs a first-in, first-out (FIFO) queue structure. After each shearing operation, the latest three-dimensional transient tension feature tensor of the liquid bag and the actual cutting size deviation are written into the buffer. When the buffer accumulates to a set capacity, an incremental parameter update is triggered. The learning rate for the incremental update is reduced to 0.01 to 0.1 times the initial training learning rate, ensuring that the network weights are adjusted to the new distribution only within a small range. This effectively tracks changes in the routing feature distribution caused by material parameter drift while avoiding catastrophic forgetting caused by large weight updates. Simultaneously, the viscoelastic property constraint equation incorporates the material's elastic modulus... With viscosity coefficient By directly embedding the gating weight calculation in the form of a material property correction factor, when material parameter drift occurs, the material property correction factor will automatically adjust the dimensional relationship of the routing features, so that the gating route has a certain material self-adaptation capability before incremental update, reducing the probability of route mismatch, and thus maintaining the accuracy of tension prediction compensation during continuous production.

[0058] Specifically, the principle of this invention is as follows: the fundamental reason why this invention can solve the above-mentioned technical problems lies in the synergistic design of three levels. The first level is the high temporal resolution of signal acquisition and feature construction. The streaming sliding window high-frequency sampling unit concurrently acquires material elongation, instantaneous acceleration, and three-dimensional tension at a frequency of 100-2000Hz, constructing... The three-dimensional transient tension feature tensor, with a temporal resolution higher than the propagation period of the liquid shock wave within the liquid bag, ensures that the features contain precursory information before the shock wave reaches the gripper, a physical prerequisite for advanced prediction. The second layer involves the structural design of a hybrid expert routing tension prediction network that satisfies real-time constraints for both prediction and routing. A spatial orthogonal decoupling layer obtains an orthogonal transformation matrix through singular value decomposition, eliminating linear correlations between different tension components, ensuring gated routing receives non-redundant input, and reducing routing misjudgment rates. The viscoelastic property constraint equations directly embed material elastic modulus, viscosity coefficient, and other physical priors into the gating weight calculation, maintaining routing stability even with batch changes in material parameters. The Fibonacci golden section dynamic routing optimization algorithm maps the gating activation threshold space of the four expert networks to a one-dimensional interval, using iterative sampling at two points along the golden section to converge to the optimal activation threshold. Its iteration count is only logarithmic, far lower than the calculation of all candidate thresholds by a fully connected classifier, significantly reducing routing decision latency to the shock wave propagation period and ensuring real-time output of feedforward commands. The adaptive exponentially decaying sliding window operator applies exponentially decaying weights to historical tension features, retaining only weighted historical information within a fixed-length window. This compresses memory usage from the linear order of the entire sequence to the constant order of the fixed window, fundamentally eliminating the risk of memory saturation at edge nodes. The residual dense jump matrix directly superimposes the original low-frequency tension trend with the high-frequency impact features extracted by the expert network, ensuring that the predicted compensation amount simultaneously covers both slowly varying tension drift and instantaneous impact disturbances. The third layer involves the synergistic effect of the lossless dimensionality reduction operator for high-dimensional data and the dynamic tension compensation gain adjustment function. The lossless dimensionality reduction operator for high-dimensional data performs truncated singular value decomposition on the tension correlation matrix, projecting the high-dimensional concurrent tension data to a lower-order subspace, eliminating memory saturation while preserving the geometric topology of the multidimensional feature space. The dynamic tension compensation gain adjustment function comprehensively adjusts the predicted tension compensation amount. The spectral radius of the norm and low-order eigenma matrix is ​​used to calculate the dimensionless gain scalar. The clamping force adjustment is divided into three levels of feedforward compensation modes (low, medium, and high) according to the gain range, avoiding overcompensation or undercompensation of a single fixed gain under different impact intensity conditions. The above three levels together constitute a closed-loop shear size control link, enabling the present invention to achieve advanced prediction and adaptive hierarchical compensation for instantaneous tension changes under liquid impact transient conditions, thereby solving the technical problem of excessive shear size deviation.

[0059] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0060] The specific implementation of step S01 is as follows: A high-frequency sampling unit with a sliding window is deployed on the gripper of the liquid bag shearing device. The sampling frequency range is 100–2000 Hz, the window length is 20–200 ms, and the step interval is 0.5–5 ms. Within each sliding window, the outputs of the tension sensor, displacement sensor, and acceleration sensor are concurrently acquired and buffered to form a continuous streaming data frame. The longitudinal elongation, transverse elongation, shear tension, instantaneous acceleration rate of change, and timestamp of the material are aligned along the batch dimension, time dimension, and feature dimension to construct a three-dimensional transient tension feature tensor of the liquid bag. ,in For batch size, For time steps, The feature dimension, with values ​​ranging from 16 to 128. Instantaneous rate of change of acceleration. Defined as the output of the accelerometer sensor The first-order difference along the time axis is expressed as follows:

[0061] ;

[0062] In the formula, For the first Instantaneous acceleration at time (unit) ), For the first Instantaneous acceleration at time (unit) ), The time interval (in seconds) between adjacent sampling times. The unit is .

[0063] The specific implementation of step S02 is as follows. The three-dimensional transient tension characteristic tensor of the liquid bag is... Input a hybrid expert routing tension prediction network. Pre-installed spatial orthogonal decoupling layer. Apply an orthogonal transformation matrix along the feature dimension By projecting each tension component onto mutually orthogonal base directions, the linear correlation between different tension components is eliminated, resulting in the reference vector matrix:

[0064] ;

[0065] In the formula, The reference vector matrix, The singular value decomposition (SVD) is performed on a matrix composed of at least 5000 frames of historical data. The specific process of SVD is as follows: [The text then abruptly shifts to a different topic:] ...the historical data matrix... ( Decomposition is performed for historical frames. Take the left singular vector matrix The former Column as ,in It is a singular value diagonal matrix. This is a right singular vector matrix. The gated routing layer introduces the viscoelastic property constraint equations for the liquid bag material:

[0066] ;

[0067] In the formula, The measured tension (in N) at the current moment. The elastic modulus of the liquid bag material (unit: MPa). The current elongation strain is dimensionless. strain rate (unit) ), The viscosity coefficient of the material (unit) ), The unit is N, which reflects the deviation between the measured tension of the material and the predicted tension based on linear elasticity plus viscosity. and The property correction factor was obtained through dynamic mechanical analysis experiments on liquid bag material samples, with experimental frequencies covering 1–100 Hz and temperature ranges covering the actual operating temperature range. The formula for calculating the property correction factor is as follows:

[0068] ;

[0069] In the formula, Learnable weight vector (unit) ,make (a dimensionless quantity) The hyperbolic tangent activation function is used. This is a property correction factor matrix, with each element taking values ​​ranging from 1 to 2. The routing feature matrix is ​​obtained by element-wise multiplication:

[0070] ;

[0071] In the formula, This is the routing feature matrix. This is the element-wise multiplication operator. The Fibonacci golden ratio dynamic routing optimization algorithm is executed in the gated routing layer, mapping the gating activation threshold space of the four expert networks to a one-dimensional interval. Based on the current input features The Frobenius norm of the matrix Taking two golden section sampling points within the interval, the formula is as follows:

[0072] ;

[0073] ;

[0074] In the formula, and These represent the lower and upper bounds of the gating activation threshold space interval, respectively. The golden ratio, derived from the Fibonacci sequence. The approximate value of the ratio of adjacent terms is determined. , These are the two sampling points for the current iteration step. In each iteration, the gating response values ​​at both sampling points are calculated, the interval with the larger response is retained, and the sampling points are updated. After 5–12 iterations, the process converges to the optimal activation threshold interval, and a sparse weight vector is output. Activation calculations are performed only on the expert networks of the top two scorers. Each expert network is a three-layer fully connected structure with modified linear units as the inter-layer activation function and 64–256 hidden layer neurons. The outputs of each expert network are then weighted and fused using sparse weights to obtain the expert fusion output matrix.

[0075] ;

[0076] In the formula, For the first Sparse weights corresponding to each expert network ( And at most two (non-zero) For the first An expert network analyzes the routing feature matrix. The output, For expert fusion output matrix. Adaptive exponential decay sliding window operator pair Apply decay weights along the time dimension:

[0077] ;

[0078] In the formula, The attenuation coefficient is... Index for the current time step, For historical time step indexing, For the first The decay weight (dimensionless) corresponding to each time step; the weight decreases as the historical time step becomes further back. Window length. Frequency of sparse weight vector update in gated routing layer satisfy:

[0079] ;

[0080] In the formula, The proportionality coefficient (unit: ) Same, make (These are dimensionless integers), with an empirical range of 800 to 3200. The update frequency (in Hz) for route weights. For floor operation, The empirical value range is 20–80. Batch size With attenuation coefficient satisfy:

[0081] ;

[0082] In the formula, The correction factor (dimensionless) has an empirical range of 0.5 to 2.0. This relationship results in slower decay of historical weights when the batch size is large, and faster decay when the batch size is small. When the real-time prediction error rate... (Dimensionless) Exceeds the preset threshold When this occurs, the hidden layer's dynamic backtracking mechanism is triggered, and the information flow redirects back to the front-end gating layer to dynamically adjust the routing weights. The initial reference range is 0.05–0.15, determined by the 95th quantile of the prediction error distribution on the validation set. The output layer employs a residual-dense jump structure.

[0083] ;

[0084] In the formula, The tension prediction compensation amount at the current moment (Unit: N) For the residual jump matrix (with dimensional coefficients, making...) and (Unity consistent) Given the original input feature vector, this structure superimposes low-frequency trend components with high-frequency impact features in the output. The network loss function is:

[0085] ;

[0086] In the formula, The mean square error (in units) between the predicted value and the actual tension compensation. ), For the residual term of physical property constraints (unit) ), The weighting coefficients for physical property constraints (dimensionless) have an empirical range of 0.01 to 0.1, determined through a grid search on the validation set. The unit is .

[0087] The specific implementation of step S03 is as follows: Construct a lossless dimensionality reduction operator for high-dimensional data, and reduce the feature dimension... Calculate the tension correlation matrix , of which Line 1 Column elements For the first Wei and Di Pearson correlation coefficient of tension characteristics (dimensionless, range of values) ).right Perform truncated singular value decomposition, retaining the first... The projection matrix is ​​formed by the left singular vectors corresponding to the singular values. The low-order tension correlation feature matrix is ​​obtained:

[0088] ;

[0089] In the formula, This is a low-order tension correlation feature matrix. The number of singular values ​​is determined by the cumulative contribution rate of singular values ​​exceeding 95%, with a practical reference range of 4–32. The lossless dimensionality reduction requirement satisfies the reconstruction error constraint.

[0090] ;

[0091] In the formula, The preset reconstruction error threshold (dimensionless) has an empirical range of 0.01 to 0.05, which is calibrated through validation set experiments. The left side of the equation is the dimensionless reconstruction error ratio.

[0092] The specific implementation method of step S04 is as follows. Based on... and Calculate the dynamic tension compensation gain adjustment function:

[0093] ;

[0094] In the formula, For tension prediction compensation amount Norm (unit: N) To calibrate the full-scale tension value (in N). Low-order tension correlation characteristic matrix The spectral radius is the maximum singular value (dimensionless, because...). (After projection by Pearson correlation coefficient, it becomes a dimensionless matrix). It is the maximum value of the spectral radius at all historical moments (dimensionless). It is a dimensionless scalar. Based on... The corresponding level of feedforward tension compensation command is output within the specified interval: when At that time, the clamping force adjustment amount is ;when At that time, the clamping force adjustment amount is ;when At that time, the clamping force adjustment amount is ;in , , , All are dimensionless gain coefficients. , For dimensionless thresholds, empirical values ​​are 0.15–0.35 and 0.55–0.80, respectively. The empirical value range is 0.8 to 1.2. The empirical value ranges from 1.5 to 2.5. The empirical value range is 3.0 to 5.0. The above parameters are calibrated under actual production line conditions through no less than 3 rounds of iterative experiments.

[0095] The specific implementation of step S05 is as follows: The gripper adjusts the clamping force in advance according to the feedforward tension compensation command before detecting the tension deviation, thereby offsetting the instantaneous tension change caused by liquid impact and realizing closed-loop control of the shearing dimension.

[0096] The specific implementation of step S06 is as follows. After each shearing is completed, the three-dimensional transient tension feature tensor of the liquid bag and the corresponding actual cutting size deviation are written into the online update data buffer. The buffer is a first-in-first-out queue-style memory area with a capacity of 500 to 5000 frames. After the set capacity is reached, the incremental update of network parameters is triggered. The incremental update uses the same loss function as the initial training. The learning rate is reduced to 0.01 to 0.1 times the initial training learning rate to prevent catastrophic forgetting and to continuously optimize. The accuracy of the prediction.

[0097] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0098] Using a continuous shearing production line for flexible liquid packaging bags as the test scenario, the nominal elastic modulus of the liquid bag material is 1.8 MPa, and the nominal viscosity coefficient is 12. The filling rate is set to 65%, the shearing speed is set to 0.8 m / s, the gripper clamping force is set to 500 N, and the full-scale tension value is [not specified]. The value is 300N. During the test, there was a batch drift of approximately 0.3MPa in the elastic modulus between batches of liquid bags, which was used to verify the tension compensation stability of the present invention under conditions of varying material parameters.

[0099] The sampling frequency of the streaming sliding window high-frequency sampling unit is set to 1000Hz, the window length to 50ms, and the step interval to 1ms. The sensor has 12 channels, which are expanded to 20 channels after feature engineering, and the feature dimensions are... Take 32. Batch size Set to 32, time steps The dimension of the three-dimensional transient tension feature tensor of the liquid bag is set to 50. After data acquisition, each channel's data undergoes normalization by subtracting the mean and dividing by the standard deviation. Time axis alignment is achieved using timestamp interpolation.

[0100] The training dataset for the hybrid expert routing tension prediction network consists of 7200 shearing experiments conducted under 36 different parameter combinations: 3 elastic moduli (1.0 MPa, 1.8 MPa, 3.0 MPa), 3 filling rates (40%, 65%, 85%), and 4 shear velocities (0.2, 0.5, 0.8, 1.5 m / s). After outlier frame removal (using a 3x standard deviation criterion), the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. The orthogonal transformation matrix used in the spatial orthogonal decoupling layer is also included. The network was trained using singular value decomposition on 5000 frames of historical data. An adaptive moment estimation optimizer was used for training, with an initial learning rate of [missing information]. The batch size is 32, the maximum number of training rounds is 300, and the training is stopped early when the validation set loss no longer decreases after 20 consecutive rounds. The weighting coefficients for physical property constraints are... Set to 0.05, preset threshold. The value is determined by the 95th percentile of the prediction error on the validation set, and is set to 0.08.

[0101] In lossless dimensionality reduction operators for high-dimensional data, the tension correlation matrix The value obtained by calculating the Pearson correlation coefficient from historical data is the value when the cumulative contribution rate of the truncated singular value decomposition exceeds 95%. Value, in the test scenario Take 12, projection matrix The dimension of the low-order tension correlation feature matrix is The reconstruction error was calibrated to 0.023 on the validation set, which is lower than the preset reconstruction error threshold of 0.03.

[0102] The parameters of the dynamic tension compensation gain adjustment function were calibrated through three rounds of iterative experiments and finally determined. It is 0.25. The gain coefficient is 0.65. It is 1.0. It is version 2.0. The value is 4.0. During the test, the tension prediction compensation amount before and after the liquid shock occurred. Gain adjustment value The clamping force adjustment amount is shown in Table 1.

[0103] Table 1. Statistical table of tension compensation parameters during a typical shear cycle.

[0104]

[0105] As shown in Table 1, the gain adjustment value increases with increasing liquid impact intensity. As the gain increases, the compensation mode switches from low gain to high gain, and the clamping force adjustment increases proportionally with the impact intensity. The actual cutting size deviation remains at a low level under all impact intensities, indicating that the graded gain mechanism can adaptively match the compensation level under different impact intensities, avoiding the problem of overcompensation in weak impacts or undercompensation in strong impacts caused by the fixed gain method.

[0106] The online update data buffer capacity is set to 2000 frames, and an incremental parameter update is triggered approximately every 40 shearing cycles. The incremental update learning rate is... After the batch elastic modulus drifted by approximately 0.3 MPa, the network underwent two incremental updates (cumulatively covering approximately 80 shear cycles) to determine the tension prediction compensation. The fact that the mean absolute error recovered to the level before the drift indicates that the online update mechanism can continuously track the drift of material parameters and maintain prediction accuracy.

[0107] like Figure 2 As shown, the Fibonacci golden section dynamic routing optimization algorithm exhibits its iterative convergence process within a typical shearing cycle. The horizontal axis represents the number of iterations, and the vertical axis represents the search interval width of the gated activation threshold. The curve shows an exponential contraction trend with increasing iteration count. After 8 iterations, the search interval width converges to approximately 2% of the initial interval width, verifying the algorithm's ability to quickly lock the optimal activation threshold within a limited number of iterations and ensuring the real-time performance of routing decisions. Since the routing decision is completed within the streaming sliding window sampling interval (1 ms), the feedforward tension compensation command can be issued before the liquid shock wave reaches the clamping area, thereby achieving advance compensation for instantaneous tension mutations. This is the core principle behind this invention's ability to reduce shear size deviations compared to traditional feedback control.

[0108] like Figure 3 As shown, this graph illustrates the relationship between the measured tension curve and the predicted tension compensation amount during continuous shearing, with the horizontal axis representing the shearing sequence number and the vertical axis representing the tension or predicted compensation amount (N). Figure 3 It is evident that the measured tension curve exhibits a significant pulse-like abrupt change at each liquid impact moment, while the tension prediction compensation amount... The corresponding compensation value was output before the impact occurred. The peak positions of the two curves showed a lead of approximately 2-3 sampling steps, indicating that the hybrid expert routing tension prediction network can utilize impact precursor information to predict the tension mutation amplitude in advance, enabling the feedforward tension compensation command to be executed before the liquid shock wave reaches the clamping zone. As the shear sequence number increases, the measured tension mutation amplitude fluctuates due to the periodic fluctuations in the liquid filling volume. However, the tracking trend of the tension prediction compensation amount remains highly consistent with the measured tension. This indicates that the weighted application of historical information by the adaptive exponential decay sliding window operator and the stable route selection of the Fibonacci golden section dynamic routing optimization algorithm jointly ensure the robust tracking capability of the prediction curve under different impact intensities.

[0109] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic tension compensation during the shearing process of a liquid bag, characterized in that, Includes the following steps: A high-frequency sampling unit with a flow window is deployed on the gripper of the liquid bag shearing device to collect material elongation, instantaneous acceleration and three-dimensional transient tension data, and to construct the three-dimensional transient tension characteristic tensor of the liquid bag; The three-dimensional transient tension feature tensor of the liquid bag is input into the hybrid expert routing tension prediction network. The tension is transformed into a reference vector through spatial orthogonal decoupling. The gated routing is driven by the Fibonacci golden section dynamic routing optimization algorithm, which activates the top two expert networks and outputs the tension prediction compensation amount for the next moment. A lossless dimensionality reduction operator for high-dimensional data is constructed. The tension, velocity, and displacement concurrent data acquired by streaming are subjected to a low-order projection of the tension correlation matrix, and the low-order tension correlation feature matrix is ​​extracted to eliminate memory saturation. Based on the tension prediction compensation amount and the low-order tension correlation feature matrix, the dynamic tension compensation gain adjustment function is calculated, the gain adjustment value is output, and different levels of feedforward tension compensation commands are fed back to the gripper according to the interval to which the gain adjustment value belongs.

2. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 1, characterized in that, The streaming sliding window high-frequency sampling unit refers to the sliding time window sampling module deployed on the edge computing node, which concurrently acquires and buffers the outputs of tension sensor, displacement sensor and acceleration sensor in each window to form a continuous streaming data frame.

3. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 2, characterized in that, The three-dimensional transient tension feature tensor of the liquid bag is a high-dimensional numerical matrix formed by aligning the longitudinal elongation, transverse elongation, shear tension, instantaneous acceleration rate of change, and timestamp of the material in three-dimensional space. The dimensional arrangement is a tensor composed of batch dimension, time dimension, and feature dimension.

4. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 3, characterized in that, Spatial orthogonal decoupling refers to applying an orthogonal transformation matrix to the three-dimensional transient tension characteristic tensor of the liquid bag along the characteristic dimension, projecting each tension component onto mutually orthogonal basis directions to obtain a reference vector. The orthogonal transformation matrix is ​​obtained by performing singular value decomposition on historical data.

5. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 4, characterized in that, The hybrid expert routing tension prediction network includes a gated routing layer. The gated routing layer introduces the viscoelastic property constraint equation of the liquid bag material to calculate the property correction factor and performs matrix multiplication with the input features to obtain the routing features.

6. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 5, characterized in that, The Fibonacci golden section dynamic routing optimization algorithm maps the gating activation threshold space of four independent expert networks to a one-dimensional interval. It uses a sequence of golden section points to perform two-point sampling and iterative optimization of the energy distribution of routing features. Each iteration shrinks the interval and updates the conditional branch selection of the gating unit, outputting a sparse weight vector. The calculation is only performed on the expert network for the top two scorers.

7. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 6, characterized in that, The hybrid expert routing tension prediction network also includes an adaptive exponentially decaying sliding window operator, which applies weights to the expert fusion output matrix along the time dimension, causing the feature attention weights of historical time steps to decay exponentially with the time span, compressing the memory usage from the full sequence level to the fixed window level.

8. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 7, characterized in that, The hybrid expert routing tension prediction network also includes a dynamic backtracking mechanism. When the real-time prediction error rate exceeds a preset threshold, the information flow jumps back to the front-end gating layer to dynamically adjust the routing weights.

9. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 8, characterized in that, The output of the hybrid expert routing tension prediction network uses a residual dense jump matrix to directly add the original input and the expert fusion output to output the tension prediction compensation.

10. The method for dynamic tension compensation during the shearing process of a liquid bag according to claim 9, characterized in that, Also includes: The gripper dynamically adjusts the clamping force based on the feedforward tension compensation command to compensate for the instantaneous tension change caused by the liquid impact during shearing of the liquid bag, thereby achieving closed-loop control of the shearing size.

Images